Digital twin reservoir flood control scheduling method

Abstract

This application provides a digital twin reservoir flood control scheduling method, comprising: acquiring multi-source reservoir monitoring data and a reservoir spatiotemporal grid; performing spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate; generating time-series grid data based on the digital twin reservoir base plate, and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence; inputting these into a scheduling optimization model, performing candidate scheduling encoding processing and group iterative optimization on a preset reservoir gate opening range and scheduling time unit to obtain a target scheduling sequence; and converting the target scheduling sequence into physical reservoir control commands to control the physical reservoir. This method effectively improves the spatiotemporal alignment effect of multi-source data and the decision-making adaptability of flood control scheduling.

Description

Technical Field

[0001] This application relates to the field of smart water conservancy and flood control scheduling technology, and more specifically, to a digital twin reservoir flood control scheduling method. Background Technology

[0002] With the continuous deepening of the smart water conservancy system, reservoir hub projects are undertaking increasingly heavy tasks in regional flood control, disaster reduction and water resource allocation. Building a digital twin "four-prevention" system covering forecasting, early warning, rehearsal and contingency plan, and promoting the transformation of reservoir scheduling from traditional experience management to proactive digital prevention and control, has become an urgent need to ensure regional people's livelihood and safety and improve flood season disaster prevention capabilities.

[0003] In existing reservoir flood control and dispatch schemes, a combination of static data integration and conventional hydrological simulation is typically employed. First, scattered raw monitoring data, including meteorological, rainfall, and hydrological data, are collected from different monitoring stations and centrally stored in a relational database. Then, this archived data serves as an offline input source, driving a standard hydrodynamic simulation model to calculate the expected flood evolution trend and water level changes. Finally, based on the calculated water level thresholds, a pre-defined static dispatch rule table is matched, and the corresponding gate opening adjustment parameters are output to execute flood discharge dispatch.

[0004] However, this traditional simulation-based scheduling scheme has significant technical shortcomings. Due to the lack of strict spatiotemporal alignment for multi-source heterogeneous data, the raw monitoring data is difficult to seamlessly integrate into a dynamic base with unified spatial topological associations, resulting in lag and spatial mapping fragmentation during data utilization. At the same time, the flood prediction stage is disconnected from the generation of static rules for mechanical scheduling, lacking a unified coding and group iterative optimization mechanism for multi-gate linkage operations under complex safety constraints. This leads to low adaptability of the generated scheduling strategy to sudden flood fluctuations, making it impossible to achieve dynamic optimization closed-loop control for physical reservoirs. Summary of the Invention

[0005] This application provides a digital twin reservoir flood control scheduling method to at least alleviate the above-mentioned technical problems.

[0006] A digital twin reservoir flood control scheduling method includes:

[0007] Acquire multi-source reservoir monitoring data and reservoir spatiotemporal grids; Based on the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid, the multi-source reservoir monitoring data are spatially topologically bound to generate a digital twin reservoir base plate. Based on the digital twin reservoir bottom plate, time-series grid data is generated, and a deep time-series neural network is used to perform time series prediction on the time-series grid data to obtain the predicted water level sequence and the predicted inflow sequence. The predicted water level sequence and the predicted inflow sequence are input into the scheduling optimization model. Based on the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rules in the scheduling optimization model, candidate scheduling coding is performed on the preset reservoir gate opening range and scheduling time unit to obtain an initial group composed of multiple candidate scheduling sequence codes. The initial group is then subjected to group iterative optimization to obtain the target scheduling sequence. The target scheduling sequence is converted into physical reservoir control instructions, and the physical reservoir control terminal is controlled to perform flood control scheduling actions according to the physical reservoir control instructions.

[0008] The technical advantages of the technical solution provided in this application are: This application's digital twin reservoir flood control scheduling method addresses the technical shortcomings of traditional flood control scheduling schemes, such as the lack of spatiotemporal alignment of multi-source data and the difficulty in fusing monitoring data to form a dynamic baseline. By spatially binding multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules within the reservoir's spatiotemporal grid, a digital twin reservoir baseline is generated. This solves the problems of data lag and spatial fragmentation inherent in traditional schemes. Compared to the traditional approach of relying on relational databases for offline static aggregation of single data acquisition and processing, this application's spatial topology binding mechanism enables seamless mapping and alignment of dispersed multi-source monitoring data under a unified three-dimensional spatial scale and time reference. This ensures a global perspective for data fusion, resulting in significantly improved spatiotemporal correlation and feature extraction completeness in the subsequently generated time-series grid data, laying a reliable data foundation for accurate water level and flow prediction.

[0009] Furthermore, addressing the technical shortcomings of traditional flood control scheduling schemes—such as the disconnect between the prediction stage and the static rule-based scheduling generation, making it difficult to achieve dynamic swarm intelligence optimization closed-loop control under complex constraints—this application inputs the predicted water level sequence and predicted inflow sequence generated by a deep temporal neural network into the scheduling optimization model. Based on the maximum water level constraint, the discharge flow amplitude constraint, and optimization rules, candidate scheduling codes are generated for the preset reservoir gate opening range and scheduling time units. Through swarm iterative optimization of the initial group, the target scheduling sequence is finally generated. Compared to the traditional scheduling generation mode that presets static threshold rules and mechanically outputs operating parameters, this application, under the premise of satisfying multiple rigid constraints such as flood discharge safety and flood control limit water levels, utilizes a swarm iterative optimization mechanism to dynamically search for the optimal spatiotemporal operation combination of gates. This makes the generated scheduling sequence more closely aligned with the real-time safety requirements and discharge stability of reservoir flood control, significantly improving the decision-making adaptability to complex flood emergencies and the safety of actual physical reservoir control, truly achieving a closed-loop linkage of "perception-prediction-optimization-control." Attached Figure Description

[0010] Figure 1 This application provides an embodiment of a digital twin reservoir flood control scheduling scenario. Figure 2 This application provides an embodiment of a digital twin reservoir flood control scheduling method. Figure 3 This application provides an embodiment of a digital twin reservoir flood control scheduling device. Figure 4 An electronic device is described in an embodiment of this application; Figure 5 This is a computer-readable storage medium according to an embodiment of the present application. Detailed Implementation

[0011] like Figure 1 The image shows a digital twin reservoir flood control scheduling scenario according to an embodiment of this application; as shown... Figure 2The image shows an embodiment of a digital twin reservoir flood control scheduling method, which includes the following steps: acquiring multi-source reservoir monitoring data and a reservoir spatiotemporal grid; performing spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate; generating time-series grid data based on the digital twin reservoir base plate, and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence; and combining the predicted water level sequence with the... The predicted inflow sequence is input into the scheduling optimization model. Based on the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the discharge flow amplitude constraint, and the scheduling sequence optimization rules in the scheduling optimization model, candidate scheduling encoding processing is performed on the preset reservoir gate opening range and scheduling time unit to obtain an initial group composed of multiple candidate scheduling sequence codes. The initial group is then subjected to group iterative optimization processing to obtain the target scheduling sequence. The target scheduling sequence is converted into physical reservoir control commands, and the physical reservoir control terminal is controlled to execute flood control scheduling actions according to the physical reservoir control commands.

[0012] Optionally, before the step of spatially binding the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate, the method further includes: acquiring historical reservoir documents; using a data extraction engine to extract data from the historical reservoir documents to obtain feature data units; performing normalization and cleaning on the feature data units to obtain standard structured data; constructing a reservoir data spatial mapping relationship between the standard structured data and the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid, and using the reservoir data spatial mapping relationship as the binding basis for the spatial topology binding rules.

[0013] Preferably, before performing spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid, historical reservoir documents corresponding to the target reservoir are first obtained; the historical reservoir documents include reservoir engineering status survey data, hydrogeological data, annual flood control scheduling data, annual water supply scheduling data, engineering operation and maintenance records, meteorological monitoring reference data, current scheduling plan text, reservoir area topographic map, core area vector boundary map, key section measurement map, reservoir area aerial imagery results, river channel digital elevation model data, river channel digital orthophoto imagery data, and engineering structure data corresponding to the reservoir dam, spillway, water conveyance pipeline, water intake, and gate. After obtaining the historical reservoir documents, they are cataloged according to document source, document type, document time, engineering object, and spatial coverage to form a historical reservoir document catalog record. The historical reservoir document catalog record includes the document source, document type, document time, engineering object, and spatial coverage. The historical reservoir document catalog record is then input into the data extraction engine, enabling the data extraction engine to determine the data extraction method corresponding to each historical reservoir document based on the historical reservoir document catalog record, and to ensure that the data extraction method corresponds to the document type and engineering object of the historical reservoir document, thereby avoiding the processing of hydrological data, engineering structure data, and spatial basic data using the same extraction method.

[0014] Preferably, the data extraction engine includes a hierarchical field reading rule for XML files, a column field reading rule for comma-separated value files, and a cell area reading rule for spreadsheet files; after reading the historical reservoir document directory record, the data extraction engine first determines the file organization method of the historical reservoir document, and then extracts data from the historical reservoir document according to the reading rule corresponding to the file organization method. For the reservoir flood control indicators recorded in tabular form, the data extraction engine extracts the crest elevation,校核洪水位, design flood level, normal high water level, dead water level, total reservoir capacity, flood control reservoir capacity, historical highest reservoir level, historical maximum inflow, historical maximum outflow, low water level alarm value, and start forecast flow standard according to the corresponding relationship between row and column headings; for the reservoir area topographic map, core area vector boundary map, key section measurement map, reservoir area aerial flight image results, river digital elevation model data, and river digital orthophoto image data recorded in the form of maps or spatial files, the data extraction engine extracts the spatial range, coordinate reference, elevation datum, spatial resolution, measurement time, and spatial object type; for the historical reservoir documents recorded in the form of dispatching plan texts or engineering operation and maintenance records, the data extraction engine extracts the dispatching period, dispatching object, gate object, water level condition, flow condition, and execution record. The data extracted by the data extraction engine as described above is used as the historical reservoir document extraction data, and the historical reservoir document extraction data is combined according to the source document, field meaning, data type, measurement unit, time mark, and spatial mark to form the characteristic data unit.

[0015] Note: "校核洪水位" in the original text is not translated as there is no English equivalent provided. You may need to provide the appropriate English term for it to complete the translation accurately.Preferably, the feature data unit is not a single numerical value, but a computable data fragment formed for the reservoir flood control scheduling scenario; each feature data unit includes at least a data name, data meaning, data source, data type, unit of measurement, time stamp, spatial stamp, engineering object stamp, and scheduling purpose stamp. The data name is used to record the water level, reservoir capacity, flow rate, gate, cross-section, river channel, or reservoir area spatial object corresponding to the feature data unit; the data meaning is used to record the physical meaning of the feature data unit in reservoir flood control scheduling; the data source is used to record that the feature data unit originates from reservoir engineering status survey data, historical flood control scheduling data, reservoir area topographic maps, or other historical reservoir documents; the data type is used to record the data format and executable data processing method corresponding to the feature data unit; the unit of measurement is used to record the water level unit, reservoir capacity unit, flow rate unit, or spatial scale unit corresponding to the feature data unit; the time stamp is used to express the observation time, occurrence time, or document formation time corresponding to the feature data unit; the spatial stamp is used to express the reservoir location, gate location, cross-section location, river channel location, or reservoir area range corresponding to the feature data unit; the engineering object stamp is used to express the reservoir dam, spillway, water conveyance pipeline, water intake, or gate corresponding to the feature data unit; the scheduling purpose stamp is used to express the purpose of the feature data unit in constructing reservoir data spatial mapping relationships, generating spatial topology binding rules, or subsequent flood control scheduling calculations. The data name, data meaning, data source, data type, unit of measurement, time stamp, spatial stamp, engineering object stamp, and scheduling purpose stamp together constitute the field structure of the feature data unit, enabling the feature data unit to continue to undergo normalization and cleaning.

[0016] Preferably, after obtaining the feature data units, the feature data units are standardized and cleaned to generate the standard structured data. The standardized cleaning includes uniqueness processing, integrity processing, legality processing, and source priority processing; wherein, the uniqueness processing identifies duplicate feature data units according to reservoir code, station code, gate number, spatial marker, and time marker, and merges duplicate feature data units while retaining traceable sources; the integrity processing reads the missing fields in the feature data units and fills in the identifiable data fields according to historical reservoir documents and multi-source reservoir monitoring data under the same project object; the legality processing verifies the feature data units according to field type, unit of measurement, water level range, flow range, time format, and spatial coordinate range, and marks feature data units that do not meet the verification conditions as problematic data units; the source priority processing, for data with the same name appearing in different historical reservoir documents for the same project object, reads the data source, document time, field integrity status, and problematic data unit mark in the feature data unit, and generates a priority source mark according to the preset source order among engineering design data, operation and maintenance records, monitoring records, and dispatch plan text. After the above-mentioned standardized cleaning, the feature data units are organized into the standard structured data. The standard structured data continues to retain the data source, the time stamp, the spatial stamp, the engineering object stamp, the scheduling purpose stamp, the problem data unit, and the priority source stamp, so as to subsequently construct the spatial mapping relationship of the reservoir data.

[0017] Preferably, when performing standardized cleaning on the feature data units, data conversion processing is also performed on the feature data units; the data conversion processing includes unified encoding, unified identification, and unified format. Unified encoding converts reservoir type codes, engineering object codes, station codes, and gate codes from different sources into unified codes that can be recognized by the reservoir spatiotemporal grid; unified identification merges different names of the same reservoir dam, spillway, gate, cross-section, or station in different historical reservoir documents into the same engineering object tag; unified format converts date records, time records, coordinate records, elevation records, water level records, reservoir capacity records, and flow records into a unified data format. The standard structured data obtained through the data transformation process has consistent data fields, unified units of measurement, and traceable engineering object markers. These consistent data fields, unified units of measurement, and traceable engineering object markers continue to participate in the data comparison and association between the standard structured data and the multi-source reservoir monitoring data, enabling the standard structured data to be compared and associated with the multi-source reservoir monitoring data under the same data semantics, rather than simply being used as offline archival data storage.

[0018] Preferably, before constructing the spatial mapping relationship of the reservoir data, the standard structured data is first subjected to spatial reference normalization processing according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid. The spatial reference normalization processing uses the grid time reference in the reservoir spatiotemporal grid as the unified reference for time recording, the plane coordinate reference corresponding to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid as the unified reference for plane position, and the elevation coordinate reference corresponding to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid as the unified reference for elevation position. When the standard structured data comes from reservoir area topographic maps, core area vector boundary maps, key section measurement maps, reservoir area aerial imagery results, river channel digital elevation model data, or river channel digital orthophoto imagery data, the spatial markers in the standard structured data are converted to the three-dimensional spatial coordinate system. When the standard structured data comes from water level, reservoir capacity, flow rate, gate, or dispatch plan text, the engineering object markers in the standard structured data are associated with the reservoir dam, spillway, intake, gate, river channel section, or reservoir area in the three-dimensional spatial coordinate system. The standard structured data, after undergoing the spatial reference normalization process, includes a unified spatial marker and a unified temporal marker. The unified spatial marker and the unified temporal marker continue to participate in the spatial and temporal correspondence between the standard structured data and the multi-source reservoir monitoring data, so that the standard structured data can establish a computable correspondence with the multi-source reservoir monitoring data in terms of spatial and temporal location.

[0019] Preferably, the spatial mapping relationship of the reservoir data is constructed based on the object correspondence, time correspondence, and spatial correspondence between the standard structured data and the multi-source reservoir monitoring data. During construction, firstly, the target reservoir object is determined according to the engineering object markers in the standard structured data. The target reservoir object includes the reservoir dam, spillway, water pipeline, intake, gate object, monitoring station object, river cross-section, reservoir area, and upstream catchment area. Then, the target reservoir object is located to a spatial grid cell in the three-dimensional spatial coordinate system according to the spatial markers in the standard structured data. Subsequently, based on the time markers in the standard structured data and the acquisition time of the multi-source reservoir monitoring data, the historical water level, historical flow, historical reservoir capacity, gate status, rainfall records, and real-time monitoring records corresponding to the target reservoir object are aligned to the grid time reference. Finally, the standard structured data and the multi-source reservoir monitoring data under the same spatial grid cell, the same grid time reference, and the same target reservoir object are associated to obtain the spatial mapping relationship of the reservoir data. The reservoir data spatial mapping relationship expresses the correspondence between the engineering attributes, spatial attributes, and scheduling attributes in historical reservoir documents and the multi-source reservoir monitoring data. The reservoir data spatial mapping relationship is subsequently used as the binding basis for the spatial topology binding rules.

[0020] Preferably, when constructing the spatial mapping relationship of the reservoir data, data association and comparison processing are also performed on the standard structured data and the multi-source reservoir monitoring data. The data association uses engineering object markers, spatial markers, and time markers as association keys to associate the reservoir flood control indicators, gate engineering parameters, cross-sectional measurement data, and river spatial data in the standard structured data with the rainfall data, water level data, flow data, and gate feedback data in the multi-source reservoir monitoring data. The data comparison performs field content comparison, field format comparison, and spatial location comparison on the associated standard structured data and the multi-source reservoir monitoring data to identify whether there are data with inconsistent time, coordinates, units, or engineering object markers under the same target reservoir object. After the data association and data comparison processing, the reservoir data spatial mapping relationship includes the priority source marker, association status, comparison status, and problem data unit marker. The problem data unit marker is generated based on the problem data unit and is used to restrict conflicting data from directly participating in the generation of the spatial topology binding rules. The priority source marker is used to determine the data source that can participate in the generation of the spatial topology binding rules when there is data from multiple sources.

[0021] Preferably, the spatial topology binding rules are generated based on the spatial mapping relationship of the reservoir data and are used to limit the binding object, binding location, binding time, and binding method when the multi-source reservoir monitoring data enters the digital twin reservoir bottom plate. Specifically, the spatial topology binding rules read the target reservoir object, spatial grid unit, grid time reference, and priority source marker in the spatial mapping relationship of the reservoir data, bind the rainfall data to the corresponding reservoir area and upstream catchment area, bind the water level data to the corresponding water level station and reservoir surface grid, bind the inflow data to the corresponding inflow river section, bind the outflow data to the corresponding spillway, water conveyance pipeline, or gate, and bind the gate feedback data to the corresponding gate object. The reservoir area and upstream catchment area, the water level monitoring station, the reservoir water surface grid, the inflow river section, the spillway, the water conveyance pipeline, the gate, and the gate object are all spatial representation objects of the target reservoir object in the three-dimensional spatial coordinate system. Through the above binding method, the spatial topology binding rules no longer rely solely on the acquisition fields of the monitoring data itself, but use the spatial mapping relationship of the reservoir data as the binding basis, and use historical engineering attributes, spatial location relationships, and real-time monitoring records together for the generation of the digital twin reservoir base plate.

[0022] Preferably, the spatial topology binding rules further include topology relationship checking rules, which verify the connection relationships of spatial points, spatial lines, and spatial surfaces based on the spatial mapping relationship of the reservoir data. For water level stations, rainfall stations, and gate objects, the topology relationship checking rules verify whether the corresponding spatial points fall within the reservoir area, river cross-section, or engineering structure area of ​​the target reservoir object; for river centerlines, water conveyance pipelines, and flood discharge channels, the topology relationship checking rules verify whether the corresponding spatial lines maintain continuous connection with the reservoir dam, spillway, water intake, or downstream river channel; for the reservoir area, upstream catchment area, and reservoir water surface area, the topology relationship checking rules verify whether the corresponding spatial surfaces maintain a covering relationship with the spatial grid cells in the three-dimensional spatial coordinate system. The water level monitoring station, the rainfall monitoring station, the gate object, the river centerline, the water conveyance pipeline, the flood discharge channel, the reservoir area, the catchment area, and the water surface area all participate in the verification of the topology relationship checking rules as spatial topology components of the target reservoir object. The spatial topology binding rules after verification by the topology relationship checking rules can exclude the reservoir data spatial mapping relationship with topological errors from the binding basis, and use the verified reservoir data spatial mapping relationship as the binding basis for generating the digital twin reservoir base plate from the multi-source reservoir monitoring data.

[0023] Preferably, the spatial mapping relationship between the standard structured data and the multi-source reservoir monitoring data can further incorporate data identification processing to enhance the spatial topology binding rules' ability to identify target reservoir objects. The data identification processing adds basic tags to the reservoir dam, spillway, intake, gate, monitoring station, river section, reservoir area, and upstream catchment area in the standard structured data according to the tagging rules in the reservoir flood control scheduling scenario. It also adds business tags based on the association status between the standard structured data and the multi-source reservoir monitoring data. The basic tags express the engineering and spatial categories of the target reservoir object, while the business tags express the target reservoir object's purpose in flood control scheduling, such as water level monitoring, flow monitoring, flood discharge control, or scheduling verification. The basic tags and business tags are written into the reservoir data spatial mapping relationship, enabling the spatial topology binding rules to determine the spatial grid cells and target reservoir objects where data enters the digital twin reservoir base plate when binding the multi-source reservoir monitoring data, based on the basic tags and business tags.

