Spatiotemporal prediction method and system for state of drainage system based on neural network

Abstract

The application discloses a neural network-based space-time prediction method and system for the state of a drainage system, and belongs to the technical field of computer system modeling and artificial intelligence, which comprises the following steps: obtaining pipe network geographic structure data and operation observation data, and performing time labeling, event-driven space-time alignment and missing data completion to generate an event alignment sequence; fusing pipeline connection relationship and flow direction rules to construct a reachable topological graph; combining the event alignment sequence to perform joint feature extraction, generating a topological space-time representation tensor, and applying constraints based on mass conservation and capacity limitation to calculate and generate a physical constraint deviation amount; fusing the topological space-time representation tensor and the physical constraint deviation amount to generate a space-time prediction field, and performing secondary mapping based on uncertainty quantification and error feedback to generate an adaptive correction field. Through the integration of event-driven data alignment, physical topology construction, constraint fusion learning and error correction, the future state of the drainage system can be predicted in space-time.

Description

Technical Field

[0001] This invention relates to the fields of computer system modeling and artificial intelligence technology, and in particular to a spatiotemporal prediction method and system for the state of drainage systems based on neural networks. Background Technology

[0002] With the advancement of smart city construction, the intelligent management of urban drainage systems has become crucial. Large-scale drainage systems are typical complex dynamic networks, whose states, such as pipeline flow and manhole levels, are influenced by various factors including rainfall, tides, and upstream water inflow, exhibiting highly nonlinear spatiotemporal coupling characteristics. Utilizing artificial intelligence technology, particularly neural network-based computational methods, to accurately predict the future state of drainage systems is of great significance for achieving urban flooding early warning, optimizing pump station scheduling, and improving emergency response efficiency.

[0003] In existing technologies, the prediction of drainage system status mainly relies on two types of methods. One type is the traditional simulation method based on hydraulic models, such as using software like the SWMM (Sudden Storm Flood Management Model). While the physical mechanisms of these methods are well-defined, they are computationally extremely costly on large-scale pipe networks and require extensive fine-tuning of parameters, making them unsuitable for real-time prediction. The other type is the data-driven method, which directly employs neural network models, such as Long Short-Term Memory (LSTM) networks or Convolutional Neural Networks (CNNs), to learn from historical data from monitoring points to predict the future. These methods can, to some extent, capture the dynamic trends of time-series data.

[0004] However, existing data-driven methods typically assume that the input data is complete and uniformly sampled. In reality, pipeline network operation observation data often suffers from inconsistent sampling frequencies, communication delays, and data gaps, leading to performance degradation when directly applied. Secondly, most neural network models, when processing pipeline network data, either treat it as an isolated time series or use only simple geographical adjacency relationships, failing to effectively incorporate the physical topology of the pipeline network and the directional rules that water flow must follow. This results in a lack of understanding of the overall hydraulic transmission process of the system. Finally, purely data-driven neural network models lack the guidance of physical laws during training. While their predictions may be statistically reasonable, they often violate basic mass conservation or capacity constraints, producing physically invalid predictions. This significantly limits their reliability in practical engineering decision-making. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a spatiotemporal prediction method and system for the state of drainage systems based on neural networks. By integrating event-driven data alignment, physical topology construction, constraint fusion learning, and error correction, it is possible to achieve spatiotemporal prediction of the future state of drainage systems.

[0006] The above objectives can be achieved through the following approach:

[0007] A spatiotemporal prediction method for the state of a drainage system based on neural networks includes: acquiring geographical structure data and operational observation data of the pipeline network, and time-annotating the data to generate a time axis annotation set; performing event-driven spatiotemporal alignment and missing data completion on the time axis annotation set to generate an event alignment sequence; constructing a reachable topology graph with directionality and connectivity based on the event alignment sequence and fusing pipeline connection relationships and flow direction rules; performing joint feature extraction on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor; applying constraints based on mass conservation and capacity limitations to the topological spatiotemporal representation tensor to calculate and generate a physical constraint deviation; fusing the topological spatiotemporal representation tensor and the physical constraint deviation to generate a spatiotemporal prediction field; performing uncertainty quantification and error feedback on the spatiotemporal prediction field to generate correction control parameters, and using the correction control parameters to perform a secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field.

[0008] Optionally, generating the time axis annotation set includes: acquiring pipeline network geographical structure data and operational observation data, and performing data format transformation to generate a unified format set; cleaning abnormal and duplicate records in the unified format set, and performing signal validity screening to generate a valid observation set; and parsing the timestamps and sampling intervals of the valid observation set to generate the time axis annotation set.

