Yellow river rainwater condition fusion prediction method and system

Abstract

This invention discloses a method and system for fusion prediction of Yellow River rainfall and water conditions, belonging to the field of rainfall and water condition data processing technology. The method includes constructing a Yellow River network topology map and generating multi-source time-series data of Yellow River rainfall and water conditions. Based on the Yellow River network topology map, initial values ​​of Yellow River side propagation delay are calculated, mapping consistency is verified, and a monotonically bounded time mapping of the Yellow River is learned. An upstream sequence alignment sequence is generated according to the Yellow River monotonically bounded time mapping, and the fusion prediction results of Yellow River rainfall and water conditions and the Yellow River variable time delay interpretation quantity are output. This invention uses the Yellow River network topology as a priori, combined with a learnable monotonically bounded variable time delay mapping and topology consistency verification, to achieve online alignment and confidence-driven fusion of multi-source rainfall and water conditions on a unified time axis. It also outputs key hydrological elements through multi-step regression of topology convergence, while providing traceable variable time delay interpretation quantities to support business review and scheduling analysis.

Description

Technical Field

[0001] This invention relates to the field of rainfall and water data processing technology, specifically to a method and system for fusion prediction of rainfall and water conditions along the Yellow River. Background Technology

[0002] In recent years, rainfall and water situation forecasting for watershed flood prediction has gradually evolved from empirical statistics and conceptual hydrological models based primarily on rain gauges and hydrological stations to a multi-source heterogeneous data processing system centered on computers. This system integrates information from rain gauges, radar quantitative precipitation, satellite precipitation estimation, control section water level and flow, and reservoir scheduling. By combining time series modeling, data assimilation, and machine learning, it enables rolling predictions of multi-step flow, water level, and flood peak elements. Simultaneously, spatiotemporal deep networks and graph structure modeling are being used to characterize watershed spatial dependencies and upstream-downstream transmission relationships, driving rainfall and water situation forecasting towards high-resolution, near real-time, and scalable electronic data processing.

[0003] Existing technologies still have shortcomings in scenarios like the Yellow River, which are significantly affected by differences in river segmentation and reservoir scheduling. Multi-source data suffers from storage delays, inconsistent timestamps, and missing outliers. Common approaches often rely on static alignment or simple window matching, which struggles to reliably support real-time rolling forecasts. The propagation time lags from upstream rainfall to downstream hydrological conditions and from upstream cross-sections to downstream cross-sections dynamically change with flow velocity, river morphology, and scheduling operations. Fixed lags or offline dynamic time warping cannot simultaneously ensure linearity and monotonic availability, and are prone to time-reversal mismatches. Existing alignment methods often optimize on a one-sided or local station basis, lacking end-to-end consistency constraints on the river network topology. This can lead to conflicting alignment results for the same upstream information on different paths, resulting in non-repeatable fusion features. Fusion weights often do not explicitly incorporate interpolation confidence and delay factors, causing low-confidence inputs to participate in fusion and leading to prediction fluctuations. Therefore, existing technologies struggle to achieve topologically consistent, monotonically bounded, and iteratively verifiable variable-delay alignment within a unified computing framework, and further generate traceable rainfall and hydrological fusion prediction outputs. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing rainwater fusion prediction methods suffer from peak time deviation due to fixed propagation time delay modeling, alignment instability due to multi-source timestamp delay and missing outliers, mapping conflicts due to lack of end-to-end consistency constraints on river network topology, and the problem of how to achieve online monotonically bounded variable time delay alignment fusion prediction for the Yellow River network and synchronously output traceable explanatory quantities.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a Yellow River rainfall and water situation fusion prediction method, including constructing a Yellow River network topology map and generating multi-source time series data of Yellow River rainfall and water situation.

[0007] The initial value of the propagation delay along the Yellow River is calculated based on the Yellow River network topology map. The consistency of the mapping is verified and the monotonic bounded time mapping of the Yellow River is learned.

[0008] The upstream sequence is aligned with the monotonic bounded time mapping of the Yellow River, and the fusion prediction results of the Yellow River rainfall and water conditions and the time delay explanation of the Yellow River are output.

[0009] As a preferred embodiment of the Yellow River rainfall and water situation fusion prediction method described in this invention, the construction of the Yellow River network topology map includes: reading the Yellow River main stream and tributary segment codes, station coordinates, control section identifiers, and confluence relationship tables; uniformly incorporating rain gauges, hydrological stations, sub-basin nodes, and reservoir control points into a node set; establishing directed connectivity relationships between nodes based on the confluence direction; and marking directed connectivity relationships pointing to any downstream node as incoming edges of the downstream node. For each directed connectivity relationship, the river segment distance, river segment category, and control section type are written as edge attributes. A topology consistency check is performed on the node set, and an adjacency data structure and a path index data structure are generated and stored in memory.

[0010] As a preferred embodiment of the Yellow River rainfall and water level fusion prediction method described in this invention, the generation of multi-source time-series data for Yellow River rainfall and water level includes collecting rainfall sequences from rain gauge stations, quantitative precipitation sequences from radar, estimated precipitation sequences from satellites, flow sequences from control sections, water level sequences from control sections, and reservoir discharge sequences. Each sequence is then time-stamped and its sampling interval is standardized according to the unified timeline for Yellow River operations. Missing data is marked with missing data tags and outliers are removed according to preset outlier detection rules. The retained data is assembled into node feature sequences according to a fixed field order. The node feature sequences are then stored in association with station identifiers and timestamp indices.

[0011] As a preferred embodiment of the Yellow River rainfall and water situation fusion prediction method described in this invention, the calculation of the initial value of the Yellow River side propagation delay based on the Yellow River network topology map includes: reading the corresponding river segment distance and river segment category for each directed connectivity relationship in the Yellow River network topology map; extracting the flow sequence, water level sequence, and reservoir discharge sequence of the upstream and downstream of the river segment within the recent sliding window; obtaining the equivalent propagation velocity according to the preset Yellow River segment category and propagation velocity conversion rules; and then calculating the candidate propagation delay using the river segment distance and the equivalent propagation velocity. A recalculation is triggered when a change in the Yellow River scheduling change flag is detected or the flow increment within the window exceeds the Yellow River flow mutation judgment threshold. The candidate propagation delay is written to the memory with a limit based on the maximum propagation delay of the Yellow River segment, serving as the initial value of the Yellow River side propagation delay.