[0024] Preferably, when the spatial mapping relationship of the reservoir data is used as the binding basis for the spatial topology binding rule, it is specifically read in the form of mapping records. Each mapping record includes at least the target reservoir object, the spatial grid cell corresponding to the target reservoir object, the grid time reference corresponding to the target reservoir object, the standard structured data field corresponding to the target reservoir object, the multi-source reservoir monitoring data field corresponding to the target reservoir object, the priority source marker, the association status, the comparison status, and the topology check status. When performing spatial topology binding, firstly, the target reservoir object corresponding to the multi-source reservoir monitoring data is determined according to the target reservoir object in the mapping record; then, the spatial position of the multi-source reservoir monitoring data entering the digital twin reservoir base plate is determined according to the spatial grid cell in the mapping record; subsequently, the time position of the multi-source reservoir monitoring data entering the digital twin reservoir base plate is determined according to the grid time reference in the mapping record; finally, the multi-source reservoir monitoring data that can participate in the binding is selected according to the topology check status in the mapping record. Therefore, the spatial topology binding rule uses the spatial mapping relationship of the reservoir data as the binding basis, enabling the digital twin reservoir base plate to simultaneously bear historical engineering attributes, unified spatial location, and real-time monitoring status.

[0025] Optionally, the step of performing spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate includes: obtaining the quadtree structure in the three-dimensional basic platform of the digital twin reservoir; performing coordinate normalization processing on the spatial nodes in the quadtree structure according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid; and constructing a hierarchical detail pyramid model according to the coordinate-normalized quadtree structure; determining the resolution level according to the observation point distance parameter; and according to the resolution level, The grid time reference and spatial topology binding rules in the reservoir spatiotemporal grid organize the multi-source reservoir monitoring data into a bottom-level tile matrix with time index and spatial binding position. The bottom-level tile matrix with time index and spatial binding position is mapped to the resolution level corresponding to the hierarchical detail pyramid model, so that the multi-source reservoir monitoring data forms a spatial topology binding result under the joint constraints of the three-dimensional spatial coordinate system, the grid time reference and the spatial topology binding rules, so as to generate the digital twin reservoir base plate carrying the spatial topology binding result.

[0026] Preferably, in the specific implementation process of generating a digital twin reservoir base plate by spatially binding the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid, the quadtree structure in the three-dimensional basic platform of the digital twin reservoir is first obtained; the three-dimensional basic platform of the digital twin reservoir is used to carry the three-dimensional terrain of the reservoir area, image tiles, vector data, building information model data, river digital elevation data, river digital orthophoto data, and the three-dimensional engineering objects corresponding to the reservoir dam, spillway, water pipeline, water intake, and gate. The quadtree structure is used to recursively divide the spatial range of the reservoir area in the three-dimensional basic platform of the digital twin reservoir. The reason why the quadtree structure adopts a "four" branch division is that the spatial range of the reservoir area in the spatiotemporal grid has a horizontal coordinate direction and a vertical coordinate direction in the three-dimensional spatial coordinate system. After dividing into two parts along the horizontal coordinate direction and the vertical coordinate direction respectively, each recursive spatial division forms a first sub-region formed by the intersection of the first horizontal sub-region and the first vertical sub-region, a second sub-region formed by the intersection of the second horizontal sub-region and the first vertical sub-region, a third sub-region formed by the intersection of the first horizontal sub-region and the second vertical sub-region, and a fourth sub-region formed by the intersection of the second horizontal sub-region and the second vertical sub-region. The first sub-region, the second sub-region, the third sub-region, and the fourth sub-region correspond to the four child nodes under the same parent node in the quadtree structure. The root node of the quadtree structure corresponds to the overall planar spatial range of the reservoir area covered by the reservoir spatiotemporal grid. The intermediate nodes of the quadtree structure correspond to the reservoir sub-ranges after hierarchical division. The leaf nodes of the quadtree structure correspond to spatial nodes that can directly carry the multi-source reservoir monitoring data. Therefore, the quadtree structure is not simply a graphical display structure, but a spatial index structure that transforms the reservoir planar spatial range, river channel range, gate objects, and monitoring station objects into addressable spatial nodes. The addressable spatial nodes are used to provide spatial writing locations for the multi-source reservoir monitoring data in subsequent processing and to provide a spatial division basis for the resolution level of the hierarchical detail pyramid model.

[0027] Preferably, after obtaining the quadtree structure, the spatial nodes in the quadtree structure are normalized according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid. The coordinate normalization process first reads the node plane boundary, elevation range, and node level corresponding to each spatial node, then maps the node plane boundary to the plane coordinate components in the three-dimensional spatial coordinate system, maps the elevation range to the elevation coordinate components in the three-dimensional spatial coordinate system, and maintains the correspondence between the node level and the parent and child nodes in the quadtree structure. Since each parent node in the quadtree structure is divided into four child nodes according to the horizontal and vertical coordinate directions, the coordinate normalization process can convert the node plane boundaries corresponding to each of the four child nodes into non-overlapping normalized spatial ranges that can be pieced together to restore the node plane boundaries corresponding to the parent node, thus ensuring that the quadtree structure maintains a continuous reservoir spatial coverage relationship after recursive partitioning. The quadtree structure, after coordinate normalization, enables the 3D topography of the reservoir area, the image tiles, the vector data, and the building information model data to be expressed in the same 3D spatial coordinate system. This avoids the same gate object, the same river cross-section, or the same station object falling into different spatial nodes due to different data sources. The quadtree structure after coordinate normalization continues to serve as the spatial index skeleton of the hierarchical detail pyramid model, ensuring that each resolution level subsequently constructed has planar coordinate components, elevation coordinate components, and node levels corresponding to the spatial nodes.

[0028] Preferably, the quadtree structure after coordinate normalization is further used to construct the hierarchical detail pyramid model. In the hierarchical detail pyramid model, "hierarchy" refers to the multi-level spatial representation structure formed by arranging different node levels in the quadtree structure according to their spatial coverage from largest to smallest. "Detail" in the hierarchical detail pyramid model refers to the different spatial granularities expressed by different resolution levels for the same reservoir area. Coarser resolution levels are used to express large-scale objects such as the reservoir area, upstream catchment area, and downstream river channel area, while finer resolution levels are used to express local objects such as reservoir dams, spillways, intakes, gates, monitoring stations, and river cross-sections. In the hierarchical detail pyramid model, "pyramid" means that coarser resolution levels contain a smaller number of spatial nodes with a larger coverage area, while finer resolution levels contain a larger number of spatial nodes with a smaller coverage area, thus forming a spatial representation relationship that expands hierarchically from a general upper level to a finer lower level. Each resolution level reads spatial nodes of the corresponding node level from the quadtree structure, allowing the hierarchical detail pyramid model to utilize the spatial index structure of the quadtree. Since the quadtree structure divides the parent node into four child nodes at each node level, the hierarchical detail pyramid model can form a spatial inheritance relationship between adjacent resolution levels. That is, one spatial node in the previous resolution level corresponds to four spatial nodes in the next resolution level. This ensures that when the spatial range of the reservoir area gradually transitions from coarse-grained to fine-grained expression, the spatial continuity of the same reservoir area, the same river channel, and the same gate object is maintained. The technical function of the hierarchical detail pyramid model is to decompose the same digital twin reservoir base plate into multiple spatial expression levels that can be read on demand. This allows the multi-source reservoir monitoring data to be bound not only to the spatial range of the reservoir area but also to local objects such as the gate object, the river cross-section, and the monitoring station.

[0029] Preferably, each resolution level in the hierarchical detail pyramid model includes a resolution level identifier, a node level index, a spatial node list, and a parent-child node mapping relationship. The resolution level identifier distinguishes different spatial representation levels from coarse to fine in the hierarchical detail pyramid model. The node level index records the quadtree structure node level corresponding to the resolution level identifier. The spatial node list records the spatial nodes in that resolution level that can receive the multi-source reservoir monitoring data. The parent-child node mapping relationship records the correspondence between spatial nodes in the previous resolution level and four spatial nodes in the next resolution level. The resolution level identifier, the node level index, the spatial node list, and the parent-child node mapping relationship together constitute the hierarchical description data of the hierarchical detail pyramid model. This hierarchical description data is subsequently used to determine which resolution level, which spatial node, and which time index position should be written to the elements in the bottom tile matrix. Therefore, the hierarchical detail pyramid model is not simply a three-dimensional display layer, but a data organization structure used to support the mapping relationship between the multi-source reservoir monitoring data, the spatial nodes, and the grid time reference.

[0030] Preferably, when determining the resolution level based on the observation point distance parameter, the observation point position and observation direction are first read from the digital twin reservoir 3D base platform. The observation point position is used to express the spatial position of the current 3D scene display viewpoint in the 3D spatial coordinate system, and the observation direction is used to express the direction of the current 3D scene display viewpoint toward the reservoir area planar space range, the gate object, or the river channel cross-section. The observation point distance parameter is generated based on the spatial interval between the observation point position and the spatial node, and the corresponding resolution level is selected in the hierarchical detail pyramid model based on the observation point distance parameter. When the observation point distance parameter corresponds to a relatively distant 3D scene display viewpoint, a coarser resolution level that can express the reservoir area planar space range and the river channel range is selected; when the observation point distance parameter corresponds to a relatively close 3D scene display viewpoint, a finer resolution level that can express the gate object, the station object, and the river channel cross-section is selected. Since the four child nodes in the quadtree structure respectively cover the four spatial orientations of the node plane boundary corresponding to the parent node, the observation point distance parameter and the observation direction can jointly determine the spatial nodes mainly covered by the current 3D scene display viewpoint, and make the resolution level prioritize reading the spatial nodes corresponding to the current spatial orientation, without having to load all spatial nodes unrelated to the current 3D scene display viewpoint at the same time. After this processing, the determination process of the resolution level is directly technically linked to the observation point distance parameter, the spatial nodes, and the hierarchical detail pyramid model, rather than fixing the multi-source reservoir monitoring data into a single spatial precision graphic base map.

[0031] Preferably, after determining the resolution level, the multi-source reservoir monitoring data is organized into a bottom-level tile matrix with time index and spatial binding position according to the resolution level, the grid time reference in the reservoir spatiotemporal grid, and the spatial topology binding rules in the reservoir spatiotemporal grid. The bottom-level tile matrix is ​​called "bottom-level" because it is located at the starting point of data writing in the hierarchical detail pyramid model, used to receive the original bound data that has not yet been written to each resolution level, and serves as the basic data organization for subsequent hierarchical mapping to the hierarchical detail pyramid model. The bottom-level tile matrix is ​​called "tile" because each row corresponds to a spatial node in the quadtree structure, and each spatial node corresponds to an independently addressable spatial tile within the reservoir area's planar spatial range. The bottom-level tile matrix is ​​called "matrix" because it forms a two-dimensional arrangement based on spatial nodes and scheduling time units. Specifically, the row direction of the bottom-layer tile matrix corresponds to the spatial nodes in the quadtree structure, and the column direction of the bottom-layer tile matrix corresponds to the scheduling time unit under the grid time reference. The element at the intersection of each row and column in the bottom-layer tile matrix is ​​used to store the multi-source reservoir monitoring data under the corresponding spatial node and the corresponding scheduling time unit. Among them, rainfall data is written to the spatial node corresponding to the plane spatial range of the reservoir area or the upstream catchment area, water level data is written to the spatial node corresponding to the monitoring station or the water surface range of the reservoir area, inflow data is written to the spatial node corresponding to the inflow river section, and outflow data and gate feedback data are written to the spatial node corresponding to the spillway, the water conveyance pipeline or the gate. Since the spatial nodes of the quadtree structure originate from the recursive partitioning of the four child nodes, the row direction in the bottom-level tile matrix can be organized according to the spatial order from the parent node to the four child nodes, and then from the four child nodes to the next level child node. This ensures that adjacent rainfall data, water level data, inflow data, outflow data, and gate feedback data within the same reservoir sub-area maintain adjacent or traceable row and column intersection relationships in the bottom-level tile matrix. Through this organization method, each element in the bottom-level tile matrix simultaneously carries the time index and the spatial binding position, enabling the multi-source reservoir monitoring data to enter subsequent mapping processing under the joint constraints of the grid time reference and the spatial topology binding rules.

[0032] Preferably, each element in the underlying tile matrix includes a data payload, a time index, a spatial binding location, a target reservoir object, a data type marker, and a binding status marker. The data payload is used to store rainfall data, water level data, inflow data, outflow data, or gate feedback data. The time index is used to record the scheduling time unit corresponding to the data payload. The spatial binding location is used to record the spatial node corresponding to the data payload. The target reservoir object is used to record the reservoir area planar spatial range, upstream catchment area, monitoring station object, inflow river cross section, spillway, water conveyance pipeline, or gate object corresponding to the data payload. The data type marker is used to distinguish the data type to which the data payload belongs. The binding status marker is used to record whether the data payload satisfies the spatial topology binding rules. The data payload, the time index, the spatial binding location, the target reservoir object, the data type tag, and the binding status tag together constitute the element description structure of each element in the underlying tile matrix. The element description structure subsequently participates in the process of mapping the underlying tile matrix to the resolution level corresponding to the hierarchical detail pyramid model, so that each element retains its time source, spatial source, and object source when written to the hierarchical detail pyramid model.

[0033] Preferably, when the spatial topology binding rule participates in generating the underlying tile matrix, it reads the spatial mapping relationship of the previously generated reservoir data and determines the target reservoir object, spatial grid unit, and grid time reference corresponding to the multi-source reservoir monitoring data based on the spatial mapping relationship of the reservoir data. For multiple data sources of the same target reservoir object, the spatial topology binding rule prioritizes reading data records with priority source markers, association status, and topology relationship check status, and writes the data records that pass the topology relationship check into the corresponding row and column intersection positions in the underlying tile matrix; for data records with spatial location conflicts, temporal location conflicts, or engineering object marker conflicts, the spatial topology binding rule retains the data record in the data position corresponding to the problem data unit marker, and associates the problem data unit marker with the spatial node and scheduling time unit corresponding to the data record, instead of directly writing it into the row and column intersection positions in the underlying tile matrix that can participate in the generation of the digital twin reservoir base plate. Because the quadtree structure divides the node plane boundary corresponding to the parent node into four child nodes, the spatial topology binding rule can determine which of the four child nodes the data record should fall into based on the spatial location of the data record. When the data record falls near the boundary of a child node, the spatial topology binding rule is used to further verify the spatial dependency relationship between the data record and the reservoir dam, spillway, water pipeline, water intake, gate object, monitoring station object, or river cross-section. Therefore, the bottom-layer tile matrix not only expresses the row and column intersection positions of the multi-source reservoir monitoring data but also whether the multi-source reservoir monitoring data meets the binding conditions of the spatial topology binding rule.

[0034] Preferably, when mapping the bottom-level tile matrix with time index and spatial binding position to the resolution level corresponding to the hierarchical detail pyramid model, firstly, the spatial node corresponding to each row in the bottom-level tile matrix is ​​read, and the resolution level corresponding to the spatial node is determined according to the node level of the spatial node in the quadtree structure; then, the scheduling time unit corresponding to each column in the bottom-level tile matrix is ​​read, and the time index position of the multi-source reservoir monitoring data in the resolution level is determined according to the correspondence between the scheduling time unit and the grid time reference; then, the target reservoir object and binding status mark of each element in the bottom-level tile matrix are read, and whether the element can be written to the resolution level is determined according to the target reservoir object and the binding status mark; finally, the elements in the bottom-level tile matrix are written to the corresponding resolution level, corresponding spatial node and corresponding time index position in the hierarchical detail pyramid model to form a spatial topology binding result. The phrase "mapping to the resolution level corresponding to the hierarchical detail pyramid model" specifically refers to: first, establishing the node correspondence between the row spatial nodes in the bottom-level tile matrix and the spatial nodes in the hierarchical detail pyramid model; then, establishing the time correspondence between the column scheduling time units in the bottom-level tile matrix and the time index positions in the hierarchical detail pyramid model; subsequently, writing the elements that satisfy the spatial topology binding rules into the corresponding resolution level in the hierarchical detail pyramid model according to the node correspondence and the time correspondence. Since there is a node hierarchy correspondence between the parent node and the four child nodes in the quadtree structure, when elements in the bottom-level tile matrix are mapped to the hierarchical detail pyramid model, they can either converge upwards or subdivide downwards along the node hierarchy correspondence between the parent node and the four child nodes. Converging upwards is used to express the overall state of the reservoir area's planar spatial range at a coarser resolution level, while subdividing downwards is used to express the local state of the gate object, the station object, and the river cross-section at a finer resolution level. The spatial topology binding result preserves the correspondence between the multi-source reservoir monitoring data, the spatial nodes, the resolution levels, the scheduling time units, the target reservoir objects, and the binding status markers, enabling the continuous monitoring status on the same spatial nodes to be read directly along the grid time reference when generating subsequent time-series grid data.

[0035] Preferably, during upward aggregation, the underlying tile matrix elements corresponding to the four child nodes under the same parent node are read, and the rainfall data, water level data, inflow data, outflow data, and gate feedback data are aggregated according to the data type marker to form the aggregation state element corresponding to the parent node. The aggregation state element continues to carry the spatial node corresponding to the parent node, the scheduling time unit, the target reservoir object, and the binding state marker, and is written to the corresponding position of the coarser resolution level in the hierarchical detail pyramid model. During downward subdivision, the underlying tile matrix elements corresponding to the parent node are read, and according to the spatial coverage relationship between the parent node and the four child nodes, the underlying tile matrix elements corresponding to the parent node are allocated to child nodes that can express local objects. The child nodes continue to carry the time index, the spatial binding position, and the target reservoir object, and are written to the corresponding position of the finer resolution level in the hierarchical detail pyramid model. Through upward aggregation and downward subdivision, the underlying tile matrix can simultaneously form large-scale state representation and local state representation in the hierarchical detail pyramid model.

[0036] Preferably, after the spatial topology binding result is formed in the hierarchical detail pyramid model, spatial continuity verification and temporal continuity verification are performed on the spatial topology binding result. The spatial continuity verification reads rainfall data, water level data, inflow data, outflow data, and gate feedback data from adjacent spatial nodes, and verifies whether there is a broken binding between adjacent spatial nodes according to the parent-child node relationship in the quadtree structure; the temporal continuity verification reads multi-source reservoir monitoring data of the same spatial node in adjacent scheduling time units, and verifies whether there is a time gap in the data on the same spatial node according to the grid time reference. Since the four child nodes under the same parent node in the quadtree structure jointly cover the reservoir sub-range corresponding to the parent node in the reservoir area plane space, the spatial continuity verification can check the boundary connection relationship between the four child nodes respectively, and verify whether the cross-boundary data still corresponds to the same target reservoir object or an adjacent target reservoir object when rainfall data, water level data, inflow data, outflow data, or gate feedback data cross the child node boundary. The spatial topology binding result, after being processed by the spatial continuity verification and the temporal continuity verification, is further written into the digital twin reservoir base plate, enabling the digital twin reservoir base plate to carry multi-source reservoir monitoring data that has undergone spatial node positioning, scheduling time unit alignment, and topology relationship verification.

[0037] Preferably, the digital twin reservoir base plate is jointly generated by the hierarchical detail pyramid model, the bottom layer tile matrix, and the spatial topology binding result. The technical essence of the digital twin reservoir base plate is: under a unified three-dimensional spatial coordinate system and a unified grid time reference, a spatiotemporal data carrying structure is formed by addressably binding the spatial foundation (comprised of the reservoir area's three-dimensional terrain, image tiles, vector data, and building information model data) with the dynamic state data (comprised of multi-source reservoir monitoring data). The spatial foundation expresses the spatial morphology of the target reservoir object, the dynamic state data expresses the operational status of the target reservoir object under different scheduling time units, and the spatial topology binding result expresses the spatial node, time index, and target reservoir object corresponding to the dynamic state data when written into the spatial foundation. The hierarchical detail pyramid model provides a multi-resolution spatial representation structure for the digital twin reservoir base plate. The bottom-layer tile matrix provides a data organization structure with time indexes and spatially bound positions for the digital twin reservoir base plate. The spatial topology binding result provides the binding relationship between the multi-source reservoir monitoring data and the target reservoir object for the digital twin reservoir base plate. When generating the digital twin reservoir base plate, the 3D terrain of the reservoir area, the image tiles, the vector data, and the building information model data are used as the spatial foundation. The multi-source reservoir monitoring data in the bottom-layer tile matrix is ​​used as dynamic state data. The spatial topology binding result is used as the binding basis for the dynamic state data to enter the spatial foundation, thereby generating the digital twin reservoir base plate that carries the spatial topology binding result. Since the quadtree structure recursively expresses the planar spatial range of the reservoir area through the four child nodes, the digital twin reservoir base plate can simultaneously retain the overall expression of the planar spatial range of the reservoir area, the local expression of the local objects, and the binding relationship of the multi-source reservoir monitoring data on the same spatial basis. This allows the digital twin reservoir base plate to continue to participate in the generation of temporal grid data in subsequent processing, enabling the temporal grid data to inherit the spatial nodes, time indices, and spatial topology binding results in the digital twin reservoir base plate.

[0038] Optionally, the step of generating time-series grid data based on the digital twin reservoir base plate and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence includes: organizing the spatial topology binding results in the digital twin reservoir base plate according to the grid time reference to obtain the time-series grid data; extracting rainfall feature vectors and flow feature vectors from the time-series grid data; inputting the rainfall feature vectors and flow feature vectors into the deep temporal neural network containing long short-term memory units; extracting time delay features from the rainfall feature vectors and flow feature vectors through the long short-term memory units to obtain time delay feature expression results; and generating the predicted water level sequence and the predicted inflow sequence based on the time delay feature expression results.