[0009] Optionally, generating the event alignment sequence includes: identifying the event trigger time and associated time window in the time axis annotation set, and generating a neighborhood group set by spatial neighborhood grouping; correcting the sampling delay and clock drift of the neighborhood group set, and interpolating intermittent missing segments to generate a complete alignment set; and fusing the event boundaries and spatial adjacency relationships of the complete alignment set to generate the event alignment sequence.

[0010] Optionally, the step of constructing a reachable topology graph with directionality and connectivity based on the event alignment sequence and by fusing pipeline connection relationships and flow direction rules includes: obtaining the node set and pipeline connection information of the event alignment sequence, and performing connectivity component analysis to generate a connectivity graph; calculating the elevation difference and edge direction indication in the connectivity graph, and performing reachability derivation to generate a reachable relation set; combining the reachable relation set and pipeline connection information to set directionality and connectivity, and generating a reachable topology graph.

[0011] Optionally, the method further includes: extracting pipeline connection information and performing noise edge filtering to generate an edge cleanup set; fusing the edge attributes of the edge cleanup set with the flow rate of change of the event alignment sequence, and performing edge direction normalization to generate a direction normalization matrix; verifying the consistency between the direction normalization matrix and the reachability set, and correcting the reachability set based on the verification result.

[0012] Optionally, generating the topological spatiotemporal representation tensor includes: extracting time series data from the event alignment sequence to generate an observation attribute matrix, and extracting geometric and connectivity metrics from the reachable topology graph to generate a topological attribute matrix; mapping the observation attribute matrix and the topological attribute matrix to nodes and edges of the reachable topology graph, and performing time window slicing to generate a time slice set; and fusing the time dependency and spatial neighborhood relationship of the time slice set to generate the topological spatiotemporal representation tensor.

[0013] Optionally, generating the physical constraint deviation includes: extracting the node inflow and outflow attributes of the topological spatiotemporal representation tensor and performing node mass conservation calculations to generate a conservation deviation spectrum; extracting the edge carrying attributes and node accumulation attributes of the topological spatiotemporal representation tensor and performing feasible region delineation to generate a feasible region boundary; comparing the deviation of the conservation deviation spectrum from the feasible region boundary in the node and edge dimensions to generate the physical constraint deviation.

[0014] Optionally, the generation of the spatiotemporal prediction field includes: constructing a graph representation of the topological spatiotemporal representation tensor and a constraint representation of the physical constraint deviation, and performing normalization processing to generate normalized input pairs; fusing the normalized input pairs and performing temporal decoding to generate a temporal decoding sequence; and decoding the future state estimates of the temporal decoding sequence in the node and edge dimensions to generate the spatiotemporal prediction field.

[0015] Optionally, the step of performing uncertainty quantification and error feedback on the spatiotemporal prediction field to generate correction control parameters, and using the correction control parameters to perform secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field includes: calculating the temporal deviation between the spatiotemporal prediction field and the event alignment sequence, and performing error source decomposition to generate an error source vector; quantizing and mapping the uncertainty in the error source vector, and then performing stability constraints to generate correction control parameters; and applying the correction control parameters to the estimation of nodes and edges in the spatiotemporal prediction field to generate an adaptive correction field.

[0016] Based on the same inventive concept, this invention also provides a spatiotemporal prediction system for the state of a drainage system based on a neural network. The system includes: a data collection and annotation module for acquiring pipeline network geographical structure data and operational observation data, and time-annotating the data to generate a timeline annotation set; an event alignment module for performing event-driven spatiotemporal alignment and missing data completion on the timeline annotation set to generate an event alignment sequence; a topology construction module for constructing a reachable topology graph with directionality and connectivity based on the event alignment sequence and integrating pipeline connection relationships and flow direction rules; a joint representation module for performing joint feature extraction on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor; a physical constraint module for applying constraints based on mass conservation and capacity limitations to the topological spatiotemporal representation tensor and calculating a physical constraint deviation; a fusion prediction module for fusing the topological spatiotemporal representation tensor and the physical constraint deviation to generate a spatiotemporal prediction field; and an adaptive correction module for performing uncertainty quantification and error feedback on the spatiotemporal prediction field, generating correction control parameters, and using the correction control parameters to perform a secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field.