[0012] As a preferred embodiment of the Yellow River rainfall and water level fusion prediction method described in this invention, the step of verifying mapping consistency and learning the Yellow River monotonically bounded time mapping includes: establishing a mapping table from the target time to the upstream sampling time for each directed connectivity relationship; the mapping table uses the initial value of the Yellow River side propagation delay as a baseline; and a trainable mapper calculates the Yellow River variable time delay interpretation quantity based on the edge condition vector, which includes upstream flow and water level, downstream flow and water level, reservoir discharge, river segment category, and timestamp. An amplitude limit rule is applied to the Yellow River variable time delay interpretation quantity to ensure that the amplitude does not exceed the maximum propagation delay of the Yellow River segment. A non-negative increment constraint is applied to the mapping output of adjacent target times to ensure that the upstream sampling time corresponding to the later time is not earlier than the upstream sampling time corresponding to the previous time, satisfying the monotonic constraint. The path index based on the Yellow River network topology map verifies the end-to-end mapping consistency from any upstream node to the downstream node via intermediate nodes. When the difference between the end-to-end Yellow River variable time delay interpretation and the sum of the two Yellow River variable time delay interpretations exceeds the Yellow River topology additive consistency tolerance, iterative updates of the trainable mapper parameters are triggered, and the updated Yellow River monotone bounded time mapping is written into memory.

[0013] As a preferred embodiment of the Yellow River rainfall and water level fusion prediction method described in this invention, the step of generating an upstream sequence alignment sequence according to the Yellow River monotone bounded time mapping includes: using the target time of the downstream node corresponding to the incoming edge as a benchmark, calling the Yellow River monotone bounded time mapping to obtain the upstream sampling time corresponding to the target time; indexing two adjacent timestamp samples of the upstream node in the Yellow River rainfall and water level multi-source time series data and generating an alignment feature vector according to a preset interpolation resampling rule; and simultaneously generating an interpolation confidence value corresponding to the alignment feature vector. When the upstream sampling time falls outside the coverage area of ​​the upstream node data, the nearest boundary sample is taken according to the boundary preservation rule. The process is repeated for the same incoming edge at consecutive target times to obtain an upstream sequence alignment sequence consistent with the time axis of the downstream node, and the upstream sequence alignment sequence is associated and stored with the incoming edge identifier and the target time index.

[0014] As a preferred embodiment of the Yellow River rainfall and water level fusion prediction method described in this invention, the output of the Yellow River rainfall and water level fusion prediction result and the Yellow River variable time delay interpretation quantity includes: for each downstream node at the same target time, aligning the node feature sequence with the upstream sequence corresponding to all incoming edges; calculating the fusion weight of each input channel based on data delay information, missing data markers, and interpolation confidence, and generating a fusion feature sequence. The fusion feature sequence is input into the Yellow River topological constraint variable time delay aligned fusion multi-step sequence regression model to generate a flow prediction sequence, water level prediction sequence, and peak arrival time prediction sequence within a preset prediction step size range as the Yellow River rainfall and water level fusion prediction result. The Yellow River variable time delay interpretation quantity output by the trainable mapper is bound by the incoming edge identifier and timestamp, and synchronously output and stored in the memory along with the Yellow River rainfall and water level fusion prediction result.

[0015] As a preferred embodiment of the Yellow River rainfall and water situation fusion prediction system described in this invention, it includes: a river network modeling and multi-source sequencing module, a variable time delay mapping learning and verification module, and an alignment fusion prediction and release module.

[0016] The river network modeling and multi-source sequencing module is used to construct the Yellow River network topology map and generate multi-source time-series data on Yellow River rainfall and water conditions.

[0017] The variable delay mapping learning and verification module is used to calculate the initial value of the propagation delay along the Yellow River based on the Yellow River network topology map, verify the consistency of the mapping, and learn the monotonic bounded time mapping of the Yellow River.

[0018] The alignment fusion prediction and interpretation module is used to generate an upstream sequence alignment sequence according to the Yellow River monotonically bounded time mapping, and outputs the Yellow River rainfall and water level fusion prediction results and the Yellow River variable time delay interpretation quantity.

[0019] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement a method for integrating rainfall and water conditions in the Yellow River.

[0020] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a Yellow River rainfall and water level fusion prediction method.

[0021] The beneficial effects of this invention are: By constructing a topological map of the Yellow River network and generating multi-source time-series data on Yellow River rainfall and water conditions, this approach unifies rain gauges, hydrological stations, sub-basins, and reservoir control points into a single river network data structure, and uses adjacency and path indexing to achieve a searchable representation of upstream and downstream relationships. Simultaneously, it unifies the time axis of the multi-source sequences, corrects biases, marks missing data, and removes outliers, ensuring stable and reproducible data input for subsequent calculations. This provides a structured prior and quality-controlled sample foundation for propagation delay estimation and mapping learning, reducing systematic errors and non-reproducibility caused by heterogeneous sampling and outliers at the source.

[0022] By calculating the initial value of the propagation delay along the Yellow River and learning the monotonically bounded time mapping of the Yellow River, this method transforms dynamic information such as river segment distance, river segment type, and recent flow, water level, and discharge into an updatable propagation delay baseline. Based on this, a trainable mapper outputs a variable delay interpretation. Amplitude limiting and non-negative increment constraints ensure the monotonicity and usability of the mapping, and end-to-end mapping consistency verification and iterative updates are performed through path indexing, ensuring that local alignment does not disrupt the path consistency of the entire river network. This constructs an online, controllable, and topologically consistent spatiotemporal alignment mechanism, providing a traceable basis for delay adjustment in scenarios such as changes in flood peak propagation speed and sudden scheduling changes.