[0039] Preferably, in the specific implementation process of generating time-series grid data based on the digital twin reservoir base plate and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain the predicted water level sequence and the predicted inflow sequence, the spatial topology binding result in the digital twin reservoir base plate is first read. The spatial topology binding result includes spatial nodes, target reservoir objects, time indexes, multi-source reservoir monitoring data, spatial binding locations, and binding status markers. The spatial nodes are used to express the addressable locations of the reservoir area, upstream catchment area, inflow river cross-section, reservoir water surface area, monitoring station objects, spillway, water conveyance pipeline, and gate objects in the three-dimensional spatial coordinate system. The target reservoir object is used to express the reservoir area, upstream catchment area, inflow river cross-section, reservoir water surface area, monitoring station objects, spillway, water conveyance pipeline, or gate object corresponding to the spatial node. The time index is used to express the scheduling time unit to which the acquisition time corresponding to the multi-source reservoir monitoring data belongs under the grid time reference. After reading the spatial topology binding result, the monitoring data of multi-source reservoirs under the same spatial node are organized in chronological order according to the grid time reference to form a spatial node time sequence record. The spatial node time sequence record retains the spatial node, the target reservoir object, the time index, the spatial binding position, and the binding status marker. The spatial node time sequence record continues to be merged according to the binding relationship between the spatial node and the target reservoir object, so that the rainfall data, water level data, inflow data, outflow data, and gate feedback data of the same target reservoir object in a continuous scheduling time unit can be organized into the same time series chain. The time series chain is written to the corresponding spatial node according to the grid time reference to obtain the time series grid data.

[0040] Preferably, the temporal grid data is not a simple time table, but a spatiotemporal joint data structure formed by reorganizing the spatial topology binding results under the grid time reference. The row direction of the temporal grid data corresponds to the spatial node or the target reservoir object formed by multiple spatial nodes, and the column direction of the temporal grid data corresponds to the continuously arranged scheduling time units. The element at the intersection of each row and column in the temporal grid data is used to store the rainfall data, water level data, inflow data, outflow data, gate feedback data, and binding status markers under the corresponding spatial node and the corresponding scheduling time unit. Each element in the temporal grid data inherits the spatial binding position and time index in the digital twin reservoir base plate, so that the temporal grid data can express the continuous change relationship of rainfall data, water level data, and inflow data in the time dimension, and express the topological association relationship between the upstream catchment area, the inflow river section, the reservoir water surface area, the spillway, the water conveyance pipeline, and the gate object in the spatial dimension. Therefore, the temporal grid data differs from traditional offline archived data. Its technical essence is to embed dynamic monitoring status into continuous spatiotemporal samples after spatial nodes. These continuous spatiotemporal samples continue to serve as the basis for extracting rainfall feature vectors and flow feature vectors, enabling subsequent deep temporal neural networks to simultaneously read temporal continuity relationships and spatial topological relationships.

[0041] Preferably, when generating the time-series grid data, time-series interpolation and sampling alignment are also performed on the time-series grid data. The time-series interpolation process reads multi-source reservoir monitoring data of the same spatial node in adjacent scheduling time units, and identifies missing row and column intersection positions according to the binding status markers, so as to form time-series interpolation data using available monitoring data of the same target reservoir object in adjacent scheduling time units. The time-series interpolation data continues to be written into the missing row and column intersection positions in the time-series grid data, so that the missing row and column intersection positions can retain the relationship with the corresponding spatial node and the corresponding scheduling time unit in subsequent feature extraction. The sampling alignment process reads the acquisition time of different data sources, and aligns the rainfall data, water level data, inflow data, outflow data, and gate feedback data to the scheduling time units under the grid time reference, respectively, to form time-aligned time-series grid data. The time-aligned time-series grid data continues to participate in the extraction of rainfall feature vectors and flow feature vectors, so that multi-source reservoir monitoring data formed under different sampling frequencies can enter the same depth time-series neural network for time series prediction according to the same grid time reference.

[0042] Preferably, when extracting rainfall feature vectors from the time-series grid data, the spatial nodes corresponding to the reservoir area, upstream catchment area, and monitoring station are first determined based on the target reservoir object in the spatial topology binding result. Then, the rainfall data of the spatial nodes within the continuous scheduling time unit is read from the time-series grid data to form a rainfall time-series segment. The rainfall time-series segment records the rainfall intensity, cumulative rainfall status, rainfall duration, and spatial location status of the monitoring station in chronological order, and the spatial location status of the monitoring station is associated with the corresponding upstream catchment area or reservoir area based on the spatial topology binding result. Subsequently, the rainfall time-series segment is processed by rainfall lag windowing to form a rainfall feature vector. The elements in the rainfall feature vector respectively express the rainfall intensity, cumulative rainfall status, rainfall duration, and spatial location status of the monitoring station within different scheduling time units. The rainfall feature vector continues to serve as the input of a deep time-series neural network, enabling the deep time-series neural network to extract the delayed impact of rainfall on the predicted water level sequence and the predicted inflow sequence based on the rainfall feature vector.

[0043] Preferably, when extracting flow feature vectors from the time-series grid data, the spatial nodes corresponding to the inflow river cross-section, spillway, water conveyance pipeline, and gate object are first determined according to the target reservoir object in the spatial topology binding result. Then, the inflow flow data, outflow flow data, and gate feedback data of the spatial nodes in the continuous scheduling time unit are read from the time-series grid data to form a flow time-series segment. The flow time-series segment records the inflow flow change status, outflow flow change status, gate feedback change status, and cross-section spatial position status in chronological order, and binds the cross-section spatial position status with the corresponding inflow river cross-section, spillway, water conveyance pipeline, or gate object according to the spatial topology binding result. Subsequently, the flow time series segments are processed by flow change windows to form a flow feature vector. The elements in the flow feature vector represent the inflow flow change status, outflow flow change status, gate feedback change status, and cross-sectional spatial position status within different scheduling time units. The flow feature vector is then input into a deep time series neural network together with the rainfall feature vector, enabling the deep time series neural network to extract the impact of inflow flow, outflow flow, and gate feedback on subsequent predicted water level sequences and predicted inflow flow sequences based on the flow feature vector.

[0044] Preferably, the deep temporal neural network includes a temporal input processing layer, a spatial node embedding layer, a rainfall-discharge coupling layer, long short-term memory units, a prediction output layer, and a sequence processing layer. The temporal input processing layer reads the rainfall feature vector and the discharge feature vector, and organizes them into temporal input samples within the same time window according to the grid time reference. The temporal input samples are then input into the spatial node embedding layer, which reads the spatial binding locations and target reservoir objects in the temporal input samples and transforms them into spatial node embedding expressions capable of participating in temporal calculations. The spatial node embedding expressions are then input into the rainfall-discharge coupling layer, which combines the rainfall state, discharge state, and spatial node embedding expressions from the temporal input samples. The rainfall-flow coupled expression is then fused to form a rainfall-flow coupled expression. This expression is then input into the Long Short-Term Memory (LSTM) unit, which extracts time-delay features to obtain a time-delay feature expression result. This time-delay feature expression result is then input into the prediction output layer, which generates water level prediction segments and inflow prediction segments based on the time-delay feature expression result. The water level prediction segments and inflow prediction segments are then input into the sequence processing layer, which processes the water level prediction segments into a predicted water level sequence and the inflow prediction segments into a predicted inflow sequence according to the grid time reference.

[0045] Preferably, when the time-series input processing layer organizes the rainfall feature vector and the flow feature vector, it reads the rainfall feature vector and the flow feature vector within a continuous scheduling time unit according to the grid time reference, and writes the rainfall feature vector and the flow feature vector within the same scheduling time unit into the time-series input sample in parallel. Each sample unit in the time-series input sample includes the rainfall intensity, cumulative rainfall status, rainfall duration status, inflow flow change status, outflow flow change status, gate feedback change status, spatial binding location, and target reservoir object for the corresponding scheduling time unit. The time-series input sample is input into the spatial node embedding layer in the order of scheduling time units, so that the spatial node embedding layer can simultaneously read the rainfall feature vector, the flow feature vector, the spatial binding location, and the target reservoir object within the same time window, thereby providing a data foundation for the rainfall-flow coupling layer to form a rainfall-flow coupled expression.

[0046] Preferably, when the spatial node embedding layer processes the spatially bound location and the target reservoir object, it first reads the spatial node corresponding to the spatially bound location, then reads the object type and spatial topology relationship corresponding to the target reservoir object, and embeds the spatial node, the object type, and the spatial topology relationship into an embedded expression to form the spatial node embedding expression. The elements in the spatial node embedding expression are used to express the spatial affiliation and topological adjacency relationships of the spatial node within the reservoir area, upstream catchment area, inflow river section, reservoir water surface area, monitoring station object, spillway, water conveyance pipeline, and gate object. The spatial node embedding expression, along with the rainfall feature vector and the flow feature vector, is then input into the rainfall-flow coupling layer, enabling the rainfall-flow coupling layer to retain the spatial meaning of the spatial node containing the data when processing rainfall and flow states.

[0047] Preferably, when the rainfall-flow coupling layer fuses the rainfall state, flow state, and spatial node embedding expression in the time-series input sample, it first reads the rainfall feature vector, the flow feature vector, and the spatial node embedding expression within the same scheduling time unit. Then, according to the spatial topological relationship between the target reservoir objects, it associates the rainfall state corresponding to the upstream catchment area, the inflow flow change state corresponding to the inflow river section, the water level related state corresponding to the reservoir water surface area, and the gate feedback change state corresponding to the gate object to form the rainfall-flow coupling expression. The rainfall-flow coupling expression retains the correspondence between the rainfall state, flow state, and spatial node embedding expression within the same scheduling time unit and continues to be input into the Long Short-Term Memory (LSTM) unit, enabling the LSTM unit to extract the time delay features between the rainfall state and the flow state between consecutive scheduling time units.

[0048] Preferably, the Long Short-Term Memory (LSTM) unit includes an input gate processing structure, a forget gate processing structure, a candidate state processing structure, a state update processing structure, and an output gate processing structure. The input gate processing structure reads the rainfall-flow coupling expression corresponding to the current scheduling time unit and selects an input state related to the current water level change state and the current inflow change state from the rainfall-flow coupling expression. The forget gate processing structure reads the historical memory state formed in the previous scheduling time unit and determines the delay impact information to be retained in the historical memory state based on the rainfall-flow coupling expression of the current scheduling time unit. The candidate state processing structure generates candidate memory states based on the rainfall-flow coupling expression of the current scheduling time unit, and the candidate memory states express the new impact of the current rainfall state and the current flow state on the subsequent predicted water level sequence and the subsequent predicted inflow sequence. The state update processing structure fuses the historical memory state, the input state, and the candidate memory state to form the current memory state. The output gate processing structure generates the time delay feature expression result corresponding to the current scheduling time unit based on the current memory state. The time delay feature expression result is then input into the prediction output layer, enabling the prediction output layer to generate water level prediction segments and inflow prediction segments corresponding to future continuous scheduling time units, respectively.

[0049] Preferably, the Long Short-Term Memory (LSTM) unit is used to handle reservoir flood control scheduling scenarios where there is a time lag between rainfall and water level, and between rainfall and inflow. In these scenarios, after rainfall data enters the upstream catchment area, it is not entirely converted into inflow or water level data within the same scheduling time unit. Instead, it experiences delays due to the confluence process, river transport process, and reservoir storage and release process. The LSTM unit retains the rainfall and flow feature vectors from previous scheduling time units through its historical memory state, expresses the delay relationship between the current scheduling time unit and previous scheduling time units through its current memory state, and provides this delay relationship to the prediction output layer through the time delay feature expression result. Thus, the deep temporal neural network does not perform real-time calculations based solely on the rainfall data of the current scheduling time unit, but generates predicted water level sequences and predicted inflow sequences for subsequent scheduling time units based on the rainfall feature vectors, flow feature vectors, and time delay feature expression results from consecutive scheduling time units.

[0050] Preferably, when the prediction output layer generates a water level prediction segment based on the time delay feature expression result, it first reads the delay impact state related to the reservoir water surface range, the monitoring station object, and the gate object in the time delay feature expression result, and then reads the historical water level state corresponding to the reservoir water surface range, the monitoring station object, and the gate object from the time-series grid data. Subsequently, the delay impact state and the historical water level state are fused to generate a water level prediction segment oriented towards a continuous scheduling time unit. The historical water level state originates from the water level data in the time-series grid data, and the historical water level state participates in the generation of the water level prediction segment, enabling the water level prediction segment to inherit the existing water level change basis in the time-series grid data. Each element in the water level prediction segment corresponds to a scheduling time unit and is used to express the predicted water level value under that scheduling time unit. After reading the water level prediction segment, the sequence sorting layer arranges the predicted water level values ​​corresponding to each scheduling time unit in order according to the grid time reference to obtain the predicted water level sequence. The reason why the predicted water level sequence is in the form of a "sequence" is that reservoir flood control scheduling needs to continuously read the predicted water level values ​​in multiple subsequent scheduling time units, rather than judging the water level result of a single scheduling time unit.

[0051] Preferably, when the prediction output layer generates an inflow prediction segment based on the time delay feature expression result, it first reads the delay impact state related to the upstream catchment area, the inflow river section, and the reservoir area from the time delay feature expression result, and then reads the historical inflow state corresponding to the upstream catchment area, the inflow river section, and the reservoir area from the time-series grid data. Subsequently, the delay impact state and the historical inflow state are fused to generate an inflow prediction segment oriented towards a continuous scheduling time unit. The historical inflow state originates from the inflow data in the time-series grid data, and the historical inflow state participates in the generation of the inflow prediction segment, enabling the inflow prediction segment to inherit the existing inflow change basis in the time-series grid data. Each element in the inflow prediction segment corresponds to a scheduling time unit and is used to express the predicted inflow value under that scheduling time unit. After reading the predicted inflow fragment, the sequence sorting layer arranges the predicted inflow values ​​corresponding to each scheduling time unit in sequence according to the grid time base to obtain the predicted inflow sequence. The reason why the predicted inflow sequence is in the form of a "sequence" is that the inflow has a continuous impact on the water level changes and gate scheduling in the subsequent scheduling time units. The continuously arranged predicted inflow sequence can serve as the time-series input for the subsequent scheduling optimization model.

[0052] Preferably, after the predicted water level sequence and the predicted inflow sequence are generated, they are kept in correspondence with the target reservoir object, spatial node, and scheduling time unit in the time-series grid data. Specifically, the predicted water level sequence corresponds to the reservoir water surface area, monitoring station object, and gate object, while the predicted inflow sequence corresponds to the upstream catchment area, inflow river cross-section, and reservoir area. The predicted water level sequence continues to participate in subsequent tiered early warning threshold comparison and the calculation of the highest water level constraint condition in the scheduling optimization model. The predicted inflow sequence continues to participate in the candidate scheduling coding processing and post-candidate scheduling water level change deduction in the scheduling optimization model. Therefore, the predicted water level sequence and the predicted inflow sequence are not isolated prediction results, but rather scheduling input data continuously transmitted from the digital twin reservoir base, the time-series grid data, the rainfall feature vector, the flow feature vector, and the time delay feature expression result.

[0053] Optionally, before the step of inputting the rainfall feature vector and the flow feature vector into the deep time-series neural network containing long short-term memory units, the method further includes: using a local outlier factor algorithm to detect outliers in the time-series grid data, and removing outlier data units from the time-series grid data to obtain time-series grid data with outlier data units removed; extracting corresponding rainfall feature vectors and flow feature vectors from the time-series grid data with outlier data units removed, so that the rainfall feature vectors and flow feature vectors input into the deep time-series neural network originate from the time-series grid data with outlier data units removed; performing cluster analysis on historical flood peak data to obtain a reference sequence of similar flood peaks; and inputting the reference sequence of similar flood peaks as prior reference information into the deep time-series neural network, so that the deep time-series neural network performs time series prediction based on the prior reference information, the rainfall feature vector, and the flow feature vector.

[0054] Preferably, before inputting the rainfall feature vector and the flow feature vector into the deep temporal neural network containing long short-term memory units, outlier detection preparation processing is performed on the temporal grid data; the temporal grid data includes spatial nodes, target reservoir objects, scheduling time units, rainfall data, water level data, inflow data, outflow data, gate feedback data, and binding status markers; the outlier detection preparation processing reads the rainfall data, water level data, inflow data, outflow data, and gate feedback data according to the target reservoir object and the scheduling time unit. The gate feedback data is processed, and the rainfall data, water level data, inflow data, outflow data, and gate feedback data of the same spatial node within a continuous scheduling time unit are organized into a local time series detection sample. The local time series detection sample retains the spatial node, the target reservoir object, the scheduling time unit, and the binding status marker, so that the subsequent local outlier factor algorithm can determine whether the local time series detection sample deviates from the local change pattern within the same spatial node, the same target reservoir object, and adjacent scheduling time units, instead of mixing all reservoir area data and performing a unified threshold judgment.

[0055] Preferably, the local outlier factor algorithm is used to detect outliers in local time-series detection samples in the time-series grid data. The technical essence of the local outlier factor algorithm lies in the following: for each local time-series detection sample to be detected, neighborhood samples are first determined within the same target reservoir object, adjacent spatial nodes, and adjacent scheduling time units. Then, the data distribution density between the local time-series detection sample and the neighborhood samples is compared. When the local time-series detection sample exhibits a lower local density relative to the neighborhood samples and its trend is inconsistent with that of adjacent scheduling time units, the local time-series detection sample is marked as an anomalous data unit. The neighborhood samples do not originate from unrelated spatial ranges, but rather from spatial nodes with topological associations to the same target reservoir object in the spatial topology binding results. For example, for the inflow data corresponding to the inflow river section, the neighborhood samples originate from the inflow data within adjacent scheduling time units, rainfall data corresponding to the upstream catchment area, and the inflow change status of the same inflow river section within consecutive scheduling time units. This ensures that the determination process of the anomalous data unit is simultaneously constrained by spatial and temporal associations. The abnormal data unit continues to be written into the corresponding spatial node and scheduling time unit in the time-series grid data so that the source location of the abnormal data unit can be located when the abnormal data unit is subsequently removed.

[0056] Preferably, when the local outlier factor algorithm is executed, the different data fields in the local time-series detection samples are first subjected to dimension unification processing. The dimension unification processing reads the units of measurement of the rainfall data, water level data, inflow data, outflow data, and gate feedback data, respectively, and converts them into normalized detection fields that can be used for similar distance measurements according to their respective data types, thus obtaining normalized local time-series detection samples. The normalized local time-series detection samples continue to participate in the neighborhood sample screening, so that the local outlier factor algorithm will not directly mix and compare water level units, flow units, and gate feedback units. Subsequently, the local neighborhood distance state is calculated based on the normalized local time-series detection samples, and an outlier state label is generated according to the local neighborhood distance state. The outlier state label is written into the corresponding row and column intersection position in the time-series grid data. The row and column intersection position is jointly determined by the spatial node and the scheduling time unit, so that the spatial node, target reservoir object, and scheduling time unit source of the outlier data unit can be retained when the outlier data unit is subsequently removed.

[0057] Preferably, after identifying the anomalous data unit, the anomalous data unit is removed from the time-series grid data to obtain time-series grid data with the anomalous data unit removed. This removal does not delete the spatial node and scheduling time unit corresponding to the row and column intersection of the anomalous data unit, but rather removes the rainfall data, water level data, inflow data, outflow data, or gate feedback data corresponding to the anomalous data unit from the subsequent feature extraction data input, while retaining the spatial node, target reservoir object, scheduling time unit, and outlier status marker corresponding to the anomalous data unit. Thus, the time-series grid data with the anomalous data unit removed still maintains its original spatial topology and temporal arrangement structure. When extracting rainfall feature vectors and flow feature vectors from the time-series grid data with the anomalous data unit removed, data that does not conform to local temporal variation patterns can be avoided, while continuing to use the spatial binding position and time index in the digital twin reservoir base plate.

[0058] Preferably, when extracting the corresponding rainfall feature vector from the time-series grid data of the removed outlier data units, rainfall data that has passed outlier detection is read, and rainfall time-series segments after outlier removal are formed according to the correspondence between upstream catchment area, reservoir area, monitoring station, and continuous scheduling time unit. The rainfall time-series segments after outlier removal are further processed by rainfall lag window to form the rainfall feature vector. The elements in the rainfall feature vector include rainfall intensity, cumulative rainfall status, rainfall duration status, and spatial location status of the monitoring station. The spatial location status of the monitoring station continues to be bound to the upstream catchment area or the reservoir area, so that the rainfall feature vector expresses both the temporal change process of the rainfall data and the spatial source of the rainfall data before entering the reservoir confluence path. The rainfall feature vector is then input into the deep time-series neural network to participate in subsequent time series prediction.

[0059] Preferably, when extracting the corresponding flow feature vector from the time-series grid data of the outlier-removed data units, the inflow flow data, outflow flow data, and gate feedback data that have passed outlier detection are read, and an outlier-removed flow time-series segments are formed according to the correspondence between the inflow river cross-section, spillway, water conveyance pipeline, gate object, and continuous scheduling time unit. The outlier-removed flow time-series segments are further processed by flow change windowing to form the flow feature vector. The elements in the flow feature vector include the inflow flow change status, outflow flow change status, gate feedback change status, and cross-sectional spatial location status. The cross-sectional spatial location status continues to be bound to the inflow river cross-section, the spillway, the water conveyance pipeline, or the gate object, so that the flow feature vector can provide the deep time-series neural network with the flow change basis after anomaly removal. The flow feature vector continues to be input into the deep time-series neural network together with the rainfall feature vector to participate in subsequent time series prediction.