[0017] Compared with the prior art, the present invention has the following advantages:

[0018] 1. This invention solves the problems of heterogeneous, asynchronous, and incomplete raw data caused by inconsistent sensor sampling frequencies, network transmission delays, and equipment failures in large-scale drainage systems by introducing an event-driven spatiotemporal alignment and data completion mechanism. It can process raw observation data into a high-quality dataset that is synchronous in time, logically related in space, and complete and continuous in data, providing a solid and reliable data foundation for the training of subsequent neural network models and ensuring the accuracy of prediction from the source.

[0019] 2. This invention constructs a reachable topology map that integrates static geographic information and dynamic hydraulic flow rules of the pipeline network. This physical topology information is then jointly represented with time-series data as an input tensor. This allows the neural network model to transcend simple data fitting and gain a deeper understanding of the physical connectivity of the pipeline network and the true motion patterns of water flow. This integration of prior physical knowledge enhances the model's ability to capture the coordinated dynamic behavior of the entire system, improving the interpretability of the prediction model and its reasoning ability regarding complex hydraulic phenomena.

[0020] 3. This invention quantifies fundamental physical laws such as mass conservation and capacity limitation into physical constraint deviations, and uses these deviations as parallel inputs to the neural network fusion module, thereby imposing strong physical constraints during model training. This avoids the model producing predictions that violate common sense physics, such as node water volume appearing or disappearing out of thin air, or pipeline flow far exceeding its physical carrying capacity limit. This enhances the authenticity and credibility of the prediction results, allowing the prediction output to be directly used for critical engineering decisions.

[0021] 4. This invention can assess the uncertainty of its own predictions in real time and continuously self-correct based on the deviation between the prediction and the actual observation, dynamically generating correction parameters to perform a secondary mapping of the original prediction field. This gives the prediction system excellent online adaptive capabilities, enabling it to cope with sensor drift or changes in pipeline characteristics, thereby maintaining long-term prediction accuracy and system robustness.

[0022] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the spatiotemporal prediction method for the state of a drainage system based on a neural network, according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the effect of physical constraints on the model prediction results in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram illustrating the optimization effect of long-term prediction error in an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the spatiotemporal prediction system for the state of a drainage system based on a neural network, according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 One embodiment of the present invention proposes a spatiotemporal prediction method for the state of a drainage system based on a neural network. By integrating event-driven data alignment, physical topology construction, constraint fusion learning, and error correction, it is possible to make spatiotemporal predictions of the future state of the drainage system.

[0030] The method described in this embodiment specifically includes:

[0031] Acquire pipeline network geographical structure data and operational observation data, and generate a time axis annotation set by time-marking the data;

[0032] Event-driven spatiotemporal alignment and missing data completion are performed on the timeline annotation set to generate an event alignment sequence;

[0033] Based on the event alignment sequence, a reachable topology graph with directionality and connectivity is constructed by integrating pipeline connection relationships and flow direction rules;

[0034] Joint feature extraction is performed on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor;

[0035] Constraints based on mass conservation and capacity limitation are applied to the topological spatiotemporal representation tensor, and physical constraint deviations are calculated.

[0036] The topological spatiotemporal representation tensor is fused with the physical constraint deviation to generate a spatiotemporal prediction field;

[0037] Uncertainty quantification and error feedback are performed on the spatiotemporal prediction field to generate correction control parameters, and the spatiotemporal prediction field is then mapped twice using the correction control parameters to generate an adaptive correction field.

[0038] This method constructs a temporally synchronized and data-complete event alignment sequence by performing time-labeling, event-driven spatiotemporal alignment, and missing data completion operations on the acquired pipeline network geographic and operational data. Based on this sequence, and by integrating the physical connections between pipelines and hydraulic flow direction rules, the system establishes a reachable topology map that reflects the flow path and connectivity of the entire system. Subsequently, the system performs joint feature extraction on the dynamic temporal information in the event alignment sequence and the static structural information in the reachable topology map to generate a topological spatiotemporal representation tensor containing spatiotemporal dynamics and topological correlations. To ensure the physical accuracy of the prediction results, the system verifies this tensor according to the law of mass conservation and pipe capacity constraints, and quantifies its deviation from physical reality to generate a physical constraint deviation. Finally, the representation tensor carrying spatiotemporal information and the deviation representing physical rules are fused to generate an initial spatiotemporal prediction field. By quantifying the uncertainty of the spatiotemporal prediction field and performing error feedback, correction control parameters are generated. These parameters are then used to perform a secondary mapping on the initial prediction field, ultimately outputting a prediction result that has undergone adaptive correction.