[0023] By generating an aligned sequence from the upstream sequence and outputting the fusion prediction results and explanatory values, this method achieves interpolation and resampling of upstream multi-source sequences according to time mapping. This allows upstream information to form an aligned feature vector that can be directly fused at the downstream target time, and simultaneously generates interpolation confidence scores. Furthermore, data latency, missing data markers, and interpolation confidence scores are incorporated into the channel weight calculation, enabling the fusion process to explicitly suppress latency and missing data. Based on this, a regression network composed of a time-series encoder, a topology convergence layer, and a multi-step decoder outputs traffic flow, water level, and peak arrival time, and binds and stores the variable latency explanatory values ​​with incoming edge identifiers and timestamps. This forms an integrated closed loop of alignment, fusion, prediction, and traceability, facilitating online business calls and result verification. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 The above is an overall flowchart of a Yellow River rainfall and water situation fusion prediction method provided in Embodiment 1 of the present invention. Detailed Implementation

[0026] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0027] Example 1, referring to Figure 1 As an embodiment of the present invention, a Yellow River rainfall and water situation fusion prediction method is provided, comprising: S1: Construct the Yellow River network topology map and generate multi-source time-series data on Yellow River rainfall and water conditions.

[0028] The Yellow River main stream and tributary segment coding and confluence relationship table is generated from Yellow River network data according to segment coding rules, and includes upstream node identifier, downstream node identifier, segment distance field, and segment category field. The Yellow River main stream and tributary segment codes, station coordinates, control section identifiers, and confluence relationship table are read. Rain gauge stations, hydrological stations, sub-basin nodes, and reservoir control points are uniformly included in the node set, and directed connectivity relationships are established between nodes according to the confluence direction. Directed connectivity relationships pointing to any downstream node are marked as incoming edges of the downstream node. For each directed connectivity relationship, segment distance, segment category, and control section type are written as edge attributes. Topological consistency checks are performed on the node set, and adjacency data structures and path index data structures are generated and stored in memory.

[0029] In this invention, the river segment codes, station coordinates, control section identifiers, and confluence relationship tables for the Yellow River's main stream and tributaries are all derived from a comprehensive compilation of Yellow River rainfall and water level operational data and river network system data. The river segment codes and river network system geometric information are derived from the results of tailoring the basin's river system data or national basic geographic water system data within the Yellow River basin, and the main stream and tributary river segments are numbered according to a pre-defined river segment coding rule. Station coordinates are derived from the directories of rain gauges, hydrological stations, reservoir control points, and other monitoring stations, as well as the latitude and longitude information registered in the rainfall and water level operational system's station database, and are unified using a unique station identifier. Control section identifiers are derived from the long-term control section list used in basin hydrological monitoring and scheduling operations and their correspondence with monitoring stations. The confluence relationship table is generated from the aforementioned river network system data and basin confluence zoning results, including upstream node identifier, downstream node identifier, connection direction, river segment distance, and river segment category fields, and is maintained and updated by basin operational personnel or the system according to topological consistency verification rules.

[0030] Furthermore, the station coordinates are stored using a unified geographic coordinate system, and unique identifiers are set for stations with the same name and relocated stations. A mapping table is established between the unique station identifiers and the identifiers of their respective sub-basins. When a station is located at the boundary of multiple sub-basins, its affiliation is assigned according to the confluence relationship table or the nearest confluence node. The mapping table and the node set are stored together in the memory for subsequent retrieval.

[0031] Furthermore, the river segment distance is preferentially calculated using the broken line length of the river segment centerline and written into the edge attribute. When the centerline is missing in the confluence relationship table, the geodesic distance between the two node coordinates is multiplied by the Yellow River channel tortuosity correction coefficient of 1.2 as the river segment distance. The correction coefficient is derived from the average tortuosity ratio obtained from the statistical analysis of the historical channel centerlines of the Yellow River's main stream and tributaries, and is fixed at 1.2.

[0032] Furthermore, topology consistency verification includes performing in-degree statistics and topology sorting on all nodes. If the topology sorting cannot cover all nodes, a loop is identified, and the directed connectivity relationships involved in the loop are marked as abnormal connectivity relationships. Abnormal connectivity relationships are corrected first according to the upstream and downstream fields of the merge relationship table. If they still cannot be corrected, the abnormal connectivity relationship is removed from the adjacency data structure and the removal log is recorded.

[0033] Furthermore, the path index data structure includes generating a set of all upstream nodes and their corresponding path sequences for each downstream node. The path sequences are first sorted topologically to determine the start and end order of traversal, and then generated by performing a depth-first traversal for each downstream node. Additionally, a path length and path edge set index are written for each path to support rapid location of a specified three-node path during subsequent end-to-end mapping consistency verification.

[0034] Rainfall sequences from rain gauge stations, quantitative precipitation sequences from radar, estimated precipitation sequences from satellites, flow sequences from control sections, water level sequences from control sections, and reservoir discharge sequences were collected. These sequences were then time-stamped and their sampling intervals were standardized according to the unified timeline for Yellow River operations. Missing data were marked with missing data tags and outliers were removed according to pre-defined outlier detection rules. The retained data were assembled into node feature sequences based on a fixed field order. These node feature sequences were then stored in association with station identifiers and timestamp indexes.

[0035] Furthermore, the Yellow River business uses Beijing time as the unified timeline and a five-minute sampling interval. All data source timestamps are first converted to Beijing time and then aligned to the nearest five-minute interval. When the original timestamp deviates from the hour by no more than two minutes, the nearest alignment is used. If the deviation exceeds two minutes, the original timestamp is retained and a time alignment anomaly is written into the missing data flag. The two-minute threshold is derived from the 95th percentile of common transmission delay statistics in the Yellow River rainfall and water situation business system and is fixed at two minutes.

[0036] Furthermore, timestamp correction includes performing monotonicity checks on continuous records from the same site. If timestamp backtracking or duplication occurs, correction is performed using the difference between the server's receiving time and the site's reporting time. When the absolute value of the difference exceeds ten minutes, it is determined to be clock drift and triggers the entire correction process. The ten-minute threshold is derived from the maintenance constraints of clock drift in the Yellow River rainfall and water level monitoring equipment and is fixed at ten minutes. Data points that are still not monotonic after correction are written as missing data points and do not participate in the assembly.