[0060] Preferably, before performing cluster analysis on historical flood peak data, historical flood peak data is first read from historical reservoir documents and the time-series grid data from which outlier data units have been removed. This historical flood peak data includes the historical highest reservoir water level, the time of occurrence of the historical highest reservoir water level, the historical maximum inflow, the time of occurrence of the historical maximum inflow, the historical maximum outflow, the time of occurrence of the historical maximum outflow, the flood peak occurrence time, the flood process, runoff, rainfall, inflow process, outflow process, and gate feedback process. After reading the historical flood peak data, process alignment processing is performed according to the grid time reference to form historical flood peak process samples. These samples arrange the rainfall process, inflow process, outflow process, water level change process, and gate feedback process within the same flood peak process onto the same time axis, enabling subsequent cluster analysis to compare based on the temporal pattern of the flood peak process, rather than classifying based solely on a single flood peak value. These historical flood peak process samples continue to participate in flood peak morphology feature extraction, transforming the historical flood peak data into a clusterable data representation.

[0061] Preferably, when performing cluster analysis on the historical flood peak process samples, flood peak morphology features are first extracted from the historical flood peak process samples. These flood peak morphology features include the duration of flood peak rise, the duration of flood peak fall, the duration of peak value, the peak inflow rate, the outflow rate response, the water level rise, the peak occurrence time, and the rainfall concentration. The flood peak morphology features are then used in similarity measurement processing. This similarity measurement process compares different historical flood peak process samples according to the time pattern, flow change pattern, and water level change pattern of the flood peak process. Historical flood peak process samples with similar rise processes, similar peak occurrence time, similar inflow rate changes, and similar water level rise are grouped into the same flood peak category, resulting in similar flood peak samples. These similar flood peak samples retain their source historical flood peak data, flood peak morphology features, and corresponding target reservoir objects, and are used to generate similar flood peak reference sequences, enabling these similar flood peak reference sequences to inherit the process morphology information and target reservoir object information from the historical flood peak data.

[0062] Preferably, the similar flood peak reference sequence is generated based on the similar flood peak samples. During generation, multiple similar flood peak samples under the same flood peak category are first read. Then, the rainfall process, inflow process, outflow process, water level change process, and gate feedback process of the multiple similar flood peak samples are time-aligned according to the grid time reference. Subsequently, the process morphology of the time-aligned multiple similar flood peak samples is organized to form the similar flood peak reference sequence. Each sequence position in the similar flood peak reference sequence corresponds to a scheduling time unit. Each sequence position in the similar flood peak reference sequence records the reference rainfall state, reference inflow state, reference outflow state, reference water level state, and reference gate feedback state under that scheduling time unit. The similar flood peak reference sequence is in "sequence" form because the flood peak process has the time evolution characteristics of continuous rise, peak maintenance, and gradual decline. The similar flood peak reference sequence can provide this time evolution characteristic to the deep temporal neural network in the form of continuous scheduling time units. The similar flood peak reference sequence continues to serve as a source of prior reference information, used to match the rainfall feature vector and flow feature vector formed within the current scheduling time unit.

[0063] Preferably, when the similar flood peak reference sequence is input into the deep time-series neural network as prior reference information, the current flood peak morphology characteristics of the current flood peak process are first determined based on the rainfall feature vector and the flow feature vector within the current scheduling time unit; then, the current flood peak morphology characteristics are matched with the flood peak morphology characteristics corresponding to the similar flood peak reference sequence to select a similar flood peak reference sequence corresponding to the current flood peak process; subsequently, the selected similar flood peak reference sequence is input into the deep time-series neural network as the prior reference information. The prior reference information is not a prediction input replacing the rainfall feature vector and the flow feature vector, but rather enters the deep time-series neural network together with the rainfall feature vector and the flow feature vector, enabling the deep time-series neural network to read the time evolution reference of the similar flood peak reference sequence based on the current monitoring state.

[0064] Preferably, when the deep temporal neural network receives the prior reference information, the temporal input processing layer reads the similar flood peak reference sequence, the rainfall feature vector, and the flow feature vector, and organizes them into an enhanced temporal input sample within the same time window according to the grid time reference. Each sample unit in the enhanced temporal input sample includes the rainfall intensity, cumulative rainfall status, inflow change status, outflow change status, and gate feedback change status of the current scheduling time unit, as well as the reference rainfall status, reference inflow status, reference outflow status, and reference water level status corresponding to the same scheduling time unit. The enhanced temporal input sample is further input into the spatial node embedding layer, which transforms the spatially bound positions and target reservoir objects in the enhanced temporal input sample into spatial node embedding expressions. The spatial node embedding expressions are then input together with the enhanced temporal input sample into the rainfall-flow coupling layer. The enhanced temporal input sample is formed by the similar flood peak reference sequence, the rainfall feature vector, and the flow feature vector, thus enabling the enhanced temporal input sample to simultaneously represent the current monitoring process and the similar flood peak reference process.

[0065] Preferably, after receiving the enhanced time-series input samples and the spatial node embedding expression, the rainfall-flow coupling layer fuses the rainfall state in the rainfall feature vector, the flow state in the flow feature vector, the reference flood peak state in the similar flood peak reference sequence, and the spatial node embedding expression to form a reference-enhanced rainfall-flow coupling expression. This expression retains the differences between the current monitoring process and the similar flood peak reference sequence, and uses these differences as one of the input contents for subsequent long short-term memory units (LSMs). These differences express the deviation trends of the current rainfall state, the current inflow change state, and the current water level correlation state relative to the similar flood peak reference sequence. The reference-enhanced rainfall-flow coupling expression continues to be input into the LSM, enabling the LSM to simultaneously read the current time-series information and the similar flood peak reference information, and to form a time-delay feature expression result through the reference-enhanced rainfall-flow coupling expression.

[0066] Preferably, after the Long Short-Term Memory (LSTM) unit receives the reference enhanced rainfall-flow coupled expression, the input gate processing structure selects input states related to the predicted water level sequence and the predicted inflow sequence from the reference enhanced rainfall-flow coupled expression; the forget gate processing structure reads the historical memory state formed in the previous scheduling time unit and determines the delayed impact information that needs to be retained in the historical memory state according to the reference enhanced rainfall-flow coupled expression; the candidate state processing structure generates candidate memory states according to the reference enhanced rainfall-flow coupled expression, and the candidate memory states are used to express the new impact of the current rainfall state, the current flow state, and the prior reference information on the predicted water level sequence and the predicted inflow sequence; the state update processing structure merges the historical memory state, the input state, and the candidate memory state to form the current memory state; the output gate processing structure generates a time delay feature expression result according to the current memory state. The time delay feature expression result continues to be input into the prediction output layer, enabling the prediction output layer to generate water level prediction segments and inflow prediction segments respectively under the joint constraints of the current monitoring state and the same type of flood peak reference sequence.

[0067] Preferably, when the prediction output layer generates a water level prediction segment based on the time delay feature expression result, it reads the delay impact status related to the reservoir water surface range, monitoring station objects, and gate objects in the time delay feature expression result, and reads the reference water level status in the similar flood peak reference sequence; subsequently, it merges the delay impact status, the reference water level status, and the historical water level status in the time-series grid data after removing abnormal data units to generate a water level prediction segment. The water level prediction segment is then input into the sequence processing layer, which processes the predicted water level values ​​in the water level prediction segment into a predicted water level sequence according to the grid time reference. The predicted water level sequence maintains a time-position correspondence with the similar flood peak reference sequence, enabling the predicted water level sequence to express the differences in water level changes between the current rainfall data and current flow data and the similar flood peak reference sequence; the predicted water level sequence continues to participate in subsequent graded early warning threshold comparisons and maximum water level constraint calculations.

[0068] Preferably, when the prediction output layer generates the inflow prediction segment based on the time delay feature expression result, it reads the delay impact state related to the upstream catchment area, the inflow river section, and the reservoir area from the time delay feature expression result, and reads the reference inflow state from the similar flood peak reference sequence. Subsequently, the delay impact state, the reference inflow state, and the historical inflow state from the time-series grid data with outlier data units removed are fused to generate the inflow prediction segment. The inflow prediction segment is then input to the sequence processing layer, which organizes the predicted inflow values ​​in the inflow prediction segment into a predicted inflow sequence according to the grid time reference. The predicted inflow sequence continues to serve as input to the subsequent scheduling optimization model, enabling the scheduling optimization model to read the predicted inflow sequence after outlier removal and enhancement by the similar flood peak reference sequence during candidate scheduling coding processing. The predicted inflow sequence continues to participate in subsequent candidate scheduling coding processing and post-candidate scheduling water level change extrapolation.

[0069] Preferably, after the above processing, the rainfall feature vector and the flow feature vector input to the deep temporal neural network both originate from the time-series grid data of the removed outlier data units, and the prior reference information originates from the similar flood peak reference sequence obtained from the clustering analysis of historical flood peak data. The deep temporal neural network performs time series prediction based on the prior reference information, the rainfall feature vector, and the flow feature vector to obtain the predicted water level sequence and the predicted inflow sequence. The predicted water level sequence continues to participate in subsequent graded early warning threshold comparison and maximum water level constraint calculation, and the predicted inflow sequence continues to participate in subsequent candidate scheduling coding processing and post-candidate scheduling water level change deduction, so that outlier detection, historical flood peak clustering analysis, and similar flood peak reference sequences can all enter the subsequent flood control scheduling processing chain.

[0070] Optionally, the step of inputting the predicted water level sequence and the predicted inflow sequence into the scheduling optimization model, and performing candidate scheduling coding processing on the preset reservoir gate opening range and scheduling time unit according to the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rules in the scheduling optimization model, to obtain an initial group composed of multiple candidate scheduling sequence codes, and performing group iterative optimization processing on the initial group to obtain the target scheduling sequence includes: in the scheduling optimization model, according to the predicted water level sequence, the predicted inflow sequence, and the preset reservoir gate opening range and scheduling time unit, performing candidate scheduling coding processing on the preset reservoir gate opening range and scheduling time unit, to obtain an initial group composed of multiple candidate scheduling sequence codes, and performing group iterative optimization processing on the initial group to obtain the target scheduling sequence. Multiple reservoir gate opening degree combinations are generated based on the opening degree range and the scheduling time unit; each reservoir gate opening degree combination is encoded as a candidate scheduling sequence code, and the initial group is composed of multiple candidate scheduling sequence codes; candidate water level change sequences and candidate discharge flow sequences after candidate scheduling are derived based on each candidate scheduling sequence code, the predicted water level sequence, and the predicted inflow sequence; based on the highest water level constraint, the discharge flow amplitude constraint, and the scheduling sequence optimization rule, the candidate water level change sequences, the candidate discharge flow sequences, and the initial group are subjected to group iterative optimization processing to obtain the target scheduling sequence.

[0071] Preferably, in the specific implementation of inputting the predicted water level sequence and the predicted inflow sequence into the scheduling optimization model, the scheduling optimization model includes a scheduling input organization structure, a reservoir gate opening degree combination generation structure, a candidate scheduling coding structure, a candidate scheduling deduction structure, a group iterative optimization structure, and a scheduling sequence decoding structure; the scheduling input organization structure is used to read the predicted water level sequence, the predicted inflow sequence, the preset reservoir gate opening degree range, the scheduling time unit, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rule, and organizes the predicted water level sequence, the predicted inflow sequence, the preset reservoir gate opening degree range, the scheduling time unit, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rule into scheduling input samples according to the scheduling time unit; the reservoir gate... The gate opening degree combination generation structure is used to generate multiple reservoir gate opening degree combinations based on the scheduling input samples; the candidate scheduling coding structure is used to encode each reservoir gate opening degree combination into a candidate scheduling sequence code, and the initial group is composed of multiple candidate scheduling sequence codes; the candidate scheduling deduction structure is used to deduce the candidate water level change sequence and candidate discharge flow sequence after candidate scheduling based on each candidate scheduling sequence code, the predicted water level sequence, and the predicted inflow sequence; the group iterative optimization structure is used to perform group iterative optimization processing on the candidate water level change sequence, the candidate discharge flow sequence, and the initial group based on the highest water level constraint, the discharge flow amplitude constraint, and the scheduling sequence optimization rules to obtain the target candidate scheduling sequence code; the scheduling sequence decoding structure is used to decode the target candidate scheduling sequence code into the target scheduling sequence.

[0072] Preferably, after reading the predicted water level sequence and the predicted inflow sequence, the scheduling input processing structure first determines the correspondence between each predicted water level value in the predicted water level sequence and the scheduling time unit according to the grid time reference, and then determines the correspondence between each predicted inflow value in the predicted inflow sequence and the scheduling time unit according to the grid time reference. Subsequently, the scheduling input processing structure organizes the predicted water level value, predicted inflow value, preset reservoir gate opening range, highest water level constraint, and downstream flow variation constraint under the same scheduling time unit into scheduling time unit input segments. Multiple consecutive scheduling time unit input segments are arranged in chronological order to form the scheduling input sample. The scheduling input sample is then input into the reservoir gate opening combination generation structure, enabling the reservoir gate opening combination generation structure to simultaneously read the predicted water level value, predicted inflow value, and preset reservoir gate opening range under the same grid time reference, and generate multiple reservoir gate opening combinations based on the predicted water level value, predicted inflow value, and preset reservoir gate opening range.

[0073] Preferably, when the reservoir gate opening degree combination generation structure generates multiple reservoir gate opening degree combinations based on the scheduling input sample, it first reads the preset reservoir gate opening degree range and expands the preset reservoir gate opening degree range according to the gate object and scheduling time unit to form a gate time period adjustable table; the row direction in the gate time period adjustable table corresponds to the gate object, the column direction in the gate time period adjustable table corresponds to the scheduling time unit, and the row and column intersection position in the gate time period adjustable table is used to record the reservoir gate opening degree that the corresponding gate object can select within the corresponding scheduling time unit. Subsequently, the reservoir gate opening degree combination generation structure combines candidate values ​​for the reservoir gate opening degree in the gate time period adjustable table based on the predicted water level value in the predicted water level sequence and the predicted inflow value in the predicted inflow sequence, resulting in multiple reservoir gate opening degree combinations. Each reservoir gate opening degree combination represents the reservoir gate opening degree arrangement of multiple gate objects within a continuous scheduling time unit. The reservoir gate opening degree combination is then input into the candidate scheduling coding structure, enabling the candidate scheduling coding structure to convert the reservoir gate opening degree arrangement within the continuous scheduling time unit into a candidate scheduling sequence code.

[0074] Preferably, when the candidate scheduling coding structure encodes each combination of reservoir gate opening degrees, it generates candidate scheduling sequence codes according to the correspondence between the gate object, the scheduling time unit, and the reservoir gate opening degree. The candidate scheduling sequence code includes the gate object position, the execution time position, and the reservoir gate opening degree position. The gate object position expresses the gate object participating in the scheduling, the execution time position expresses the time position at which the gate object executes the corresponding reservoir gate opening degree, and the reservoir gate opening degree position expresses the reservoir gate opening degree of the gate object within the scheduling time unit. The candidate scheduling coding structure arranges multiple candidate scheduling sequence codes according to the time order of the scheduling time unit, and the initial group is composed of multiple candidate scheduling sequence codes. The initial group is then input into the candidate scheduling deduction structure, enabling the candidate scheduling deduction structure to read the candidate scheduling sequence codes one by one, and to deduce the candidate water level change sequence and candidate outflow sequence after the candidate scheduling based on the candidate scheduling sequence codes, the predicted water level sequence, and the predicted inflow sequence.

[0075] Preferably, when the candidate scheduling deduction structure deduces the candidate water level change sequence and candidate discharge flow sequence after candidate scheduling based on each candidate scheduling sequence code, the predicted water level sequence, and the predicted inflow sequence, it first reads the gate object position, execution time position, and reservoir gate opening position from a candidate scheduling sequence code, and converts the reservoir gate opening position into the candidate gate action state within the corresponding scheduling time unit. The candidate gate action state continues to be read in correspondence with the predicted inflow value of the predicted inflow sequence within the same scheduling time unit to form a candidate scheduling inflow-outflow relationship. The candidate scheduling deduction structure deduces the candidate discharge flow sequence after candidate scheduling based on the candidate scheduling inflow-outflow relationship, and temporally couples the candidate discharge flow sequence with the predicted water level sequence and the predicted inflow sequence to deduce the candidate water level change sequence after candidate scheduling. Thus, the candidate water level change sequence and the candidate discharge flow sequence both originate from the combined effect of the candidate scheduling sequence code, the predicted water level sequence, and the predicted inflow sequence, rather than being generated in isolation by the candidate scheduling sequence code.

[0076] Preferably, the derivation process of the candidate discharge flow sequence includes gate opening degree reading, gate discharge state conversion, and discharge flow time sequence arrangement. Gate opening degree reading is used to read the reservoir gate opening degree of each gate object in each scheduling time unit from the candidate scheduling sequence encoding to obtain the candidate gate action state. Gate discharge state conversion is used to obtain the candidate discharge flow value in each scheduling time unit based on the candidate gate action state, the predicted water level value in the predicted water level sequence, and the discharge relationship corresponding to the gate object. Discharge flow time sequence arrangement is used to arrange multiple candidate discharge flow values ​​sequentially according to the scheduling time unit to form the candidate discharge flow sequence. The candidate discharge flow sequence is then input into the group iterative optimization structure and participates in the fitness calculation under the discharge flow amplitude constraint condition, so that the reservoir gate opening degree change corresponding to the candidate scheduling sequence encoding can be transformed into an evaluable candidate discharge flow sequence.

[0077] Preferably, the deduction process of the candidate water level change sequence includes water volume status reading, water level change deduction, and water level time series organization. The water volume status reading is used to read the predicted inflow value from the predicted inflow sequence, the candidate outflow value from the candidate outflow sequence, and the predicted water level value from the predicted water level sequence within the same scheduling time unit, obtaining the candidate water level deduction input state. The water level change deduction is used to deduce the water level change state after candidate scheduling within the scheduling time unit based on the candidate water level deduction input state, obtaining the candidate water level change value. The water level time series organization is used to arrange multiple candidate water level change values ​​sequentially according to the scheduling time unit, forming the candidate water level change sequence. The candidate water level change sequence continues to be input into the group iterative optimization structure and participates in the fitness calculation under the highest water level constraint condition, so that the impact of the candidate scheduling sequence encoding on the reservoir water level can be evaluated in the form of a candidate water level change sequence within a continuous scheduling time unit.

[0078] Preferably, when the group iterative optimization structure performs group iterative optimization processing on the candidate water level change sequence, the candidate discharge flow sequence, and the initial group, it first reads the corresponding candidate water level change sequence and candidate discharge flow sequence for each candidate scheduling sequence encoding in the initial group; then, it calculates the fitness of the candidate water level change sequence according to the maximum water level constraint to obtain a first fitness value for the candidate scheduling sequence encoding, and calculates the fitness of the candidate discharge flow sequence according to the discharge flow amplitude constraint to obtain a second fitness value for the candidate scheduling sequence encoding. The first fitness value is used to express the degree of matching between the candidate water level change sequence and the maximum water level constraint, and the second fitness value is used to express the degree of matching between the candidate discharge flow sequence and the discharge flow amplitude constraint; the first fitness value and the second fitness value continue to participate in the comprehensive fitness value calculation, so that each candidate scheduling sequence encoding can be evaluated simultaneously in terms of both the maximum water level constraint and the discharge flow amplitude constraint.

[0079] Preferably, when determining the comprehensive fitness value based on the first fitness value and the second fitness value, the group iterative optimization structure first reads the water level constraint evaluation weight, the discharge flow amplitude evaluation weight, and the gate scheduling continuity evaluation weight in the scheduling sequence optimization rules. Then, it weights the first fitness value according to the water level constraint evaluation weight to obtain a weighted first fitness value. It then weights the second fitness value according to the discharge flow amplitude evaluation weight to obtain a weighted second fitness value. Finally, it performs a smoothing evaluation on the reservoir gate opening degree changes within adjacent scheduling time units in the candidate scheduling sequence code according to the gate scheduling continuity evaluation weight to obtain a scheduling continuity fitness value. The group iterative optimization structure integrates the weighted first fitness value, the weighted second fitness value, and the scheduling continuity fitness value to obtain the comprehensive fitness value. The comprehensive fitness value is then bound to the corresponding candidate scheduling sequence code, enabling subsequent group iterative optimization processing to select a candidate scheduling sequence code that better conforms to the scheduling sequence optimization rules based on the comprehensive fitness value.

[0080] Preferably, the scheduling sequence optimization rules include candidate scheduling sequence code selection rules, candidate scheduling sequence code crossover rules, candidate scheduling sequence code perturbation rules, and iteration stopping rules. The candidate scheduling sequence code selection rules are used to select candidate scheduling sequence codes from the initial population to participate in the next round of iteration based on the comprehensive fitness value. The candidate scheduling sequence code crossover rules are used to perform segment swapping on the reservoir gate opening degree arrangements corresponding to some gate objects in different candidate scheduling sequence codes, while maintaining the order of the scheduling time units, to generate cross-candidate scheduling sequence codes. The candidate scheduling sequence code perturbation rules are used to adjust the reservoir gate opening degree positions in the cross-candidate scheduling sequence codes within a limited range of the preset reservoir gate opening degree, to form updated candidate scheduling sequence codes. The iteration stopping rules are used to stop the population iterative optimization process when the number of iterations reaches a preset number of iterations or when the change in the comprehensive fitness value obtained from continuous iterations is less than a preset change range. The candidate scheduling sequence encoding selection rule, the candidate scheduling sequence encoding crossover rule, the candidate scheduling sequence encoding perturbation rule, and the iteration stopping rule work together to enable the group iterative optimization structure to perform round-by-round screening and updating among multiple candidate scheduling sequence encodings, and to continue inputting the screened and updated candidate scheduling sequence encodings into the candidate scheduling deduction structure.

[0081] Preferably, when the group iterative optimization structure performs group iterative optimization processing, it first sorts the candidate scheduling sequence codes in the initial group according to the comprehensive fitness value corresponding to each candidate scheduling sequence code, and selects candidate parent codes according to the candidate scheduling sequence code selection rules. The candidate parent codes are derived from the candidate scheduling sequence codes in the initial group whose comprehensive fitness values ​​meet the candidate scheduling sequence code selection rules, and continue to participate in the candidate scheduling sequence code cross-processing. Subsequently, the candidate parent codes are fragment-swapped according to the candidate scheduling sequence code cross-processing rules to obtain cross-candidate scheduling sequence codes, and the reservoir gate opening degree of the cross-candidate scheduling sequence codes is adjusted according to the candidate scheduling sequence code perturbation rules to obtain updated candidate scheduling sequence codes. The candidate scheduling deduction structure continues to deduce updated candidate water level change sequences and updated candidate outflow sequences based on the updated candidate scheduling sequence codes, the predicted water level sequence, and the predicted inflow sequence. The group iterative optimization structure continues to calculate updated comprehensive fitness values ​​based on the updated candidate water level change sequences and updated candidate outflow sequences. Through the above round-by-round processing, the initial group is updated into a multi-round iterative group, and the updated candidate scheduling sequence encoding in the multi-round iterative group is continuously associated with the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, and the outflow amplitude constraint.

[0082] Preferably, during the group iterative optimization process, if a candidate water level change sequence corresponding to a candidate scheduling sequence code has a water level change state exceeding the highest water level constraint, the group iterative optimization structure reduces its water level constraint matching degree when calculating the comprehensive fitness value of the candidate scheduling sequence code; if a candidate discharge flow sequence corresponding to a candidate scheduling sequence code has a discharge flow change state exceeding the discharge flow amplitude constraint, the group iterative optimization structure reduces its discharge flow amplitude constraint matching degree when calculating the comprehensive fitness value of the candidate scheduling sequence code; if a candidate scheduling sequence code can make the candidate water level change sequence satisfy the highest water level constraint and make the candidate discharge flow sequence satisfy the discharge flow amplitude constraint, the group iterative optimization structure retains the candidate scheduling sequence code as a constraint-matching candidate scheduling sequence code. The constraint-matching candidate scheduling sequence code continues to participate in subsequent iterations, or is determined by the group iterative optimization structure as the target candidate scheduling sequence code when the iteration stopping rule is met, so that the target scheduling sequence can originate from the candidate scheduling sequence code after continuous constraint evaluation.

[0083] Preferably, after obtaining the target candidate scheduling sequence encoding, the scheduling sequence decoding structure decodes the target candidate scheduling sequence encoding based on the gate object position, execution time position, and reservoir gate opening position to form the target scheduling sequence. The target scheduling sequence includes gate objects, target reservoir gate opening degrees, and execution time positions arranged in the order of the scheduling time units. The gate objects correspond to the gate objects associated with the physical reservoir control terminal, the target reservoir gate opening degree corresponds to the executable reservoir gate opening degree within the preset reservoir gate opening degree range, and the execution time position corresponds to the scheduling time unit. The target scheduling sequence continues to participate in the generation of subsequent physical reservoir control instructions, enabling the target candidate scheduling sequence encoding obtained by the scheduling optimization model to be converted into subsequent executable scheduling control content, rather than remaining at the level of abstract candidate scheduling sequence encoding.

[0084] Preferably, after the above processing, the scheduling optimization model uses the predicted water level sequence and the predicted inflow sequence as inputs for time series prediction, the preset reservoir gate opening range and the scheduling time unit as candidate scheduling search boundaries, the highest water level constraint and the outflow variation constraint as candidate scheduling evaluation criteria, and the scheduling sequence optimization rule as the basis for group iterative updates to generate the target scheduling sequence. The target scheduling sequence is then input into the subsequent physical reservoir control command generation process, enabling the predicted water level sequence and predicted inflow sequence obtained in the prediction stage to be transformed into a scheduling execution basis for the physical reservoir control terminal through candidate scheduling encoding, candidate scheduling deduction, and group iterative optimization processing.

[0085] Optionally, the step of performing group iterative optimization processing on the candidate water level change sequence, the candidate discharge flow rate sequence, and the initial group based on the highest water level constraint, the discharge flow rate variation constraint, and the scheduling sequence optimization rule to obtain the target scheduling sequence includes: for each candidate scheduling sequence code in the initial group, calculating the fitness of the candidate water level change sequence corresponding to the candidate scheduling sequence code according to the highest water level constraint to obtain a first fitness value of the candidate scheduling sequence code; calculating the fitness of the candidate discharge flow rate sequence corresponding to the candidate scheduling sequence code according to the discharge flow rate variation constraint to obtain a second fitness value of the candidate scheduling sequence code; determining the comprehensive fitness value of the candidate scheduling sequence code based on the first fitness value and the second fitness value; performing group iterative optimization processing on the initial group according to the comprehensive fitness value corresponding to each candidate scheduling sequence code and the scheduling sequence optimization rule to obtain the target candidate scheduling sequence code; and decoding the target candidate scheduling sequence code into the target scheduling sequence.

[0086] Preferably, in the specific implementation of the group iterative optimization process for the candidate water level change sequence, the candidate discharge flow rate sequence, and the initial group based on the highest water level constraint, the discharge flow rate variation constraint, and the scheduling sequence optimization rule, the code of each candidate scheduling sequence in the initial group is first read, and the candidate water level change sequence and the candidate discharge flow rate sequence corresponding to the candidate scheduling sequence code are read. The candidate water level change sequence includes the candidate water level change value within a continuous scheduling time unit, the execution time position corresponding to the candidate water level change value, and the target reservoir object corresponding to the candidate water level change value. The candidate discharge flow rate sequence includes the candidate discharge flow rate value within a continuous scheduling time unit, the execution time position corresponding to the candidate discharge flow rate value, and the gate object corresponding to the candidate discharge flow rate value. After reading the candidate water level change sequence and the candidate discharge flow sequence, a correspondence is established between the candidate water level change sequence and the highest water level constraint, and a correspondence is established between the candidate discharge flow sequence and the discharge flow amplitude constraint. This allows subsequent fitness calculations to be performed based on the candidate water level change process and the candidate discharge flow change process, respectively, rather than making a static judgment based solely on the reservoir gate opening degree in the candidate scheduling sequence encoding.

[0087] Preferably, when calculating the fitness of the candidate water level change sequence corresponding to the candidate scheduling sequence encoding according to the highest water level constraint, the upper limit of the constraint water level corresponding to the target reservoir object, the scheduling time unit range applicable to the constraint water level, and the water level comparison benchmark corresponding to the constraint water level are first read from the highest water level constraint. The water level comparison benchmark is used to limit the water level benchmark caliber used when comparing the candidate water level change value with the upper limit of the constraint water level in the same dimension. Then, the candidate water level change values ​​in the candidate water level change sequence are read one by one according to the execution time position. For each execution time position, the candidate water level change value at that execution time position is compared with the corresponding upper limit of the constraint water level in the same dimension according to the water level comparison benchmark to obtain the water level constraint deviation state. The water level constraint deviation state includes the water level constraint satisfied state that does not exceed the upper limit of the constraint water level and the water level constraint exceeded state that exceeds the upper limit of the constraint water level. Subsequently, the deviation states of multiple water level constraints within the continuous scheduling time unit are cumulatively evaluated in chronological order to obtain the first fitness value of the candidate scheduling sequence code. The first fitness value is used to express the degree of matching between the candidate water level change sequence corresponding to the candidate scheduling sequence code and the highest water level constraint condition, and continues to participate in the comprehensive fitness value calculation of the candidate scheduling sequence code.

[0088] Preferably, the calculation of the first fitness value not only determines whether a single candidate water level change value in the candidate water level change sequence exceeds the maximum water level constraint, but also evaluates the proximity and persistence of the candidate water level change sequence within consecutive scheduling time units. Specifically, when the candidate water level change value at a certain execution time position does not exceed the maximum water level constraint but continues to approach the upper limit of the constraint water level, the water level constraint deviation state corresponding to that execution time position is recorded as the near-upper limit state; when multiple consecutive scheduling time units are all in the near-upper limit state, the near-upper limit states corresponding to multiple consecutive scheduling time units are jointly written into the calculation process of the first fitness value. Thus, the first fitness value can reflect both whether the candidate water level change sequence meets the maximum water level constraint and the water level rise trend of the candidate water level change sequence in subsequent consecutive scheduling time units, providing calculable data basis for whether the candidate scheduling sequence encoding is suitable to continue participating in the group iterative optimization process.

[0089] Preferably, when calculating the fitness of the candidate discharge flow sequence corresponding to the candidate scheduling sequence encoding according to the discharge flow amplitude constraint, the allowable discharge flow variation range between adjacent scheduling time units, the discharge flow constraint object corresponding to the gate object, and the discharge flow change evaluation direction are first read from the discharge flow amplitude constraint. The discharge flow change evaluation direction is used to limit the candidate discharge flow value to perform incremental evaluation, decremental evaluation, or bidirectional amplitude evaluation between adjacent scheduling time units. Then, the candidate discharge flow value in the candidate discharge flow sequence is read according to the execution time position. For two adjacent execution time positions, the difference between the candidate discharge flow value corresponding to the later execution time position and the candidate discharge flow value corresponding to the earlier execution time position is calculated to obtain the candidate discharge flow amplitude state. Subsequently, the candidate discharge flow amplitude state is compared with the discharge flow amplitude constraint according to the discharge flow change evaluation direction to obtain the discharge flow amplitude constraint deviation state. The deviation states of the outflow amplitude constraints corresponding to multiple consecutive scheduling time units are continuously evaluated in chronological order to obtain the second fitness value of the candidate scheduling sequence code. The second fitness value is used to express the degree of matching between the candidate outflow sequence corresponding to the candidate scheduling sequence code and the outflow amplitude constraint condition, and continues to participate in the comprehensive fitness value calculation of the candidate scheduling sequence code.

[0090] Preferably, the calculation of the second fitness value is further evaluated in conjunction with the changes in reservoir gate opening degree in the candidate scheduling sequence encoding. Specifically, the reservoir gate opening degree position corresponding to the same gate object within adjacent execution time positions in the candidate scheduling sequence encoding is first read, and a reservoir gate opening degree change state is formed based on the changes in the reservoir gate opening degree position; then, the reservoir gate opening degree change state is correlated with the candidate discharge flow rate change state to determine whether the candidate discharge flow rate change state originates from continuous adjustment of the reservoir gate opening degree. If the candidate discharge flow rate change state exceeds the discharge flow rate change constraint condition, and the corresponding reservoir gate opening degree change state also has a large jump between adjacent execution time positions, then the discharge flow rate change deviation information corresponding to the adjacent execution time position is written into the calculation process of the second fitness value. The outflow variation deviation information includes the corresponding candidate outflow variation status, reservoir gate opening change status, and execution time position, so that the second fitness value can link the reservoir gate opening change in the candidate scheduling sequence encoding with the outflow variation in the candidate outflow sequence, so that the outflow variation constraint is no longer an isolated flow value constraint, but directly acts on the iterative screening process of the candidate scheduling sequence encoding.

[0091] Preferably, when determining the comprehensive fitness value of the candidate scheduling sequence code based on the first fitness value and the second fitness value, the water level constraint evaluation weight, the discharge flow amplitude evaluation weight, and the gate scheduling continuity evaluation weight in the scheduling sequence optimization rules are first read. The water level constraint evaluation weight is used to adjust the participation intensity of the first fitness value in the comprehensive evaluation, the discharge flow amplitude evaluation weight is used to adjust the participation intensity of the second fitness value in the comprehensive evaluation, and the gate scheduling continuity evaluation weight is used to adjust the evaluation intensity of the continuous change of the reservoir gate opening degree between adjacent execution time positions. Subsequently, the first fitness value is weighted according to the water level constraint evaluation weight to obtain a weighted first fitness value; the second fitness value is weighted according to the discharge flow amplitude evaluation weight to obtain a weighted second fitness value; and the scheduling continuity evaluation of the reservoir gate opening degree change state in the candidate scheduling sequence code is performed according to the gate scheduling continuity evaluation weight to obtain a scheduling continuity fitness value. The weighted first fitness value, the weighted second fitness value, and the scheduling continuity fitness value together form the comprehensive fitness value of the candidate scheduling sequence encoding.

[0092] Preferably, after the comprehensive fitness value is formed, it is further bound to the candidate scheduling sequence code, and the candidate scheduling sequence code, the corresponding candidate water level change sequence, the corresponding candidate discharge flow sequence, and the comprehensive fitness value corresponding to the candidate scheduling sequence code are jointly written into the group evaluation record. The group evaluation record is used to express the evaluation status of each candidate scheduling sequence code in the initial group under the constraints of the highest water level, the discharge flow amplitude, and the scheduling sequence optimization rules. The group evaluation record is further input into the group iterative optimization process, so that the group iterative optimization process can directly read the comprehensive fitness value corresponding to each candidate scheduling sequence code, and can trace the source relationship between the candidate scheduling sequence code and the candidate water level change sequence and the candidate discharge flow sequence based on the group evaluation record.

[0093] Preferably, when performing group iterative optimization on the initial population based on the comprehensive fitness value corresponding to each candidate scheduling sequence code and the scheduling sequence optimization rule, the candidate scheduling sequence codes in the initial population are first sorted according to the comprehensive fitness value to obtain a candidate scheduling sequence code sorting result; then, according to the candidate scheduling sequence code selection rule in the scheduling sequence optimization rule, candidate parent codes participating in the next round of iteration are selected from the candidate scheduling sequence code sorting result. The candidate parent codes continue to undergo segment exchange according to the candidate scheduling sequence code crossover rule in the scheduling sequence optimization rule to obtain cross-candidate scheduling sequence codes; the cross-candidate scheduling sequence codes continue to adjust the reservoir gate opening position according to the candidate scheduling sequence code perturbation rule in the scheduling sequence optimization rule to obtain updated candidate scheduling sequence codes. The updated candidate scheduling sequence codes continue to be input into the candidate scheduling deduction process to re-deduct updated candidate water level change sequences and updated candidate discharge flow sequences, and the updated candidate water level change sequences and updated candidate discharge flow sequences continue to participate in the calculation of the first fitness value, the second fitness value, and the comprehensive fitness value.

[0094] Preferably, when the candidate scheduling sequence encoding cross-rule execution segment exchange occurs, the gate object position, execution time position, and reservoir gate opening position in different candidate parent codes are read with the execution time position as the boundary. While maintaining the order of the execution time positions, the reservoir gate opening positions within consecutive execution time positions in one candidate parent code are exchanged with the corresponding reservoir gate opening positions within the execution time positions in another candidate parent code, resulting in the cross-candidate scheduling sequence encoding. The cross-candidate scheduling sequence encoding still retains the correspondence between the gate object position, execution time position, and reservoir gate opening position, allowing it to continue to be read by the candidate scheduling deduction process. When the candidate scheduling sequence encoding perturbation rule executes reservoir gate opening position adjustment, the reservoir gate opening position in the cross-candidate scheduling sequence encoding is adjusted only within the preset reservoir gate opening range, and the adjusted reservoir gate opening position is rewritten into the updated candidate scheduling sequence encoding, ensuring that the updated candidate scheduling sequence encoding still satisfies the preset reservoir gate opening range.

[0095] Preferably, after each round of group iterative optimization, the updated candidate scheduling sequence code, the updated candidate water level change sequence, the updated candidate discharge flow sequence, and the comprehensive fitness value corresponding to the updated candidate scheduling sequence code are written into the iterative group record. The iterative group record is compared with the previous round of group evaluation record to determine whether the iteration stopping rule in the scheduling sequence optimization rules is satisfied. If the iterative group record does not yet satisfy the iteration stopping rule, the iterative group record is used as the input for the next round of group iterative optimization, and the candidate scheduling sequence code selection, candidate scheduling sequence code crossover, reservoir gate opening position adjustment, candidate scheduling deduction, and fitness calculation continue to be performed. If the iterative group record satisfies the iteration stopping rule, the candidate scheduling sequence code whose comprehensive fitness value conforms to the scheduling sequence optimization rules is selected from the iterative group record as the target candidate scheduling sequence code. The target candidate scheduling sequence code is derived from the candidate scheduling sequence code after continuous evaluation of the candidate water level change sequence and the candidate discharge flow sequence, rather than from the reservoir gate opening enumeration result without scheduling deduction.

[0096] Preferably, after obtaining the target candidate scheduling sequence code, the target candidate scheduling sequence code is decoded to form the target scheduling sequence. During decoding, the gate object position, execution time position, and reservoir gate opening degree position in the target candidate scheduling sequence code are first read; then, based on the gate object position, the gate object participating in the flood control scheduling action is determined; based on the execution time position, the execution time position for that gate object to execute the reservoir gate opening degree is determined; and based on the reservoir gate opening degree position, the target reservoir gate opening degree of that gate object at the corresponding execution time position is determined. Subsequently, the target reservoir gate opening degrees corresponding to multiple gate objects at the same execution time position are organized into time-slice scheduling records, and multiple time-slice scheduling records are arranged according to the chronological order of the execution time positions to obtain the target scheduling sequence; the target scheduling sequence therefore includes the correspondence between gate objects, execution time positions, and target reservoir gate opening degrees.

[0097] Preferably, after the target scheduling sequence is formed, it continues to be associated with the candidate water level change sequence, candidate discharge flow sequence, and comprehensive fitness value corresponding to the target candidate scheduling sequence code. This ensures that the evaluation source of the target scheduling sequence can be read when the target scheduling sequence is subsequently converted into a physical reservoir control command. The gate objects in the target scheduling sequence correspond to the gate objects associated with the physical reservoir control terminal, the execution time position in the target scheduling sequence corresponds to the execution time parameter in the physical reservoir control command, and the target reservoir gate opening degree in the target scheduling sequence corresponds to the reservoir gate opening degree within the preset reservoir gate opening degree range. Thus, the target scheduling sequence can be continuously transmitted from the comprehensive fitness value, the target candidate scheduling sequence code, and the scheduling sequence optimization rules to the physical reservoir control command generation process, enabling the data results obtained from the group iterative optimization process to be transformed into the control basis for subsequent flood control scheduling actions.

[0098] Optionally, after the steps of generating time-series grid data based on the digital twin reservoir base plate and using a deep time-series neural network to perform time-series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence, the method further includes: obtaining a tiered early warning threshold; comparing and verifying the predicted water level values ​​in the predicted water level sequence with the tiered early warning threshold to obtain a predicted water level threshold verification result; generating a dynamic early warning identifier when the predicted water level value exceeds the tiered early warning threshold as indicated by the predicted water level threshold verification result; and dynamically updating the outflow amplitude constraint conditions in the scheduling optimization model according to the dynamic early warning identifier.

[0099] Preferably, after generating time-series grid data based on the digital twin reservoir base plate and using a deep time-series neural network to perform time-series prediction on the time-series grid data to obtain the predicted water level sequence and the predicted inflow sequence, a graded early warning threshold is first obtained. The graded early warning threshold is not a single water level value, but a graded threshold data structure formed for the predicted water level sequence. The graded early warning threshold includes an early warning level marker, a threshold type marker, an applicable threshold object, an applicable threshold time range, a water level threshold record, a water level interval threshold record, a water level trend threshold record, and a constraint update method marker. The warning level marker is used to express the warning level of the predicted water level value. The threshold type marker is used to distinguish between water level threshold warnings, water level interval warnings, and water level trend warnings. The applicable threshold object is used to express the reservoir water surface range, monitoring station object, or gate object corresponding to the graded warning threshold. The applicable threshold time range is used to express the scheduling time unit range in which the graded warning threshold participates in the comparison and verification. The water level threshold record is used to express the upper or lower limit of the water level under different warning levels. The water level interval threshold record is used to express the interval boundary between adjacent warning levels for the predicted water level value. The water level trend threshold record is used to express the rising or falling state of the predicted water level value within a continuous scheduling time unit. The constraint update method marker is used to express the processing method for the dynamic warning identifier to participate in the dynamic update of the discharge flow amplitude constraint conditions. The graded warning threshold continues to be read in subsequent comparison and verification so that the predicted water level values ​​in the predicted water level sequence can generate predicted water level threshold verification results according to the warning level marker, threshold type marker, and threshold applicable time range.