[0037] Furthermore, the outlier detection rules combine physical range rules and abrupt change rules. Rainfall intensity less than zero or greater than 250 mm per hour at rain gauge stations is considered an anomaly. Radar quantitative precipitation greater than 200 mm per hour is considered an anomaly. Satellite precipitation estimates greater than 150 mm per hour are considered an anomaly. A change in water level at a control section exceeding 0.8 meters within five consecutive minutes is considered an anomaly. A change in flow rate at a control section exceeding 40% of the previous time interval within five consecutive minutes is considered an anomaly. A change in reservoir discharge exceeding 1000 cubic meters per second within five consecutive minutes is considered an anomaly. These thresholds are determined by the smaller of the 99.5 percentile of similar data from the Yellow River over the past five years and the upper limit of the instrument range, and are fixed as values ​​written into the system configuration. Rainfall intensity is in millimeters per hour, water level in meters, and flow rate and discharge in cubic meters per second. The 99.5 percentile is obtained based on the empirical distribution of valid samples from the same station and field over the past five years, after five-minute resampling.

[0038] It should be noted that this step first establishes node and inbound edge connectivity using the Yellow River network as the framework, and then unifies rainfall, radar, satellite, flow, water level and discharge into the same business time axis. Combined with clock correction, missing measurement marking and outlier threshold cleaning, a searchable adjacency and path index and node feature sequence are formed, which provides stable input for subsequent variable time delay mapping learning and end-to-end consistency verification.

[0039] S2: Calculate the initial value of the propagation delay along the Yellow River based on the Yellow River network topology map, verify the consistency of the mapping, and learn the monotonic bounded time mapping of the Yellow River.

[0040] The adjacency data structure, path index data structure, and node feature sequence written to memory by S1 are read as inputs for calculation and learning in this step. For each directed connectivity relationship in the Yellow River network topology, the corresponding river segment distance and river segment category are read. The flow sequence, water level sequence, and reservoir discharge sequence of the upstream and downstream of the river segment within the recent sliding window are extracted. The equivalent propagation velocity is obtained according to the preset Yellow River segment category and propagation velocity conversion rules. Then, the candidate propagation delay is calculated using the river segment distance and the equivalent propagation velocity. Recalculation is triggered when a change in the Yellow River scheduling change flag is detected or the flow increment within the window exceeds the Yellow River flow change judgment value. The candidate propagation delay is written to memory with a limit based on the maximum propagation delay of the Yellow River segment as the initial value of the Yellow River side propagation delay.

[0041] Furthermore, the recent sliding window is fixed at 60 minutes with a sampling step of 5 minutes. The arithmetic mean of the flow rate sequence, water level sequence, and reservoir discharge sequence within the window is used as the window statistic. When the proportion of missing data markers within the window exceeds 30%, the window statistic is deemed unreliable and the initial value update of the edge propagation delay is suspended until the proportion of missing data markers does not exceed 30%.

[0042] Furthermore, the conversion rules for Yellow River section categories and propagation speeds include classifying river sections into four types: canyon type, plain type, reservoir type, and scour-deposition type, with corresponding baseline propagation speeds of 1.8 m / s, 1.2 m / s, 0.8 m / s, and 1.0 m / s, respectively. A minimum propagation speed of 0.2 m / s is also configured. The maximum propagation delay for a Yellow River section is taken as the smaller of the section distance divided by 0.2 m / s and 48 hours. Both the minimum propagation speed and the maximum propagation delay for the Yellow River section are globally configured and written into the system configuration.

[0043] Furthermore, the Yellow River dispatch change indicator is generated based on the reservoir discharge sequence and gate status. When the absolute value of the difference between two consecutive 5-minute reservoir discharge samples exceeds 300 cubic meters per second and continues for two sampling cycles, or when the gate status changes from open to closed, the Yellow River dispatch change indicator is set to a changed state. The Yellow River flow change criterion is the greater of 40% of the previous flow rate and 300 cubic meters per second. A recalculation is triggered when any sample within the window satisfies the condition that the absolute value of the flow increase exceeds this criterion. When gate status is missing, the Yellow River dispatch change indicator is generated solely based on the reservoir discharge difference and its duration.

[0044] To ensure that the candidate propagation delay provides computable support for the variation of flow velocity with inflow and scheduling, the equivalent propagation velocity is first obtained by fusing the baseline propagation velocity for each river segment category with window statistics. The equivalent propagation velocity is calculated as follows:

[0045] in, Represents directed connectivity The equivalent propagation speed for the corresponding river section. Represents directed connectivity The baseline propagation speed for the corresponding river segment category. This indicates the inflow correction factor, which is preset according to the river section category. This represents the scheduling correction coefficient, which is preset according to the river section category. This represents the average flow sequence of upstream nodes within the recent sliding window. Indicates the reference flow rate for the river section category. This represents the average value of the reservoir discharge sequence within the recent sliding window. This represents the reference discharge for the river segment category; if this segment does not contain any reservoir control points, it is set to zero. When this segment does not contain any reservoir control points, the scheduling correction term is not calculated. and order Used for reserving positions. When the edge contains reservoir control points, it is selected according to the reservoir area type. cubic meters per second and sorted by category .

[0046] The inflow correction factor and the dispatch correction factor are fixed values ​​according to the river section category: 0.3 and 0.2 for canyon type, 0.25 and 0.15 for plain type, 0.2 and 0.25 for reservoir type, and 0.35 and 0.15 for scour and sedimentation type. The reference flow rate for each river section category is 1000, 800, 500, and 600 cubic meters per second, respectively. The reference discharge rate for each river section category is 800 cubic meters per second for reservoir type and zero for other categories.

[0047] Within each 5-minute update cycle, window statistics are read for each directed connectivity relationship, and the equivalent propagation velocity is calculated. When the Yellow River scheduling change flag is in a changing state, the equivalent propagation velocity is recalculated using the latest two window statistics to reflect the sudden change in propagation velocity caused by the scheduling.