[0100] Preferably, the tiered early warning threshold is jointly generated from historical reservoir documents, standard structured data, and the digital twin reservoir base plate, and read upon retrieval. The historical reservoir documents provide water level control data, scheduling plan data, and historical flood peak data already generated during reservoir operation. The standard structured data provides standardized and cleaned threshold fields, time fields, and target reservoir object fields. The digital twin reservoir base plate provides the spatial nodes, spatial binding locations, and binding status markers corresponding to the applicable objects of the thresholds. When generating the tiered early warning threshold, water level fields related to water level control are first read from the standard structured data, and the water level fields are bound to the reservoir water surface range, monitoring station objects, or gate objects in the digital twin reservoir base plate according to the target reservoir object. Subsequently, the bound water level fields are organized according to early warning level markers, threshold type markers, applicable objects of the thresholds, and applicable time ranges of the thresholds to form tiered early warning thresholds that can be directly invoked from the predicted water level sequence. Therefore, the graded early warning threshold is not a temporarily input empirical threshold, but a data processing result jointly defined by the water level field, the target reservoir object, the spatial node, and the scheduling time unit; the graded early warning threshold continues to participate in the generation of the predicted water level threshold verification sample, so that the graded early warning threshold can establish a one-to-one verification relationship with the predicted water level value in the predicted water level sequence.

[0101] Preferably, before comparing and verifying the predicted water level values ​​in the predicted water level sequence with the graded early warning thresholds, a threshold verification preparation process is first performed on the predicted water level sequence. The predicted water level sequence includes predicted water level values ​​within a continuous scheduling time unit, the execution time position corresponding to the predicted water level values, and the target reservoir object corresponding to the predicted water level values. The threshold verification preparation process reads the applicable threshold objects corresponding to the target reservoir objects and reads the applicable threshold time range corresponding to the execution time position to form a predicted water level threshold verification sample. The predicted water level threshold verification sample includes predicted water level values, execution time positions, target reservoir objects, early warning level markers, threshold type markers, water level threshold records, water level interval threshold records, and water level trend threshold records. The predicted water level threshold verification sample continues to participate in the comparison and verification, so that each predicted water level value can be verified under the corresponding target reservoir object and the corresponding execution time position, instead of being compared with all graded early warning thresholds. The predicted water level threshold verification sample also retains the applicable object and the applicable time range of the threshold, so that the subsequently generated predicted water level threshold verification results can be traced back to the corresponding spatial node and execution time position.

[0102] Preferably, when comparing and verifying the predicted water level values ​​in the predicted water level sequence with the graded early warning thresholds, verification states are formed according to water level threshold early warning, water level interval early warning, and water level trend early warning. For the water level threshold early warning, the predicted water level values ​​and water level threshold records in the predicted water level threshold verification sample are read, and the predicted water level values ​​and water level threshold records are compared with each other on the same dimension to generate a water level threshold comparison state; for the water level interval early warning, the predicted water level values ​​and water level interval threshold records are read, and the early warning interval into which the predicted water level values ​​fall is determined to generate a water level interval comparison state; for the water level trend early warning, multiple predicted water level values ​​and water level trend threshold records within a continuous scheduling time unit are read, and a water level trend comparison state is generated based on the direction and magnitude of change of the predicted water level values ​​within adjacent scheduling time units. The water level threshold comparison status, the water level interval comparison status, and the water level trend comparison status together form the predicted water level threshold verification result, enabling the predicted water level threshold verification result to express three types of technical information: single-point water level, interval location, and continuous change trend. The predicted water level threshold verification result is further used to generate dynamic early warning indicators, so that the water level threshold comparison status, the water level interval comparison status, and the water level trend comparison status can all enter the subsequent constraint update processing.

[0103] Preferably, the predicted water level threshold verification result includes the target reservoir object, execution time location, predicted water level value, water level threshold comparison status, water level interval comparison status, water level trend comparison status, warning level marker, threshold type marker, and dynamic update trigger status. The water level threshold comparison status indicates whether the predicted water level value exceeds the recorded water level threshold; the water level interval comparison status indicates the warning interval to which the predicted water level value belongs; the water level trend comparison status indicates whether the predicted water level value is rising or falling within a continuous scheduling time unit; the warning level marker indicates the warning level corresponding to the predicted water level value; the threshold type marker indicates whether the predicted water level threshold verification result originates from a water level threshold warning, a water level interval warning, or a water level trend warning; and the dynamic update trigger status indicates whether the predicted water level threshold verification result needs to participate in the dynamic update of the discharge flow amplitude constraint conditions. After generating the predicted water level threshold verification result, the predicted water level threshold verification result is bound to the execution time position in the predicted water level sequence, and the predicted water level threshold verification result is further input into the dynamic warning identifier generation process, so that the subsequent dynamic warning identifier can be traced back to the specific predicted water level value, execution time position and threshold type marker.

[0104] Preferably, when the predicted water level threshold verification result indicates that the predicted water level value exceeds the graded early warning threshold, a dynamic early warning identifier is generated. The dynamic early warning identifier includes the target reservoir object, execution time location, early warning level marker, threshold type marker, water level exceedance status, trend change status, and constraint update method marker. The water level exceedance status originates from the water level threshold comparison status and is used to express the exceedance relationship between the predicted water level value and the water level threshold record. The trend change status originates from the water level trend comparison status and is used to express the direction of change of the predicted water level value within a continuous scheduling time unit. The constraint update method marker originates from the graded early warning threshold and is used to limit the processing method in which the dynamic early warning identifier subsequently participates in the dynamic update of the discharge flow amplitude constraint. The dynamic early warning identifier is not simply a prompt message, but a computable state record carrying the target reservoir object, execution time location, early warning level marker, threshold type marker, and constraint update method marker. The dynamic early warning identifier continues to enter the scheduling input processing structure in the scheduling optimization model, enabling the scheduling input processing structure to process the input before dynamically updating the discharge flow amplitude constraint based on the dynamic early warning identifier.

[0105] Preferably, after the dynamic early warning identifier is generated, early warning status merging is performed according to the early warning level marker and the threshold type marker. The early warning status merging process first reads multiple dynamic early warning identifiers generated within consecutive execution time positions for the same target reservoir object, then determines whether the multiple dynamic early warning identifiers belong to the same early warning level based on the early warning level marker, and determines whether the multiple dynamic early warning identifiers originate from water level threshold early warning, water level interval early warning, or water level trend early warning based on the threshold type marker. If multiple dynamic early warning identifiers correspond to the same target reservoir object and appear consecutively at adjacent execution time positions, then the multiple dynamic early warning identifiers are organized into a dynamic early warning identifier sequence; the dynamic early warning identifier sequence includes consecutive execution time positions, corresponding dynamic early warning identifiers, corresponding early warning level markers, and corresponding threshold type markers. The dynamic early warning identifier sequence continues to be input into the scheduling optimization model, enabling the scheduling optimization model to read the water level early warning change status within consecutive scheduling time units, rather than only reading dynamic early warning identifiers at a single execution time position; the dynamic early warning identifier sequence also continues to participate in the generation of early warning constraint update input samples, enabling the dynamic early warning identifiers within consecutive scheduling time units to jointly influence the dynamic update of the discharge flow amplitude constraint conditions.

[0106] Preferably, when dynamically updating the discharge flow variation constraint conditions in the scheduling optimization model based on the dynamic early warning identifier, the scheduling input processing structure in the scheduling optimization model first reads the dynamic early warning identifier, the dynamic early warning identifier sequence, the predicted water level sequence, the predicted inflow sequence, and the discharge flow variation constraint conditions before the update. The discharge flow variation constraint conditions before the update are the discharge flow variation constraint conditions that participated in the calculation of the scheduling optimization model before the dynamic update. The discharge flow variation constraint conditions before the update include the allowable range of discharge flow variation between adjacent scheduling time units, the discharge flow constraint object corresponding to the gate object, and the evaluation direction of discharge flow variation. The scheduling input processing structure establishes a correspondence between the dynamic early warning identifier and the predicted water level value in the predicted water level sequence, the predicted inflow value in the predicted inflow sequence, and the discharge flow variation constraint conditions before the update, based on the target reservoir object and execution time position in the dynamic early warning identifier, to form an early warning constraint update input sample. The input sample for the early warning constraint update includes a dynamic early warning identifier, a dynamic early warning identifier sequence, a predicted water level value, a predicted inflow value, a discharge flow rate variation constraint condition before the update, and the execution time position. The input sample for the early warning constraint update continues to input the discharge flow rate variation constraint condition for dynamic update processing, so that the dynamic update of the discharge flow rate variation constraint condition can be based on the water level exceedance status, trend change status, early warning level marker, and inflow flow prediction process simultaneously.

[0107] Preferably, the dynamic update process for the discharge flow amplitude constraint condition generates updated discharge flow amplitude constraint conditions based on the early warning constraint update input sample; when the dynamic early warning indicator represents that the predicted water level exceeds the higher-level graded early warning threshold and the trend change state represents that the predicted water level continues to rise within the continuous scheduling time unit, the dynamic update process for the discharge flow amplitude constraint condition reconfigures the allowable discharge flow variation range in the discharge flow amplitude constraint condition before the update, so that the candidate discharge flow sequence within the corresponding execution time position corresponds to a larger allowable discharge flow variation range during evaluation, and makes the candidate scheduling The sequence encoding can form an evaluation direction biased towards releasing reservoir capacity in the group iterative optimization structure. When the dynamic early warning indicator represents that the predicted water level falls into a lower-level warning range and the trend change status represents that the predicted water level falls back, the dynamic update processing of the discharge flow amplitude constraint condition smooths out the allowable discharge flow variation range in the discharge flow amplitude constraint condition before the update, so that the candidate discharge flow sequence at the corresponding execution time position corresponds to a smaller allowable discharge flow variation range during evaluation, and enables the candidate scheduling sequence encoding to form an evaluation direction biased towards reducing discharge fluctuations in the group iterative optimization structure. The updated discharge flow amplitude constraint condition formed after the above processing continues to enter the group iterative optimization structure and replaces the previous discharge flow amplitude constraint condition at the same execution time position in the second fitness value calculation.

[0108] Preferably, the updated discharge flow variation constraint includes the target reservoir object, the gate object, the execution time and location, the updated allowable discharge flow variation range, the updated discharge flow variation evaluation direction, the warning level marker, the threshold type marker, and the source of the dynamic warning identifier. The source of the dynamic warning identifier records which dynamic warning identifier triggered the updated discharge flow variation constraint. The warning level marker records the warning level corresponding to the updated discharge flow variation constraint. The threshold type marker records whether the updated discharge flow variation constraint originates from a water level threshold warning, a water level interval warning, or a water level trend warning. The updated allowable discharge flow variation range is used for dimensional comparison of candidate discharge flow variation states. The updated discharge flow variation evaluation direction limits the evaluation direction of candidate discharge flow variation states in fitness calculation. The updated discharge flow amplitude constraint continues to be bound to the execution time position, so that the group iterative optimization structure can read the corresponding constraint according to the time position when calculating the second fitness value; the updated discharge flow amplitude constraint also continues to be written into the early warning constraint update record, so that the source of the dynamic early warning identifier can be traced in the subsequent target scheduling sequence generation process.

[0109] Preferably, after reading the updated discharge flow amplitude constraint, the group iterative optimization structure recalculates the second fitness value based on the candidate discharge flow sequence corresponding to the candidate scheduling sequence encoding. Specifically, the group iterative optimization structure first reads the candidate discharge flow values ​​corresponding to adjacent execution time positions in the candidate discharge flow sequence to generate candidate discharge flow amplitude states; then it reads the updated discharge flow amplitude constraint corresponding to the same execution time position and compares the candidate discharge flow amplitude states with the updated allowable discharge flow change range on the same dimension to generate updated discharge flow amplitude constraint deviation states. The updated discharge flow amplitude constraint deviation states continue to participate in the calculation of the second fitness value, so that the second fitness value can reflect the influence of the dynamic warning indicator on the discharge flow amplitude constraint; the second fitness value continues to participate in the comprehensive fitness value calculation together with the first fitness value, so that the dynamic warning indicator can indirectly participate in the selection of the target scheduling sequence. The updated discharge flow amplitude constraint deviation state is also associated with the source of the dynamic early warning identifier, enabling the group iterative optimization structure to trace the source of the early warning trigger for the change in the second fitness value during the candidate scheduling sequence encoding and evaluation process.

[0110] Preferably, the dynamic early warning identifier also participates in the local update of the scheduling sequence optimization rule. After reading the early warning level marker, threshold type marker, and constraint update method marker in the dynamic early warning identifier, the scheduling sequence optimization rule performs local configuration on the candidate scheduling sequence coding selection rule and the candidate scheduling sequence coding perturbation rule. When the early warning level marker corresponds to a higher water level early warning state, the candidate scheduling sequence coding selection rule prioritizes retaining the candidate scheduling sequence codes that enable the candidate water level change sequence to meet the highest water level constraint condition and the candidate discharge flow sequence to meet the updated discharge flow amplitude constraint condition. When the early warning level marker corresponds to a lower water level early warning state, the candidate scheduling sequence coding perturbation rule makes a small adjustment to the reservoir gate opening position within the preset reservoir gate opening range to reduce the change of the candidate discharge flow sequence between adjacent execution time positions. Through the above processing, the dynamic early warning identifier not only updates the discharge flow amplitude constraint, but also enters the candidate scheduling sequence coding screening process in the scheduling sequence optimization rule; the candidate scheduling sequence coding screening process continues to generate target candidate scheduling sequence codes, so that the dynamic early warning identifier can participate in the generation of the target scheduling sequence through the scheduling sequence optimization rule.

[0111] Preferably, after the dynamic update of the discharge flow amplitude constraint is completed, the dynamic early warning identifier, the dynamic early warning identifier sequence, the predicted water level threshold verification result, the updated discharge flow amplitude constraint, and the corresponding execution time position are jointly written into the early warning constraint update record. The early warning constraint update record is used to express the source relationship between the predicted water level threshold verification result, the dynamic early warning identifier, the dynamic early warning identifier sequence, and the updated discharge flow amplitude constraint. The early warning constraint update record continues to input the candidate scheduling deduction structure and the group iterative optimization structure in the scheduling optimization model, so that the candidate scheduling deduction structure can read the dynamically updated constraint boundary when deducing the candidate discharge flow sequence, and the group iterative optimization structure can read the dynamically updated evaluation basis when calculating the second fitness value and the comprehensive fitness value. Therefore, the predicted water level sequence is compared with the graded early warning threshold to form the predicted water level threshold verification result. The predicted water level threshold verification result generates a dynamic early warning identifier. The dynamic early warning identifier further forms an updated discharge flow amplitude constraint condition. The updated discharge flow amplitude constraint condition enters the second fitness numerical calculation and the comprehensive fitness numerical calculation, so that the graded early warning processing result can enter the subsequent target scheduling sequence generation process.

[0112] Optionally, before the step of converting the target scheduling sequence into physical reservoir control instructions, the method further includes: importing the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence into a digital twin reservoir 3D rendering engine; controlling a virtual gate model to perform motion simulation in the digital twin reservoir base according to the target scheduling sequence, and providing an initial water level state and an inflow driving state for the motion simulation of the virtual gate model according to the predicted water level sequence and the predicted inflow sequence; generating a virtual simulation screen of water surface evolution using a pixel-stream rendering mechanism; performing flood control safety verification on the target scheduling sequence under the constraints of the highest water level and the variation of the outflow based on the virtual simulation screen of water surface evolution, and obtaining a flood control safety verification result; and executing the step of converting the target scheduling sequence into physical reservoir control instructions when the flood control safety verification result is verified as passed.

[0113] Preferably, in the specific implementation process of importing the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence into the digital twin reservoir 3D rendering engine, the spatial topology binding result in the digital twin reservoir base plate is first read, and the spatial nodes of the reservoir water surface range, upstream catchment range, inflow river cross section, spillway, water conveyance pipeline, gate object, station object, and river cross section in the three-dimensional spatial coordinate system are determined according to the spatial topology binding result. The digital twin reservoir 3D rendering engine is used to read the digital twin 3D spatial objects and scheduling dynamic status data in the digital twin reservoir base plate, and organize the digital twin 3D spatial objects and the scheduling dynamic status data into a renderable scheduling pre-simulation 3D scene; the digital twin 3D spatial objects include the reservoir 3D terrain, reservoir water surface range, river cross section, spillway, water conveyance pipeline, gate object, and virtual gate model, and the scheduling dynamic status data includes the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence. During import, the execution time position in the target scheduling sequence, the predicted water level value in the predicted water level sequence, and the predicted inflow value in the predicted inflow sequence are first time-aligned according to the grid time reference to form a scheduling pre-simulation input record. The scheduling pre-simulation input record is then written into the digital twin reservoir 3D rendering engine, so that when the virtual gate model simulates actions, it can read the corresponding target reservoir gate opening degree, predicted water level value, and predicted inflow value within the same scheduling time unit, and enable the scheduling pre-simulation 3D scene to simultaneously inherit the spatial nodes of the digital twin 3D spatial object and the time position of the scheduling dynamic status data.

[0114] Preferably, the pre-scheduling input record includes a target scheduling sequence, a predicted water level sequence, a predicted inflow sequence, a scheduling time unit, a gate object, a target reservoir gate opening degree, a reservoir water surface area, an inflow river cross-section, and a spatial binding location. The target scheduling sequence expresses the gate opening degree arrangement of the target reservoir gate within a continuous scheduling time unit; the predicted water level sequence expresses the predicted water level value of the reservoir water surface area or a monitoring station within a continuous scheduling time unit; the predicted inflow sequence expresses the predicted inflow value of the inflow river cross-section within a continuous scheduling time unit; and the spatial binding location binds the gate object, the reservoir water surface area, and the inflow river cross-section to a spatial node in the digital twin reservoir base plate. After the scheduling pre-simulation input record enters the digital twin reservoir 3D rendering engine, the digital twin reservoir 3D rendering engine reads the scheduling pre-simulation input record frame by frame according to the scheduling time unit, and writes the data corresponding to each scheduling time unit into the scheduling pre-simulation 3D scene cache; the scheduling pre-simulation 3D scene cache continues to participate in the virtual gate model action simulation and water surface evolution virtual simulation screen generation, so that the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence are not discrete data displayed separately, but data input that can be jointly pre-simulated under the same scheduling pre-simulation 3D scene and the same grid time reference.

[0115] Preferably, when simulating the action of a virtual gate model controlled according to the target scheduling sequence in the digital twin reservoir base plate, the virtual gate model corresponding to the gate object is first read according to the spatial topology binding result. The virtual gate model is a three-dimensional, driveable representation of the gate object in the digital twin reservoir base plate. The virtual gate model includes spatial nodes of the gate object, gate opening direction, gate opening degree mapping relationship, and gate action state. The gate opening degree mapping relationship is used to convert the target reservoir gate opening degree in the target scheduling sequence into the virtual gate opening position of the virtual gate model in the three-dimensional spatial coordinate system. The gate action state is used to express the virtual gate opening position of the virtual gate model in the corresponding scheduling time unit and the opening change between adjacent scheduling time units. During the motion simulation, the digital twin reservoir 3D rendering engine reads the target reservoir gate opening degree in the target scheduling sequence according to the scheduling time unit, and updates the position of the virtual gate model according to the gate opening degree mapping relationship to form a virtual gate action sequence. The virtual gate action sequence continues to participate in the pre-simulation of the discharge flow visualization processing, so that the gate scheduling action corresponding to the target scheduling sequence can form a continuous 3D action process in the bottom plate of the digital twin reservoir, and the virtual gate opening position of the virtual gate model can be read by the subsequent water surface evolution state calculation.

[0116] Preferably, when providing the initial water level state and inflow driving state for the virtual gate model's action simulation based on the predicted water level sequence and the predicted inflow sequence, the predicted water level values ​​corresponding to the reservoir water surface area, monitoring station objects, and gate objects in the predicted water level sequence are first read, and these predicted water level values ​​are written into the initial water level state in the scheduling pre-simulation 3D scene cache. The initial water level state is used to express the water level height and water surface boundary position of the reservoir water surface area at the beginning of the current scheduling time unit. Subsequently, the predicted inflow values ​​corresponding to the upstream catchment area and inflow river cross-section in the predicted inflow sequence are read, and these predicted inflow values ​​are written into the inflow driving state in the scheduling pre-simulation 3D scene cache. The inflow driving state is used to express the flow input process entering the reservoir water surface area within the current scheduling time unit. The initial water level state and the inflow-driven state jointly participate in the calculation of the water surface evolution state, so that the virtual gate model action simulation can be simultaneously constrained by the target reservoir gate opening degree, the predicted water level value, and the predicted inflow value, rather than simply displaying the gate object statically according to the target scheduling sequence.

[0117] Preferably, the water surface evolution state calculation takes the initial water level state, the inflow-driven state, and the virtual gate action sequence as inputs, and generates a water surface evolution state record according to the scheduling time unit. The water surface evolution state record includes the scheduling time unit, the reservoir water surface range, the predicted water level value, the predicted inflow value, the target reservoir gate opening degree, the pre-simulated discharge flow value, and the water surface boundary change state; wherein, the pre-simulated discharge flow value is derived from the discharge state jointly determined by the target reservoir gate opening degree corresponding to the virtual gate action sequence and the predicted water level value, and the water surface boundary change state is used to express the water surface range of the reservoir area within the continuous scheduling time unit, whether the water surface rises, falls, or remains constant. After generating the water surface evolution status record, the digital twin reservoir 3D rendering engine writes the water surface evolution status record into the spatial nodes corresponding to the reservoir water surface range, so that a readable pre-simulation relationship is formed between the reservoir water surface range, gate objects, inflow river cross sections and scheduling time units; the water surface evolution status record continues to enter the pixel stream rendering mechanism to generate a virtual simulation screen of water surface evolution, and enables the pre-simulated discharge flow value to establish a correspondence with the discharge flow amplitude constraint conditions in subsequent flood control safety verification.