[0048] Candidate propagation delays are calculated based on equivalent propagation velocity and river segment distance. After being limited to the maximum propagation delay of the Yellow River segment, the delay is written into memory as the initial value for propagation delay along the Yellow River. The propagation delay is expressed as:

[0049] in, Represents directed connectivity The candidate propagation delay for the corresponding river segment. Represents directed connectivity The distance between the corresponding river segments. This represents the minimum propagation speed configuration value. This represents the initial value of the propagation delay along the Yellow River. This represents the maximum propagation delay configuration value for the Yellow River section.

[0050] When the Yellow River dispatch change flag changes or the Yellow River flow change judgment is triggered, the propagation delay of the affected side is recalculated and the initial value of the Yellow River side propagation delay in memory is overwritten. If no trigger occurs, the initial value of the Yellow River side propagation delay from the previous cycle is kept unchanged to reduce jitter.

[0051] For each directed connectivity relationship, a mapping table is established from the target time to the upstream sampling time. The mapping table uses the initial value of the Yellow River's propagation delay as the baseline. A trainable mapper calculates the Yellow River's variable time delay interpretation based on the edge condition vector, which includes upstream flow and water level, downstream flow and water level, reservoir discharge, river segment category, and timestamp. An amplitude limit rule is applied to the Yellow River's variable time delay interpretation to ensure that the amplitude does not exceed the maximum propagation delay of the Yellow River segment. A non-negative increment constraint is applied to the mapping output of adjacent target times to ensure that the upstream sampling time corresponding to the later time is not earlier than the upstream sampling time corresponding to the previous time, satisfying the monotonic constraint. Based on the path index of the Yellow River network topology, the end-to-end mapping consistency from any upstream node to the downstream node via intermediate nodes is verified. The end-to-end Yellow River variable time delay interpretation is generated through the three-node path determined by the path index. The end-to-end mapping output is obtained by first calculating the upstream sampling time from the intermediate node to the downstream node, and then substituting it as the target time into the mapping table from the upstream node to the intermediate node. The initial value of the end-to-end Yellow River propagation delay is the sum of the initial values ​​of the propagation delays of the two Yellow River segments along the path. When the difference between the end-to-end Yellow River variable delay interpretation and the sum of the variable delay interpretations of the two Yellow River segments exceeds the Yellow River topology additive consistency tolerance, iterative updates to the trainable mapper parameters are triggered, and the updated Yellow River monotonically bounded time map is written to memory.

[0052] Furthermore, the mapping table from the target time to the upstream sampling time is generated hourly with a time step of 5 minutes. The baseline sampling time of the mapping table is equal to the target time minus the initial value of the Yellow River side propagation delay. In addition to upstream flow and water level, downstream flow and water level, reservoir discharge, river section type and timestamp, the edge condition vector also includes Yellow River scheduling change indicators and the proportion of missing data markers. The proportion of missing data markers is used as a suppression factor input to the trainable mapper to reduce the variable time delay fluctuations caused by missing data.

[0053] Furthermore, the trainable mapper employs a three-layer fully connected structure. The input is an edge condition vector, with 64 neurons in the first hidden layer and 32 neurons in the second hidden layer. The activation function is a linear rectified function, and the output is a single scalar explanation of the Yellow River's variable time delay. When the Yellow River topological additive consistency tolerance is triggered, the trainable mapper undergoes a maximum of 20 parameter updates using an adaptive moment estimation optimizer with a fixed learning rate of 0.1%. After each update, the mapping table of the corresponding edge is overwritten into memory, and the mapping version number is recorded. The parameter update aims to minimize the sum of squared topological additive consistency residuals of all triggered three-node paths at the current target time. The residuals are recalculated after each update until the residuals do not exceed the Yellow River topological additive consistency tolerance or the maximum number of updates is reached.

[0054] Furthermore, the Yellow River topology additive consistency tolerance is fixed at 10 minutes, derived from the 95th percentile of the statistical error between the initial side propagation delay of typical flood events in the Yellow River over the past five years and the historical peak time difference. End-to-end mapping consistency verification enumerates the three-node path from any upstream node through intermediate nodes to the downstream node according to the path index. An iterative update is triggered when the difference exceeds 10 minutes, and iterative updates stop when the difference does not exceed 10 minutes or when 20 updates are reached.

[0055] To ensure that the variable time delay interpretation of the Yellow River is within the physically feasible range and consistent with the maximum propagation time delay of the Yellow River section, a bounded mapping is performed on the output of the trainable mapper. This generates a variable time delay interpretation that satisfies the amplitude limitation requirement. Constrained Yellow River variable time delay explanatory power:

[0056] in, Represents a node To the node The time delay explanation of the Yellow River variation. This indicates the target moment on the unified timeline of the Yellow River business. Represents directed connectivity The corresponding trainable mapper forward computation function. Indicates the target time The edge condition vector. This represents the hyperbolic tangent function, used to compress the mapper output to a bounded range.

[0057] When generating the mapping table from the target time to the upstream sampling time, a method is constructed for each target time. ,through After obtaining the unconstrained output, the following formula for interpreting the variable time delay in the Yellow River system is used: It is used to correct the baseline sampling time to reflect the dynamic changes in propagation delay.

[0058] To satisfy the monotonic constraint, ensuring that the upstream sampling time corresponding to the later target time is no earlier than the upstream sampling time corresponding to the previous target time, the mapping table is generated recursively using non-negative increments, as follows:

[0059] in, Indicates the target time The corresponding upstream sampling time mapping output. Indicates the target time The corresponding upstream sampling time mapping output. This indicates the time step for generating the mapping table. This represents a non-negative activation function, used to ensure that the increment is positive. This represents a monotonically increasing branch function, which can be used with... Shared parameters or independent configuration. The monotonic incremental branch function can share the first two layers of parameters with the trainable mapper or have an independent configuration.

[0060] When generating a mapping table of consecutive target times for the same directed connectivity relation, using Starting from the recursive base, it is generated according to the non-negative increment recursive generation formula. This ensures that the upstream sampling time corresponding to the next target time is not earlier than the upstream sampling time corresponding to the previous target time.