[0118] Preferably, when using a pixel-stream rendering mechanism to generate a virtual simulation of water surface evolution, the pixel-stream rendering mechanism first reads the water surface evolution state record, the virtual gate action sequence, and the digital twin 3D spatial object in the digital twin reservoir bottom plate. It then maps the water surface evolution state record to the water surface evolution rendering state of the reservoir area, and maps the virtual gate action sequence to the virtual gate opening rendering state of the gate object. Subsequently, the pixel-stream rendering mechanism generates continuous image frames according to the scheduling time unit. Each continuous image frame includes the reservoir water surface area, virtual gate model, inflow river cross-section, spillway, water conveyance pipeline, and corresponding water surface boundary change state under the corresponding scheduling time unit. The continuous image frames are then encoded into a virtual simulation of water surface evolution in chronological order. The technical essence of the pixel stream rendering mechanism lies in outputting the three-dimensional rendering results in the form of a pixel frame sequence, so that the virtual simulation screen of water surface evolution can express the water level changes and gate action changes within a continuous scheduling time unit; the virtual simulation screen of water surface evolution continues to participate in flood control safety verification, so that the flood control safety verification can read the visualized water surface evolution state corresponding to the spatial node and scheduling time unit, and the visualized water surface evolution state is jointly formed by the water surface evolution rendering state and the virtual gate opening rendering state.

[0119] Preferably, when performing flood control safety verification of the target scheduling sequence under the constraints of the highest water level and the discharge flow variation based on the virtual simulation image of water surface evolution, the water surface evolution status record is first read from the continuous image frames corresponding to the virtual simulation image of water surface evolution, and the predicted water level value, water surface boundary change status, target reservoir gate opening degree, and pre-simulated discharge flow value corresponding to each scheduling time unit are extracted from the water surface evolution status record. Subsequently, the predicted water level value is compared with the highest water level constraint in the same dimension to generate a water level safety verification status; the change in pre-simulated discharge flow value between adjacent scheduling time units is compared with the discharge flow variation constraint in the same dimension to generate a discharge flow safety verification status. The water level safety verification status is used to express whether the water level evolution process corresponding to the target scheduling sequence meets the highest water level constraint condition, and the discharge flow safety verification status is used to express whether the discharge flow change process corresponding to the target scheduling sequence meets the discharge flow amplitude constraint condition; the water level safety verification status and the discharge flow safety verification status together form the flood control safety verification result, and enable the flood control safety verification result to simultaneously reflect the water surface evolution process and the discharge flow change process.

[0120] Preferably, the flood control safety verification result includes a target scheduling sequence, scheduling time unit, water level safety verification status, discharge flow safety verification status, constraint deviation position, constraint deviation type, and verification status marker. The constraint deviation position indicates the execution time location where the highest water level constraint or the discharge flow amplitude constraint is not met. The constraint deviation type indicates that the constraint deviation originates from the predicted water level value, water surface boundary change state, or pre-simulated discharge flow value. The verification status marker indicates whether the target scheduling sequence passes the flood control safety verification. After generating the flood control safety verification result, it is associated with the water surface evolution virtual simulation screen, the water surface evolution status record, and the target scheduling sequence, enabling the flood control safety verification result to be traced back to the gate object, execution time location, and target reservoir gate opening degree of the target scheduling sequence. The flood control safety verification result continues to participate in the judgment of whether to execute the conversion of the target scheduling sequence into a physical reservoir control command, and allows the constraint deviation position and constraint deviation type to continue participating in the generation of scheduling rollback verification records when verification fails.

[0121] Preferably, when the verification status marker in the flood control safety verification result indicates that the verification has passed, the step of converting the target scheduling sequence into a physical reservoir control command is executed. Before execution, the target scheduling sequence, gate object, execution time position, target reservoir gate opening degree, water level safety verification status, and discharge flow safety verification status in the flood control safety verification result are read first, and the water level safety verification status and the discharge flow safety verification status are used as the verification basis for the target scheduling sequence to enter the physical execution process. Subsequently, the gate object in the target scheduling sequence is mapped to a control object identifier that the physical reservoir control terminal can recognize, the execution time position is mapped to an execution time parameter that the physical reservoir control terminal can recognize, and the target reservoir gate opening degree is mapped to a reservoir gate opening degree control parameter that the physical reservoir control terminal can recognize, so as to form a physical reservoir control command. The physical reservoir control command includes the control object identifier, the execution time parameter, and the reservoir gate opening degree control parameter. The control object identifier, the execution time parameter, and the reservoir gate opening degree control parameter together define the control object, execution time, and reservoir gate opening degree of the subsequent flood control scheduling actions executed by the physical reservoir control terminal, so that the target scheduling sequence enters the physical execution process after 3D rendering pre-simulation and flood control safety verification.

[0122] Preferably, when the verification status marker in the flood control safety verification result indicates that the verification has failed, the step of converting the target scheduling sequence into a physical reservoir control command is not executed. Instead, a scheduling rollback verification record is generated based on the constraint deviation position and constraint deviation type in the flood control safety verification result. The scheduling rollback verification record includes the target scheduling sequence, constraint deviation position, constraint deviation type, water level safety verification status, discharge flow safety verification status, and the corresponding water surface evolution status record. The scheduling rollback verification record continues to be fed back to the group iterative optimization structure in the scheduling optimization model, enabling the group iterative optimization structure to adjust the evaluation direction of the candidate scheduling sequence encoding according to the constraint deviation position and constraint deviation type, and regenerate the target scheduling sequence that meets the flood control safety verification requirements. Therefore, the virtual simulation screen of water surface evolution is not only used to display the scheduling pre-rehearsal results, but also participates in the verification process before the target scheduling sequence enters physical execution through the flood control safety verification results and the scheduling rollback verification record; the scheduling rollback verification record continues to retain the correspondence between the target scheduling sequence and the water surface evolution status record, so that the group iterative optimization structure can read the water surface evolution process and the discharge flow change process that failed the verification.

[0123] Preferably, after the above processing, the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence first enter the digital twin reservoir 3D rendering engine to form the scheduling pre-simulation input record; the scheduling pre-simulation input record continues to drive the virtual gate model to generate the virtual gate action sequence, and combines the initial water level state and the inflow driving state to generate the water surface evolution state record; the water surface evolution state record forms the water surface evolution virtual simulation screen through the pixel stream rendering mechanism; the water surface evolution virtual simulation screen continues to be used to generate the flood control safety verification result; when the flood control safety verification result characterization verification passes, the target scheduling sequence is then converted into the physical reservoir control command. Through this process, before the target scheduling sequence enters the physical reservoir control terminal, it has already completed a three-dimensional pre-simulation and flood control safety verification with the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, and the outflow amplitude constraint. This ensures that subsequent flood control scheduling actions have a data source corresponding to the scheduling pre-simulation input record, a spatial source corresponding to the spatial topology binding result, a time source corresponding to the grid time reference, and a constraint verification source corresponding to the flood control safety verification result.

[0124] Optionally, after the step of controlling the physical reservoir control terminal to perform flood control scheduling actions according to the physical reservoir control command, the method further includes: acquiring real-time measured water level data fed back by the physical reservoir control terminal; performing time-series alignment processing on the real-time measured water level data according to the grid time reference to obtain a real-time measured water level sequence corresponding to the predicted water level sequence; calculating the predicted water level error digital sequence between the predicted water level sequence and the real-time measured water level sequence; extracting the error change characteristics of the predicted water level error digital sequence using an autoregressive algorithm; and correcting the parameters of the deep temporal neural network according to the error change characteristics.

[0125] Preferably, after the physical reservoir control terminal is controlled to perform flood control scheduling actions according to the physical reservoir control command, the real-time measured water level data fed back by the physical reservoir control terminal is first acquired; the real-time measured water level data comes from the water level feedback record formed by the physical reservoir control terminal after executing the physical reservoir control command, and the real-time measured water level data includes the target reservoir object, the station object, the spatial node, the measured water level value, the measured acquisition time, the gate object, the execution status of the physical reservoir control command, and the feedback status marker. The target reservoir object and the monitoring station object are used to bind the real-time measured water level data to the spatial topology binding result in the digital twin reservoir base plate. The spatial node is used to express the spatial source of the real-time measured water level data in the three-dimensional spatial coordinate system. The measured acquisition time is used to express the acquisition time source of the real-time measured water level data entering the feedback processing process. The physical reservoir control command execution status is used to express whether the gate object corresponding to the physical reservoir control command has executed the target reservoir gate opening degree according to the target scheduling sequence. The feedback status flag is used to express whether the real-time measured water level data can participate in subsequent time sequence alignment processing. After reading the real-time measured water level data, a correspondence is established between the real-time measured water level data and the physical reservoir control command, the target scheduling sequence, and the predicted water level sequence, so that subsequent error calculation can be traced back to the corresponding gate object, execution time position, and predicted water level value.

[0126] Preferably, after acquiring the real-time measured water level data, the real-time measured water level data undergoes feedback data normalization processing. This normalization process first reads the measured water level value, measured acquisition time, target reservoir object, monitoring station object, spatial node, and feedback status marker from the real-time measured water level data. Then, based on the target reservoir object and the monitoring station object, it checks whether the measured water level value corresponds to the reservoir water surface range or monitoring station object described by the predicted water level sequence. For real-time measured water level data that can participate in subsequent processing, represented by feedback status markers, the measured water level value is converted to the same water level measurement caliber as the predicted water level value in the predicted water level sequence. The converted measured water level value, along with the target reservoir object, the monitoring station object, the spatial node, the measured acquisition time, and the feedback status marker, is organized into a normalized measured water level record. This normalized measured water level record continues to participate in time-series alignment processing, ensuring that the real-time measured water level data corresponds to the predicted water level sequence in terms of water level measurement caliber, spatial source, and object attribution before entering error calculation.

[0127] Preferably, when performing time-series alignment processing on the real-time measured water level data according to the grid time reference, the measured acquisition time in the standardized measured water level record is first read, and the execution time position to which the measured acquisition time belongs is determined according to the grid time reference. Subsequently, according to the target reservoir object, the station object, and the execution time position, the standardized measured water level record is written into the execution time position corresponding to the predicted water level sequence to form a real-time measured water level sequence. Each sequence position in the real-time measured water level sequence corresponds to an execution time position, and each sequence position in the real-time measured water level sequence records the measured water level value, target reservoir object, station object, spatial node, measured acquisition time, and feedback status flag at that execution time position. The real-time measured water level sequence and the predicted water level sequence use the same grid time reference, the same target reservoir object, and the same execution time position, so that water level values ​​at different execution time positions or different spatial nodes will not be mixed and compared when calculating the predicted water level error digital sequence later.

[0128] Preferably, when generating the real-time measured water level sequence, a feedback gap marking process is also performed on the real-time measured water level sequence. The feedback gap marking process reads the execution time position in the predicted water level sequence and checks whether the real-time measured water level sequence has a measured water level value at the corresponding execution time position. If the real-time measured water level sequence lacks a measured water level value at a certain execution time position, a feedback gap mark is written at that execution time position, and the feedback gap mark is bound to the corresponding target reservoir object, station object, spatial node, and execution time position. If the real-time measured water level sequence has a measured water level value at a certain execution time position, an error calculation relationship is established between the measured water level value and the predicted water level value at the same execution time position in the predicted water level sequence. The feedback gap mark continues to participate in the generation of the predicted water level error sequence, ensuring that the execution time position lacking a measured water level value is not mistakenly identified as the true source of error between the predicted water level sequence and the real-time measured water level sequence.

[0129] Preferably, when calculating the predicted water level error sequence between the predicted water level sequence and the real-time measured water level sequence, the predicted water level values ​​in the predicted water level sequence are first read one by one according to the target reservoir object, the measuring station object, and the execution time position. Then, the measured water level values ​​for the same target reservoir object, the same measuring station object, and the same execution time position in the real-time measured water level sequence are read. Subsequently, the difference in dimensionality between the predicted water level values ​​and the measured water level values ​​at the same execution time position is calculated to obtain the predicted water level error value corresponding to that execution time position. The predicted water level error values ​​are then arranged in chronological order according to the execution time positions to form the predicted water level error sequence. Each sequence position in the predicted water level error sequence includes the predicted water level error value, the target reservoir object, the measuring station object, a spatial node, the execution time position, the predicted water level value, the measured water level value, and a feedback status flag, so that the predicted water level error sequence expresses both the magnitude of the prediction deviation and the source of the prediction deviation.

[0130] Preferably, after the predicted water level error digital sequence is formed, it undergoes error direction identification processing and error continuity processing. The error direction identification processing reads the predicted water level error value corresponding to each execution time position and generates an error direction state based on the positive or negative deviation of the predicted water level error value. The error continuity processing reads multiple predicted water level error values ​​within consecutive execution time positions and generates a continuous error change state based on the increase or decrease relationship between these multiple predicted water level error values. The error direction state and the continuous error change state are then written into the predicted water level error digital sequence, so that the predicted water level error digital sequence not only expresses the predicted water level error value at a single execution time position but also expresses the error evolution direction within consecutive execution time positions. Therefore, the subsequent autoregressive algorithm can read the error change basis with a temporal relationship, rather than only adjusting parameters based on the predicted water level error value at a certain moment.

[0131] Preferably, when using an autoregressive algorithm to extract the error change features of the predicted water level error digital sequence, multiple predicted water level error values ​​prior to the current execution time position are first read from the predicted water level error digital sequence, and these multiple predicted water level error values ​​are arranged according to the grid time reference to form an error backtracking segment; the error backtracking segment is used to express the change trajectory of the predicted water level error digital sequence within the previous execution time position. The autoregressive algorithm performs time correlation analysis on the error backtracking segment, uses the predicted water level error values ​​within the previous execution time position as a reference input for the error state at the current execution time position, and generates an error regression weight state based on the degree of influence of the previous predicted water level error values ​​on the current predicted water level error values; the error regression weight state is further combined with the error direction state and the error continuous change state to form error change features. The error change features include the error regression weight state, the error direction state, the error continuous change state, and the error residual state. The error residual state is used to express the remaining error change portion that cannot be explained by the previous predicted water level error values, and the error change features continue to serve as the input basis for parameter correction of the deep temporal neural network.

[0132] Preferably, the error variation features continue to establish a correspondence with the time series prediction process of the deep temporal neural network; the deep temporal neural network includes a time series input processing layer, a spatial node embedding layer, a rainfall-flow coupling layer, a long short-term memory unit, a prediction output layer, and a sequence processing layer. When correcting the parameters of the deep temporal neural network based on the error variation features, the error variation features are first aligned with the predicted water level sequence according to the execution time position, and the aligned error variation features are written into the network correction input record; the network correction input record includes the error variation features, the predicted water level sequence, the real-time measured water level sequence, the rainfall feature vector, the flow feature vector, and the time delay feature expression result. The network correction input record continues to enter the deep temporal neural network, enabling the deep temporal neural network to adjust the subsequent time series prediction process according to the error variation features between the predicted water level sequence and the real-time measured water level sequence.

[0133] Preferably, when correcting the parameters of the deep temporal neural network, the temporal input processing layer first reads the error change features, rainfall feature vector, and flow feature vector from the network correction input record, and supplements the error change features into the temporal input sample at the corresponding execution time position according to the grid time reference to form an error-enhanced temporal input sample. The spatial node embedding layer reads the target reservoir object, spatial node, and spatial binding position from the error-enhanced temporal input sample, and converts the spatial node and spatial binding position into a spatial node embedding expression; the spatial node embedding expression continues to be input into the rainfall-flow coupling layer along with the error-enhanced temporal input sample. The rainfall-flow coupling layer reads the rainfall state, flow state, and error change features from the error-enhanced temporal input sample, and fuses the rainfall state, flow state, and error change features accordingly to form an error-enhanced rainfall-flow coupling expression; the error-enhanced rainfall-flow coupling expression continues to be input into the long short-term memory unit, enabling the long short-term memory unit to read the error change features formed by real-time measured water level data feedback on the basis of the original rainfall-flow coupling.

[0134] Preferably, after the Long Short-Term Memory (LSTM) unit receives the error-enhanced rainfall-flow coupled expression, the input gate processing structure reads the input state related to subsequent water level changes from the error-enhanced rainfall-flow coupled expression, and takes the error direction state and error continuous change state in the error change features as part of the input state; the forget gate processing structure reads the historical memory state formed at the previous execution time position, and determines the delay impact information that needs to be retained in the historical memory state according to the error regression weight state in the error change features; the candidate state processing structure generates candidate memory states according to the error-enhanced rainfall-flow coupled expression, and the candidate memory states are used to express the new impact of the current rainfall state, the current flow state, and the predicted water level error sequence on the subsequent predicted water level sequence; the state update processing structure merges the historical memory state, the input state, and the candidate memory state to form a corrected current memory state; the output gate processing structure generates a corrected time delay feature expression result according to the corrected current memory state. The corrected time delay feature expression result continues to be input into the prediction output layer, enabling the prediction output layer to generate subsequent predicted water level sequences and subsequent predicted inflow sequences based on the error change features after feedback from real-time measured water level data.

[0135] Preferably, when the prediction output layer corrects parameters based on the corrected time delay feature expression result, it reads the prediction water level error values ​​related to the reservoir water surface range, station objects, and gate objects in the prediction water level error digital sequence, and maps the prediction water level error values ​​to the corrected time delay feature expression result to adjust the output bias state of the water level prediction segment in the prediction output layer; the output bias state is used to express the water level deviation direction that the prediction output layer needs to correct when generating subsequent water level prediction segments. Subsequently, the output bias state is written into the parameter correction record of the prediction output layer, and the prediction output layer generates the corrected water level prediction segment based on the parameter correction record and the corrected time delay feature expression result; the sequence organizing layer organizes the corrected water level prediction segment into a corrected prediction water level sequence according to the grid time reference, and continues to form a new prediction water level error digital sequence with the subsequent real-time measured water level sequence. Through this process, the parameter correction does not directly replace the prediction result of the deep time series neural network. Instead, it transmits the error change features extracted from the digital sequence of predicted water level error layer by layer to the time series input processing layer, the rainfall-flow coupling layer, the long short-term memory unit, the prediction output layer, and the sequence processing layer. This enables the subsequent time series prediction process of the deep time series neural network to read the error change features formed by the feedback of real-time measured water level data.

[0136] Preferably, after correcting the parameters of the deep time-series neural network based on the error change characteristics, the error change characteristics, the corrected time delay feature expression result, the corrected predicted water level sequence, and the corresponding real-time measured water level sequence are jointly written into a prediction correction record. The prediction correction record includes the target reservoir object, the station object, the spatial node, the execution time location, the predicted water level error numerical sequence, the error change characteristics, the network correction input record, the corrected time delay feature expression result, and the corrected predicted water level sequence. The prediction correction record continues to participate in subsequent time-series prediction and scheduling optimization model input updates, enabling the parameter correction result of the deep time-series neural network to be transmitted to the predicted water level sequence generation process, and allowing the subsequent scheduling optimization model to read the predicted water level sequence corrected by the real-time measured water level data feedback. Thus, the real-time measured water level data, the real-time measured water level sequence, the predicted water level error numerical sequence, the error change characteristics, and the parameter correction process of the deep time-series neural network form a continuous data processing chain, preventing the data fed back from the physical reservoir control terminal from remaining at the simple display or recording level.

[0137] Preferably, before using the deep temporal neural network for time series prediction, training samples for training the deep temporal neural network are first constructed. These training samples originate from historical data such as the digital twin reservoir floor, historical time-series grid data, historical real-time measured water level sequences, historical measured inflow sequences, historical measured outflow sequences, historical gate feedback data, and historical flood peak data. When constructing the training samples, the spatial topology binding results in the digital twin reservoir floor are first read, and the rainfall data, water level data, inflow data, outflow data, and gate feedback data in the spatial topology binding results are organized into historical time-series grid data according to the grid time reference. The historical time-series grid data undergoes outlier detection and abnormal data unit removal processing to form historical time-series grid data with removed abnormal data units. This historical time-series grid data with removed abnormal data units continues to serve as the input data source for the training samples, ensuring that the input content in subsequent training samples maintains the same data organization caliber as the time-series grid data with removed abnormal data units used in the aforementioned time series prediction stage.

[0138] Preferably, the training samples include training input samples and training label samples; the training input samples are derived from the historical time-series grid data of the removed abnormal data units, and the training label samples are derived from the historical real-time measured water level sequence and the historical measured inflow sequence. When generating the training input samples, historical rainfall data, historical inflow data, historical outflow data, and historical gate feedback data from the historical time-series grid data of the removed abnormal data units are first read according to continuous scheduling time units. Then, the historical rainfall data, historical inflow data, historical outflow data, and historical gate feedback data are arranged chronologically according to spatial nodes, target reservoir objects, and execution time positions to form historical time-series input segments. These historical time-series input segments are further used to extract historical rainfall feature vectors and historical flow feature vectors, enabling the training input samples to correspond in data source, temporal arrangement, and spatial binding relationship to the rainfall feature vectors and flow feature vectors input into the deep time-series neural network during actual prediction.

[0139] Preferably, when extracting the historical rainfall feature vector, historical rainfall data corresponding to the upstream catchment area, reservoir area, and monitoring station are first read from the historical time-series input segments, and the historical rainfall data is organized into historical rainfall time-series segments according to continuous scheduling time units. The historical rainfall time-series segments include historical rainfall intensity, historical cumulative rainfall status, historical rainfall duration status, and spatial location status of the monitoring station. Subsequently, the historical rainfall time-series segments are processed by rainfall lag windows to form the historical rainfall feature vector. The historical rainfall feature vector continues to serve as part of the training input samples, enabling the time-series input processing layer to read the changes in historical rainfall status within continuous scheduling time units, and enabling the Long Short-Term Memory (LSTM) unit to extract the delayed impact of historical rainfall status on subsequent water level changes and subsequent inflow changes during the training phase.