[0061] To incorporate river network topology constraints into time-mapping learning, end-to-end mapping consistency checks are performed on the three-node paths, with the Yellow River topology additive consistency tolerance used as the trigger condition. The topology additive consistency residual is expressed as:

[0062] in, Indicates the target time The path of the next three nodes The topology can be made consistent with residuals. Represents a node To the node The Yellow River variation time delay explanation quantity Represents a node To the node The time delay explanation of the Yellow River variation. Represents a node To the node The time delay explanation of the Yellow River variation. This represents the node identifier of the upstream, intermediate, and downstream nodes determined by the path index.

[0063] After updating the full edge mapping table at each target time step, traverse the three-node path using the path index and calculate... .when When the topological consistency tolerance of the Yellow River is exceeded, the parameters of the trainable mapper involving the edges are iteratively updated, and the updated Yellow River monotonically bounded time map is written to memory.

[0064] It should be noted that this step uses the river network topology as a priori, decomposalizing the propagation delay into a computable initial value for edge propagation delay and a learnable variable delay interpretation. First, the equivalent propagation velocity is derived from the river segment distance, river segment type, and near-window flow level and discharge, forming an initial value. Then, a mapping table from the target time to the upstream sampling time is generated using this initial value as a baseline. The mapping table is bounded and monotonic through amplitude limiting and non-negative incremental recursion, avoiding time reversal during alignment. Subsequently, the path index is used to perform end-to-end topological additive consistency checks on the three-node paths, triggering parameter iterative updates only for edges related to conflicting paths, achieving online adaptiveness. Compared to fixed lag or offline DTW, this invention can maintain the alignment consistency of the entire river network during scheduling abrupt changes and peak propagation velocity variations, and outputs a traceable variable delay interpretation to locate the source of propagation delay changes.

[0065] S3: Generate an upstream sequence aligned with the monotonic bounded time mapping of the Yellow River, and output the Yellow River rainfall and water situation fusion prediction results and the Yellow River variable time delay interpretation quantity.

[0066] Using the target time of the downstream node corresponding to the incoming edge as a benchmark, the upstream sampling time corresponding to the target time is obtained by calling the Yellow River monotonic bounded time mapping. In the Yellow River multi-source rainfall and water situation time series data, two adjacent timestamp samples of the upstream node are indexed, and an alignment feature vector is generated according to a preset interpolation resampling rule. Simultaneously, an interpolation confidence score corresponding to the alignment feature vector is generated. When only one valid sample can be indexed, an alignment feature vector is generated according to the boundary preservation rule, and the interpolation confidence score is calculated by extrapolating the time interval. When the upstream sampling time falls outside the coverage area of ​​the upstream node data, the nearest boundary sample is taken according to the boundary preservation rule. This process is repeated for the same incoming edge at consecutive target times to obtain an upstream sequence alignment sequence consistent with the downstream node's time axis. The upstream sequence alignment sequence is then associated and stored with the incoming edge identifier and the target time index.

[0067] Furthermore, the interpolation resampling rules employ consistent executable rules across different fields. For continuous quantities such as flow rate, water level, and reservoir discharge, linear interpolation is used. For fields stored as cumulative amounts over time periods, such as station rainfall and radar / satellite precipitation, the data is first converted to unit time intensity before linear interpolation, and then restored to cumulative amounts according to the target sampling interval. The unit time intensity is calculated using the difference between two adjacent original timestamps as the original sampling interval; if the original sampling interval is missing, it is calculated using a five-minute unified sampling interval. Category fields are preserved based on proximity. When the sample interval between two adjacent timestamps exceeds the Yellow River interpolation support interval of 15 minutes, the interpolation is deemed unreliable, and the field is marked as missing without interpolation. The 15-minute interval is taken as three times the five-minute unified sampling interval to cover common short-term delays.

[0068] Furthermore, the interpolation confidence score is generated jointly based on time proximity and data availability. When the target time coincides with the upstream sample timestamp, the confidence score is set to one. When interpolation is required, the confidence score decreases linearly with the time distance from the target time to the samples on both sides. When the trigger boundary maintains a regular pattern, the confidence score decays according to the boundary extrapolation time. When the boundary extrapolation time exceeds the maximum permissible time interval of 30 minutes maintained by the Yellow River boundary, the confidence score is set to zero and a missing data marker is added. The 30 minutes is derived from the 99th percentile of the transmission delay of the Yellow River rainfall and water level monitoring network over the past five years.

[0069] Furthermore, the association storage of the upstream sequence alignment sequence adopts a composite index key. The composite index key includes at least the downstream node identifier, the incoming edge identifier, and the target time index. Each record also stores the alignment feature vector, interpolation confidence, timestamps on both sides of the original sample, and missing test markers. When the same composite index key is written repeatedly, the latest mapping version number of the Yellow River monotonic bounded time mapping is used to overwrite the old record, thereby ensuring that subsequent fusion calls always read the alignment result consistent with the current mapping.

[0070] For each downstream node, at the same target time, the node feature sequence is aggregated and aligned with the upstream sequences corresponding to all incoming edges. Based on data delay information, missing data markers, and interpolation confidence, the fusion weights of each input channel are calculated, and a fusion feature sequence is generated. This fusion feature sequence is input into the Yellow River topological constraint variable time-delay aligned fusion multi-step sequence regression model to generate flow prediction sequences, water level prediction sequences, and peak arrival time prediction sequences within a preset prediction step size, serving as the Yellow River rainfall and water level fusion prediction results. The Yellow River variable time-delay interpreter output by the trainable mapper is bound by incoming edge identifiers and timestamps, and output synchronously with the Yellow River rainfall and water level fusion prediction results and stored in memory.

[0071] Furthermore, the calculation of the fusion weights involves calculating and normalizing the latency factor, missing data factor, and confidence factor for each input channel. The latency factor is calculated linearly by subtracting the ratio of data latency to the Yellow River data latency removal interval from 1, truncated to between 0 and 1. The missing data factor is either 0 or 1, and the confidence factor is the interpolated confidence level. Data latency is the difference between the server's data entry time and the data acquisition timestamp. The initial weights are the product of these three factors, normalized for all channels at the same target time. When the data latency exceeds 20 minutes of the Yellow River data latency removal interval, the channel weight is reset to zero and a channel removal flag is added. The 20-minute interval is derived from the 95th percentile of the real-time data entry latency statistics from the Yellow River Rainfall and Water Resources Management System over the past five years. Similarly, when the missing data flag indicates a missing key field or the interpolated confidence level is zero, the weight is also reset to zero. Initial weights for retained channels are calculated based on confidence level and delay factor, and then normalized. If the normalized channel weight is less than 0.05 of the lower limit weight of the Yellow River fusion channel, it is incorporated into the missing measurement processing to avoid weak signal disturbance. Key fields are the main measurement fields of the input channels: the rain gauge channel is the station rainfall field, the radar channel is the radar quantitative precipitation field, the satellite channel is the satellite precipitation estimation field, the hydrological channel is the control section flow and control section water level fields, and the reservoir channel is the reservoir discharge field. The lower limit weight of the Yellow River fusion channel is taken as the 5th percentile of the normalized weight distribution of the effective channels in the offline training samples and fixed at 0.05.

[0072] Furthermore, the Yellow River topologically constrained variable-delay aligned fusion multi-step sequence regression model is a computer-executable neural network, comprising a time encoder, a topology convergence layer, and a multi-step decoder connected in series. The time encoder receives the most recent 12 5-minute steps of the fused feature sequence as input, employing two layers of gated recurrent units (GRUs): the first layer has 64 hidden units and the second layer has 32 hidden units, outputting a time-encoded vector. The topology convergence layer performs attention-weighted convergence on the time-encoded vectors of each upstream aligned sequence based on the set of incoming edges from downstream nodes. The attention weights are prior to the fusion weights and are forced to be non-negative during normalization. The multi-step decoder is a two-layer fully connected structure, outputting a flow prediction sequence and a water level prediction sequence for the next 12 steps. Another branch converts the step length corresponding to the maximum value of the flow prediction sequence for the next 12 steps into a flood peak arrival time prediction sequence. The output is bound to the downstream node identifier and the target time index and written to memory. The flood peak arrival time prediction sequence is a one-dimensional sequence, whose elements are the target time plus the timestamp obtained by multiplying the step length corresponding to the maximum value by 5 minutes. The step size corresponding to the maximum value is the step number in the 12-step flow prediction sequence where the maximum value is taken.

[0073] Furthermore, model training employs a separation of offline training and online inference. Offline training uses historical flood events as samples, with the input being the fused feature sequence generated according to S1 and this step. The supervision labels are the measured flow and water level sequences of the corresponding downstream nodes and the peak arrival time extracted from the measured sequences. Before training, each continuous field is normalized according to the historical quantiles of the stations, and the normalization parameters are retained for reuse in online inference. Training uses mini-batch gradient descent with a batch size of 64. The optimizer uses adaptive moment estimation with a fixed learning rate of 0.1%. If the validation set loss does not decrease for five consecutive rounds, the training stops early, and the model parameters are fixed. During online inference, only the fixed parameters are loaded, and the model parameters are not updated, thus ensuring stable and reproducible real-time output. The peak arrival time is derived from the post-processing of the predicted flow sequence and does not participate in the backpropagation of network parameters. The training loss consists only of the mean square error between the predicted flow sequence and the predicted water level sequence. The peak arrival time is only used for offline evaluation and threshold calibration.

[0074] Furthermore, the synchronous output of the Yellow River variable time delay interpretation adopts a timestamp system consistent with the alignment sequence: for each incoming edge at each target time, the Yellow River variable time delay interpretation output by the trainable mapper is read and stored in conjunction with the incoming edge identifier, target time index, and mapping version number. When the prediction result is written to the memory, the set of Yellow River variable time delay interpretations for each incoming edge at the same target time is also written, so that the prediction result can be traced and matched with the corresponding alignment criteria, and supports retrieval and playback by downstream node, incoming edge identifier, and time range.

[0075] It should be noted that this step uses a monotonically bounded mapping to align the upstream multi-source time series to the downstream time axis, and generates reliable fusion features by combining interpolation confidence and delay missing measurement gating. Then, the multi-step sequence regression of topology convergence is used to regress the network output flow level and derive the peak arrival time. At the same time, variable delay interpretation is bound and stored simultaneously to achieve traceability of prediction output and alignment basis.

[0076] Example 2, an embodiment of the present invention, provides a Yellow River rainfall and water situation fusion prediction system, including a river network modeling and multi-source sequencing module, a variable time delay mapping learning and verification module, and an aligned fusion prediction and release module.

[0077] The river network modeling and multi-source sequencing module is used to construct the Yellow River network topology map and generate multi-source time-series data on Yellow River rainfall and water conditions.

[0078] The variable delay mapping learning and verification module is used to calculate the initial value of the propagation delay along the Yellow River based on the Yellow River network topology map, verify the consistency of the mapping, and learn the monotonic bounded time mapping of the Yellow River.

[0079] The alignment and fusion prediction and interpretation module is used to generate an upstream sequence alignment sequence according to the monotonic bounded time mapping of the Yellow River, and outputs the Yellow River rainfall and water level fusion prediction results and the Yellow River variable time delay interpretation quantity.

Claims

1. A method for fusion prediction of rainfall and water conditions in the Yellow River, characterized in that, include: Construct a topology map of the Yellow River network and generate multi-source time-series data on rainfall and water conditions of the Yellow River. The initial value of the propagation delay along the Yellow River is calculated based on the Yellow River network topology map. The consistency of the mapping is verified and the monotonic bounded time mapping of the Yellow River is learned. The upstream sequence is aligned with the monotonic bounded time mapping of the Yellow River, and the fusion prediction results of the Yellow River rainfall and water conditions and the time delay explanation of the Yellow River are output.

2. The Yellow River rainfall and water situation fusion prediction method as described in claim 1, characterized in that: The construction of the Yellow River network topology map includes, Read the codes of the main and tributary river sections of the Yellow River, station coordinates, control section identifiers and confluence relationship tables, and unify the rain gauge stations, hydrological stations, sub-basin nodes and reservoir control points into the node set. Establish directed connectivity between nodes according to the confluence direction, and mark the directed connectivity pointing to any downstream node as the incoming edge of the downstream node. For each directed connectivity relationship, the river segment distance, river segment category, and control section type are written as edge attributes; Perform topology consistency checks on the node set, and generate adjacency data structures and path index data structures, which are then stored in memory.