[0140] Preferably, when extracting the historical flow feature vector, historical inflow data, historical outflow data, and historical gate feedback data corresponding to the inflow river cross-section, spillway, water conveyance pipeline, and gate objects are first read from the historical time-series input segments. These historical inflow data, historical outflow data, and historical gate feedback data are then organized into historical flow time-series segments according to continuous scheduling time units. Each historical flow time-series segment includes historical inflow change states, historical outflow change states, historical gate feedback change states, and cross-sectional spatial location states. Subsequently, the historical flow time-series segments are processed using flow change windows to form the historical flow feature vector. This historical flow feature vector, together with the historical rainfall feature vector, forms the training input sample, enabling the rainfall-flow coupling layer to read the correspondence between historical rainfall states and historical flow states during the training phase.

[0141] Preferably, when generating the training label samples, firstly, the historical measured water level values ​​corresponding to the reservoir water surface area, monitoring station objects, and gate objects in the historical real-time measured water level sequence are read, and the historical measured water level values ​​are organized into a historical target water level sequence according to the grid time reference. Each sequence position in the historical target water level sequence corresponds to an execution time position, and the historical measured water level value, target reservoir object, monitoring station object, and spatial node at that execution time position are recorded. Subsequently, the historical measured inflow values ​​corresponding to the upstream catchment area and inflow river cross-section in the historical measured inflow sequence are read, and the historical measured inflow values ​​are organized into a historical target inflow sequence according to the grid time reference. Each sequence position in the historical target inflow sequence corresponds to an execution time position, and the historical measured inflow value, target reservoir object, and inflow river cross-section at that execution time position are recorded. The historical target water level sequence and the historical target inflow sequence together constitute the training label sample, so that the prediction output layer and the sequence sorting layer have comparable output references during the training phase.

[0142] Preferably, to establish a traceable temporal correspondence between the training input samples and the training label samples, the continuous scheduling time unit is first divided into a historical input time range and a future prediction time range when constructing the training samples. The historical input time range carries the historical rainfall feature vector and the historical flow feature vector, while the future prediction time range carries the historical target water level sequence and the historical target inflow sequence. The historical input time range and the future prediction time range are connected sequentially according to the grid time reference, so that the historical rainfall and historical flow states in the training input samples can correspond to the subsequent water level changes and subsequent inflow changes in the training label samples. Thus, the training samples do not simply pair input data at the same moment with output data at the same moment, but rather express the temporal delay transmission relationship between the rainfall confluence process, the river transport process, and the reservoir storage and release process.

[0143] Preferably, when the historical flood peak data has already undergone cluster analysis to form a similar flood peak reference sequence, the similar flood peak reference sequence corresponding to the training input sample is also written into the training sample. During writing, historical flood peak morphological features are first extracted based on the historical rainfall feature vector and historical flow feature vector within the historical input time range. Then, the historical flood peak morphological features are matched with the flood peak morphological features corresponding to the already formed similar flood peak reference sequence to select the similar flood peak reference sequence corresponding to the training input sample. The similar flood peak reference sequence, together with the historical rainfall feature vector and the historical flow feature vector, forms an enhanced training input sample, enabling the time-series input processing layer to simultaneously read historical monitoring status and the time evolution reference of historically similar flood peak processes during the training phase, and enabling the long short-term memory unit to extract the difference between the current flood peak process and the similar flood peak reference sequence.

[0144] Preferably, when training the deep temporal neural network using the training samples, the temporal input processing layer first reads the training input samples or the enhanced training input samples, and organizes the historical rainfall feature vector, the historical flow feature vector, and the similar flood peak reference sequence into training temporal input samples according to the grid time reference. Each sample unit in the training temporal input samples corresponds to an execution time position and includes historical rainfall status, historical flow status, historical gate feedback change status, spatial nodes, spatial binding positions, and target reservoir objects. The training temporal input samples are then input into the spatial node embedding layer, which transforms the spatial nodes, spatial binding positions, and target reservoir objects in the training temporal input samples into spatial node embedding expressions. The spatial node embedding expressions are then input into the rainfall-flow coupling layer along with the training temporal input samples, so that the subsequent training process simultaneously preserves the temporal sequence relationship and the spatial topological binding relationship.

[0145] Preferably, after receiving the training time-series input samples and the spatial node embedding expression, the rainfall-flow coupling layer fuses the historical rainfall state in the historical rainfall feature vector, the historical flow state in the historical flow feature vector, the reference flood peak state in the similar flood peak reference sequence, and the spatial node embedding expression to form a training rainfall-flow coupling expression. The training rainfall-flow coupling expression is then input into the Long Short-Term Memory (LSTM) unit. The LSM unit's input gate processing structure, forget gate processing structure, candidate state processing structure, state update processing structure, and output gate processing structure sequentially extract time-delay features from the training rainfall-flow coupling expression to form a training time-delay feature expression result. The training time-delay feature expression result is then input into the prediction output layer, enabling the prediction output layer to generate training water level prediction segments and training inflow prediction segments based on the time-delay relationships during the training phase.

[0146] Preferably, after the prediction output layer generates the training water level prediction segment and the training inflow prediction segment, the sequence organizing layer organizes the training water level prediction segment into a training prediction water level sequence and the training inflow prediction segment into a training prediction inflow sequence according to the grid time reference. Subsequently, the training prediction water level sequence is compared with the historical target water level sequence in the training label sample to form a water level training deviation state; the training prediction inflow sequence is compared with the historical target inflow sequence in the training label sample to form an inflow training deviation state. The water level training deviation state and the inflow training deviation state continue to participate in the calculation of the training loss state, so that the training process of the deep temporal neural network can evaluate the water level prediction deviation and the inflow prediction deviation separately, rather than evaluating the entire network based on a single output result.

[0147] Preferably, the training loss states include water level training deviation states, inflow training deviation states, trend consistency deviation states, and time continuity deviation states. The trend consistency deviation state expresses the deviation of the training predicted water level sequence from the historical target water level sequence during the rising, falling, or remaining stable process. The time continuity deviation state expresses whether there are jumps between adjacent execution time positions in the training predicted water level sequence and the training predicted inflow sequence that do not conform to the reservoir's temporal change pattern. When calculating the training loss states, the water level training deviation states and the inflow training deviation states are first read separately, and then the training predicted water level sequence, the historical target water level sequence, the training predicted inflow sequence, and the historical target inflow sequence within consecutive execution time positions are read to form the trend consistency deviation state and the time continuity deviation state. The training loss states are further used to adjust the trainable parameters in the deep temporal neural network, and the adjustment process of the trainable parameters is simultaneously constrained by water level numerical deviation, inflow numerical deviation, trend direction deviation, and time continuity deviation.

[0148] Preferably, when adjusting the trainable parameters in the deep temporal neural network according to the training loss state, the training loss state is first propagated in reverse order along the sequence arrangement layer, the prediction output layer, the long short-term memory unit, the rainfall-flow coupling layer, the spatial node embedding layer, and the temporal input arrangement layer to obtain the parameter correction state corresponding to each structural layer. Subsequently, the spatial node embedding parameters in the spatial node embedding layer, the rainfall-flow coupling parameters in the rainfall-flow coupling layer, the gating parameters in the long short-term memory unit, the prediction output parameters in the prediction output layer, and the sequence arrangement parameters in the sequence arrangement layer are adjusted according to the parameter correction state. The spatial node embedding parameters affect the spatial node embedding expression, the rainfall-flow coupling parameters affect the training rainfall-flow coupling expression, the gating parameters affect the training time delay feature expression result, the prediction output parameters affect the training water level prediction segment and the training inflow prediction segment, and the sequence arrangement parameters affect the training predicted water level sequence and the training predicted inflow sequence. The adjusted deep temporal neural network continues to read the next batch of training samples and generates training prediction water level sequences and training prediction inflow sequences again, so as to gradually reduce the training loss state through multiple rounds of training.

[0149] Preferably, during multiple rounds of training, a spatiotemporal consistency check is performed on the training samples. This check reads the target reservoir object, spatial nodes, execution time position, historical rainfall feature vector, historical flow feature vector, historical target water level sequence, and historical target inflow sequence from each training sample. It also checks whether spatial nodes within the same training sample originate from the same digital twin reservoir base and whether execution time positions within the same training sample are continuously arranged according to the grid time reference. If the binding relationship between spatial nodes and the target reservoir object in the training sample is incomplete, or if there are time breaks in the execution time position that cannot be filled by feedback gap markers, the training sample is marked as invalid. This invalid state is used to exclude the training sample from the current training round, ensuring that the training process of the deep temporal neural network does not write data with inconsistent spatial or discontinuous temporal sources into the parameter correction process. If a training sample is not marked as invalid, it is used as a valid training sample for the current training round and continues to be input into the temporal input processing layer.

[0150] Preferably, the training completion conditions of the deep temporal neural network include a training loss state convergence condition, a verification prediction deviation condition, and a temporal trend consistency condition. The training loss state convergence condition expresses that the change amplitude of the training loss state after multiple rounds of training is lower than the preset training change amplitude. The verification prediction deviation condition expresses that the water level verification deviation state between the verification prediction water level sequence and the verification target water level sequence, as well as the inflow verification deviation state between the verification prediction inflow sequence and the verification target inflow sequence, are both within the preset verification deviation range. The temporal trend consistency condition expresses that the upward trend, downward trend, or maintenance trend of the verification prediction water level sequence and the verification target water level sequence within the continuous execution time position meets the preset trend consistency requirement. When the training loss state convergence condition, the verification prediction deviation condition, and the temporal trend consistency condition are all satisfied, the adjustment of the trainable parameters in the deep temporal neural network is stopped, and the current deep temporal neural network is used as the deep temporal neural network that has completed training. When any one of the training loss state convergence condition, the verification prediction deviation condition, and the temporal trend consistency condition is not satisfied, the next batch of training samples is read, and the generation of training temporal input samples, spatial node embedding expression, training rainfall-flow coupling expression, training time delay feature expression results, training predicted water level sequence, training predicted inflow sequence, training loss state calculation, and trainable parameter adjustment are continued.

[0151] Preferably, after training is completed, the structural layer parameters of the time-series input organization layer, spatial node embedding layer, rainfall-flow coupling layer, long short-term memory unit, prediction output layer, and sequence organization layer in the trained deep temporal neural network are written into the network training record. The network training record includes the training sample source, training input samples, training label samples, training time-series input samples, training rainfall-flow coupling expression, training time delay feature expression result, training loss state, parameter correction state, training completion conditions, and the trained deep temporal neural network. The network training record continues to participate in subsequent time series prediction processes, enabling the directly invoked deep temporal neural network when acquiring new time-series grid data, and generating the predicted water level sequence and the predicted inflow sequence according to the trained time-series input organization method, spatial node embedding method, rainfall-flow coupling method, time delay feature extraction method, and sequence organization method.

[0152] like Figure 3 As shown, this is an embodiment of a digital twin reservoir flood control scheduling device according to this application. The digital twin reservoir flood control scheduling device includes: The data acquisition module is used to acquire multi-source reservoir monitoring data and reservoir spatiotemporal grids; The spatial topology binding module is used to perform spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference and spatial topology binding rules in the reservoir spatiotemporal grid, so as to generate a digital twin reservoir bottom plate; The time series prediction module is used to generate time series grid data based on the digital twin reservoir bottom plate, and use a deep time series neural network to perform time series prediction on the time series grid data to obtain the predicted water level sequence and the predicted inflow sequence. The scheduling sequence generation module is used to input the predicted water level sequence and the predicted inflow sequence into the scheduling optimization model. Based on the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rules in the scheduling optimization model, the module performs candidate scheduling coding processing on the preset reservoir gate opening range and scheduling time unit to obtain an initial group composed of multiple candidate scheduling sequence codes. The module then performs group iterative optimization processing on the initial group to obtain the target scheduling sequence. The flood control scheduling execution module is used to convert the target scheduling sequence into physical reservoir control instructions, and control the physical reservoir control terminal to execute flood control scheduling actions according to the physical reservoir control instructions.

[0153] like Figure 4 As shown, an electronic device according to an embodiment of this application includes: a processor and a memory; the memory is used to store computer programs; the processor is used to execute the program stored in the memory to implement the functions of each module of the device described in the embodiment of this application, or to implement the steps of the method described.

[0154] like Figure 5 As shown, a computer-readable storage medium is provided in an embodiment of this application, on which a computer program is stored; the processor is used to execute the program stored in the memory to implement the functions of each module of the device described in the embodiment of this application, or to implement the steps of the method described.

[0155] Figures 2-5 For an exemplary description, please refer to the above. Figure 1 This will not be elaborated upon here.

Claims

1. A digital twin reservoir flood control scheduling method, characterized in that, The digital twin reservoir flood control scheduling method includes the following steps: acquiring multi-source reservoir monitoring data and a reservoir spatiotemporal grid; performing spatial topology binding on the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate; generating time-series grid data based on the digital twin reservoir base plate, and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence; and inputting the predicted water level sequence and the predicted inflow sequence into the scheduling system. The optimization model is used to perform candidate scheduling coding processing on the preset reservoir gate opening range and scheduling time unit based on the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rules in the scheduling optimization model. This results in an initial group composed of multiple candidate scheduling sequence codes, and the initial group is subjected to group iterative optimization processing to obtain the target scheduling sequence. The target scheduling sequence is then converted into physical reservoir control commands, and the physical reservoir control terminal is controlled to execute flood control scheduling actions according to the physical reservoir control commands.

2. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, Before the step of spatially binding the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid to generate a digital twin reservoir base plate, the method further includes: acquiring historical reservoir documents, using a data extraction engine to extract data from the historical reservoir documents to obtain feature data units; performing normalization and cleaning on the feature data units to obtain standard structured data; constructing a reservoir data spatial mapping relationship between the standard structured data and the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid, and using the reservoir data spatial mapping relationship as the binding basis for the spatial topology binding rules.

3. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, The step of generating a digital twin reservoir base plate by spatially binding the multi-source reservoir monitoring data according to the three-dimensional spatial coordinate system, grid time reference, and spatial topology binding rules in the reservoir spatiotemporal grid includes: obtaining the quadtree structure in the three-dimensional basic platform of the digital twin reservoir; performing coordinate normalization processing on the spatial nodes in the quadtree structure according to the three-dimensional spatial coordinate system in the reservoir spatiotemporal grid; and constructing a hierarchical detail pyramid model according to the coordinate normalized quadtree structure; determining the resolution level according to the observation point distance parameter; and performing spatial topology binding according to the resolution level, the... The grid time reference and spatial topology binding rules in the reservoir spatiotemporal grid organize the multi-source reservoir monitoring data into a bottom-level tile matrix with time index and spatial binding position. The bottom-level tile matrix with time index and spatial binding position is mapped to the resolution level corresponding to the hierarchical detail pyramid model, so that the multi-source reservoir monitoring data forms a spatial topology binding result under the joint constraints of the three-dimensional spatial coordinate system, the grid time reference and the spatial topology binding rules, so as to generate the digital twin reservoir base plate carrying the spatial topology binding result.

4. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, The steps of generating time-series grid data based on the digital twin reservoir floor and using a deep temporal neural network to perform time series prediction on the time-series grid data to obtain predicted water level sequences and predicted inflow sequences include: organizing the spatial topology binding results in the digital twin reservoir floor according to the grid time reference to obtain the time-series grid data; extracting rainfall feature vectors and flow feature vectors from the time-series grid data; inputting the rainfall feature vectors and flow feature vectors into the deep temporal neural network containing long short-term memory units; extracting time delay features from the rainfall feature vectors and flow feature vectors through the long short-term memory units to obtain time delay feature expression results; and generating the predicted water level sequence and the predicted inflow sequence based on the time delay feature expression results.

5. The digital twin reservoir flood control scheduling method as described in claim 4, characterized in that, Before the step of inputting the rainfall feature vector and the flow feature vector into the deep time-series neural network containing long short-term memory units, the method further includes: using a local outlier factor algorithm to detect outliers in the time-series grid data and removing outlier data units from the time-series grid data to obtain time-series grid data with outlier data units removed; extracting corresponding rainfall feature vectors and flow feature vectors from the time-series grid data with outlier data units removed, so that the rainfall feature vectors and flow feature vectors input into the deep time-series neural network originate from the time-series grid data with outlier data units removed; performing cluster analysis on historical flood peak data to obtain a reference sequence of similar flood peaks; and inputting the reference sequence of similar flood peaks as prior reference information into the deep time-series neural network, so that the deep time-series neural network performs time series prediction based on the prior reference information, the rainfall feature vector, and the flow feature vector.

6. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, The step of inputting the predicted water level sequence and the predicted inflow sequence into the scheduling optimization model, and performing candidate scheduling coding processing on the preset reservoir gate opening range and scheduling time unit according to the predicted water level sequence, the predicted inflow sequence, the highest water level constraint, the outflow variation constraint, and the scheduling sequence optimization rules in the scheduling optimization model, to obtain an initial group composed of multiple candidate scheduling sequence codes, and performing group iterative optimization processing on the initial group to obtain the target scheduling sequence includes: in the scheduling optimization model, according to the predicted water level sequence, the predicted inflow sequence, and the preset reservoir gate opening range, the candidate scheduling sequence is coded as follows: The range and the scheduling time unit generate multiple reservoir gate opening degree combinations; each reservoir gate opening degree combination is encoded as a candidate scheduling sequence code, and the initial group is composed of multiple candidate scheduling sequence codes; based on each candidate scheduling sequence code, the predicted water level sequence, and the predicted inflow sequence, the candidate water level change sequence and the candidate outflow sequence after the candidate scheduling are deduced; based on the highest water level constraint, the outflow amplitude constraint, and the scheduling sequence optimization rule, the candidate water level change sequence, the candidate outflow sequence, and the initial group are subjected to group iterative optimization processing to obtain the target scheduling sequence.

7. The digital twin reservoir flood control scheduling method as described in claim 6, characterized in that, The step of performing group iterative optimization processing on the candidate water level change sequence, the candidate discharge flow rate sequence, and the initial group based on the highest water level constraint, the discharge flow rate variation constraint, and the scheduling sequence optimization rule to obtain the target scheduling sequence includes: for each candidate scheduling sequence code in the initial group, calculating the fitness of the candidate water level change sequence corresponding to the candidate scheduling sequence code according to the highest water level constraint to obtain a first fitness value of the candidate scheduling sequence code; calculating the fitness of the candidate discharge flow rate sequence corresponding to the candidate scheduling sequence code according to the discharge flow rate variation constraint to obtain a second fitness value of the candidate scheduling sequence code; determining the comprehensive fitness value of the candidate scheduling sequence code based on the first fitness value and the second fitness value; performing group iterative optimization processing on the initial group according to the comprehensive fitness value corresponding to each candidate scheduling sequence code and the scheduling sequence optimization rule to obtain the target candidate scheduling sequence code; and decoding the target candidate scheduling sequence code into the target scheduling sequence.

8. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, After the steps of generating time-series grid data based on the digital twin reservoir base plate and using a deep time-series neural network to perform time-series prediction on the time-series grid data to obtain a predicted water level sequence and a predicted inflow sequence, the method further includes: obtaining a tiered early warning threshold; comparing and verifying the predicted water level values ​​in the predicted water level sequence with the tiered early warning threshold to obtain a predicted water level threshold verification result; generating a dynamic early warning identifier when the predicted water level value exceeds the tiered early warning threshold as indicated by the predicted water level threshold verification result; and dynamically updating the outflow amplitude constraint conditions in the scheduling optimization model according to the dynamic early warning identifier.

9. The digital twin reservoir flood control scheduling method as described in any one of claims 1 to 8, characterized in that, Before the step of converting the target scheduling sequence into physical reservoir control commands, the method further includes: importing the target scheduling sequence, the predicted water level sequence, and the predicted inflow sequence into a digital twin reservoir 3D rendering engine; simulating the action of a virtual gate model in the digital twin reservoir base according to the target scheduling sequence, and providing the initial water level state and inflow driving state for the action simulation of the virtual gate model according to the predicted water level sequence and the predicted inflow sequence; generating a virtual simulation screen of water surface evolution using a pixel-stream rendering mechanism; performing flood control safety verification on the target scheduling sequence under the constraints of the highest water level and the variation of the outflow based on the virtual simulation screen of water surface evolution, and obtaining the flood control safety verification result; and executing the step of converting the target scheduling sequence into physical reservoir control commands when the flood control safety verification result is verified as passed.

10. The digital twin reservoir flood control scheduling method as described in claim 1, characterized in that, Following the step of controlling the physical reservoir control terminal to perform flood control scheduling actions according to the physical reservoir control command, the method further includes: acquiring real-time measured water level data fed back by the physical reservoir control terminal; performing time-series alignment processing on the real-time measured water level data according to the grid time reference to obtain a real-time measured water level sequence corresponding to the predicted water level sequence; calculating the predicted water level error digital sequence between the predicted water level sequence and the real-time measured water level sequence; extracting the error change characteristics of the predicted water level error digital sequence using an autoregressive algorithm; and correcting the parameters of the deep temporal neural network based on the error change characteristics.

Images