3. The Yellow River rainfall and water situation fusion prediction method as described in claim 2, characterized in that: The generated multi-source time-series data on rainfall and water conditions in the Yellow River includes... Rainfall sequences from rain gauge stations, quantitative precipitation sequences from radar, estimated precipitation sequences from satellites, flow sequences from control sections, water level sequences from control sections, and reservoir discharge sequences were collected. The sequences were then corrected for timestamps and their sampling intervals were standardized according to the unified timeline for Yellow River operations. Missing data is marked with missing data and outliers are removed according to preset outlier judgment rules. The retained data is assembled into a node feature sequence according to a fixed field order. The node feature sequence is associated with and stored with the site identifier and timestamp index.

4. The Yellow River rainfall and water situation fusion prediction method as described in claim 3, characterized in that: The calculation of the initial value of the Yellow River side propagation delay based on the Yellow River network topology map includes... For each directed connectivity relationship in the Yellow River network topology map, the corresponding river segment distance and river segment category are read. The flow sequence, water level sequence, and reservoir discharge sequence of the upstream and downstream of the river segment within the recent sliding window are extracted. The equivalent propagation speed is obtained according to the preset Yellow River segment category and propagation speed conversion rules. Then, the candidate propagation delay is calculated using the river segment distance and the equivalent propagation speed. A recalculation is triggered when a change in the Yellow River dispatching change flag is detected or the flow increment within the window exceeds the Yellow River flow change judgment value. The candidate propagation delay is limited and written into the memory according to the maximum propagation delay of the Yellow River section, and used as the initial value of the propagation delay along the Yellow River.

5. The Yellow River rainfall and water situation fusion prediction method as described in claim 4, characterized in that: The process of verifying mapping consistency and learning the Yellow River monotone bounded time mapping includes... For each directed connectivity relationship, a mapping table is established from the target time to the upstream sampling time. The mapping table uses the initial value of the Yellow River side propagation time delay as the baseline. The trainable mapper calculates the Yellow River variable time delay interpretation based on the edge condition vector. The edge condition vector includes the upstream flow rate and water level, the downstream flow rate and water level, the reservoir discharge, the river section category, and the timestamp. An amplitude limit rule is applied to the interpretation of the variable time delay in the Yellow River, so that the amplitude does not exceed the maximum propagation time delay of the Yellow River section; Apply non-negative increment constraints to the mapping output of adjacent target times, so that the upstream sampling time corresponding to the next time time is not earlier than the upstream sampling time corresponding to the previous time time, thus satisfying the monotonic constraint. The path index based on the Yellow River network topology map verifies the end-to-end mapping consistency from any upstream node to the downstream node via intermediate nodes. When the difference between the end-to-end Yellow River variable time delay interpretation and the sum of the two Yellow River variable time delay interpretations exceeds the Yellow River topology additive consistency tolerance, iterative updates of the trainable mapper parameters are triggered, and the updated Yellow River monotone bounded time mapping is written into memory.

6. The Yellow River rainfall and water situation fusion prediction method as described in claim 5, characterized in that: The process of generating upstream sequence alignment sequences according to the Yellow River monotone bounded time mapping includes... Based on the target time of the downstream node corresponding to the incoming edge, the upstream sampling time corresponding to the target time is obtained by calling the Yellow River monotone bounded time mapping. In the Yellow River rainfall and water conditions multi-source time series data, the two adjacent timestamp samples of the upstream node are indexed and the aligned feature vector is generated according to the preset interpolation resampling rules. At the same time, the interpolation confidence corresponding to the aligned feature vector is generated. When the upstream sampling time falls outside the data coverage of the upstream node, the nearest boundary sample is taken according to the boundary preservation rule. Repeatedly execute the process on the same incoming edge at consecutive target times to obtain an upstream sequence alignment sequence that is consistent with the time axis of the downstream node, and store the upstream sequence alignment sequence in association with the incoming edge identifier and the target time index.

7. The Yellow River rainfall and water situation fusion prediction method as described in claim 6, characterized in that: The output of the Yellow River rainfall and water level fusion prediction results and the Yellow River time delay explanation quantity include, For each downstream node, the node feature sequence is aggregated and aligned with the upstream sequence corresponding to all incoming edges at the same target time. The fusion weight of each input channel is calculated based on data delay information, missing test markers and interpolation confidence, and a fusion feature sequence is generated. The fused feature sequence is input into the Yellow River topological constraint variable time delay aligned fusion multi-step sequence regression model to generate the flow prediction sequence, water level prediction sequence and flood peak arrival time prediction sequence within the preset prediction step size range as the Yellow River rainfall and water situation fusion prediction result. The time delay interpretation of the Yellow River output by the trainable mapper is bound to the timestamp by the incoming edge identifier, and then output synchronously with the Yellow River rainfall and water level fusion prediction results and stored in the memory.

8. A Yellow River rainfall and water level fusion prediction system, employing the Yellow River rainfall and water level fusion prediction method as described in any one of claims 1 to 7, characterized in that: It includes a river network modeling and multi-source sequencing module, a variable time delay mapping learning and verification module, and an alignment fusion prediction and release module; The river network modeling and multi-source sequencing module is used to construct the Yellow River network topology map and generate multi-source time-series data on Yellow River rainfall and water conditions. The variable delay mapping learning and verification module is used to calculate the initial value of the propagation delay along the Yellow River based on the Yellow River network topology map, verify the mapping consistency and learn the monotonically bounded time mapping of the Yellow River. The alignment fusion prediction and interpretation module is used to generate an upstream sequence alignment sequence according to the Yellow River monotonically bounded time mapping, and outputs the Yellow River rainfall and water level fusion prediction results and the Yellow River variable time delay interpretation quantity.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the Yellow River rainfall and water situation fusion prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the Yellow River rainfall and water situation fusion prediction method as described in any one of claims 1 to 7.

Images