[0039] This method addresses the data heterogeneity and incompleteness issues in large-scale drainage systems caused by inconsistent sensor sampling frequencies, data transmission delays, or missing data through event-driven spatiotemporal alignment and completion, thus providing a high-quality input foundation for the model. By constructing a reachable topology graph that integrates flow direction rules and jointly representing this topology information with time-series data, the model moves beyond simple data fitting to understanding the physical connectivity of the pipe network and the direction of fluid movement, enhancing the ability to capture the overall system behavior. Quantifying and fusing prior physical knowledge such as mass conservation and capacity constraints into physical constraint deviations avoids predictions that violate physical principles, enhancing the realism and credibility of the prediction results. Finally, by introducing an adaptive correction closed loop with uncertainty quantification and error feedback, the system acquires self-correction and optimization capabilities, dynamically adapting to changes in system state, continuously improving prediction accuracy, and providing a more reliable basis for subsequent control decisions.

[0040] Optionally, the generation of the timeline annotation set includes:

[0041] Acquire pipeline network geographical structure data and operational observation data, and perform data format transformation to generate a unified format set;

[0042] Clean the abnormal and duplicate records in the unified format set, and perform signal validity screening to generate a valid observation set;

[0043] The timestamps and sampling intervals of the valid observation set are analyzed to generate a time axis label set.

[0044] Specifically, this step transforms diverse and formatted raw data into a unified, reliable dataset with precise time stamps, providing standardized data input for subsequent spatiotemporal prediction modeling. First, the system performs data format transformation to generate a unified format set. This operation addresses the heterogeneity of data sources, performing a unified parsing and mapping process for acquired pipeline geographic structure data, such as GIS Shapefiles, and operational observation data, such as CSV logs or database tables from SCADA systems. The system maps field names from different data sources, such as "flow rate" and "instantaneous flow rate," to the standard "FlowRate," and converts all time records from their respective time zones and formats, such as Unix timestamps or "year-month-day hour:minute:second" strings, to ISO8601 standard UTC time strings. Simultaneously, it unifies the units of physical quantities such as flow rate and liquid level to international standard units, thereby generating a unified format set with a standardized data structure and unified semantics.

[0045] Next, the system cleanses and filters this unified data set to generate a valid observation set. Invalid data caused by sensor malfunction, communication interference, or recording errors is removed. The cleaning process includes filtering out abnormal records; for example, for liquid level data, the system removes any values ​​outside a preset physical range, such as the range between 0.1 meters and 2.5 meters above and below the pipe diameter. Simultaneously, the system identifies and deletes completely duplicate records based on a combination of sensor identifier, timestamp, and measurement value. Following this is signal validity filtering, which filters out signals that, although within the measurement range, actually reflect a sensor "dead" or communication interruption. The system determines validity by calculating the signal variance within a sliding time window, considering the signal valid only if the variance exceeds a preset threshold. This process can be represented by the following formula:

[0046] ,

[0047] in, This represents the log of a signal within a certain time window. For the k-th observation within this window, This is the average of all observations within the window. The number of data points included in the window is typically set to cover an observation period of 1 to 2 hours, for example, 12 to 24 data points. Records within this time window are considered valid and included in the valid observation set only when the calculated variance V is greater than a positive threshold Vthresh set according to the sensor's accuracy.

[0048] Finally, the system generates a timeline annotation set by parsing the timestamps and sampling intervals of the valid observation set. This clarifies the precise location of each valid data point on the timeline and its temporal relationship with adjacent data points. The system iterates through each record in the valid observation set, extracts its standardized timestamp as its coordinates on the global timeline, and calculates the time difference between the current record and the previous valid record, using this difference as the sampling interval. The resulting timeline annotation set is a collection of structured information including sensor ID, standardized observation values, precise UTC timestamps, and sampling intervals. It lays a solid time series foundation for subsequent event-driven spatiotemporal alignment and missing data completion processes.

[0049] Optionally, the generation of the event alignment sequence includes:

[0050] Identify the event trigger times and associated time windows in the timeline annotation set, and generate a neighborhood group set by spatial neighborhood grouping;

[0051] Specifically, the system transforms the multiple sensor data streams generated in the preceding steps, which are not coordinated in time, into a complete sequence centered on a specific hydraulic event, fully aligned in both time and space. First, the system performs event identification and spatial neighborhood grouping to generate a neighborhood grouping set. The core of this operation is identifying key disturbances in the entire pipeline network system, i.e., the event trigger moment. The system continuously analyzes the flow data of key monitoring points, such as the city's main outfall, and uses a rate of change detection algorithm to identify moments when the flow rate rises sharply within a short period as the event trigger moment. Once the trigger moment is identified, the system immediately defines a correlation time window, typically covering 30 to 60 minutes prior to the event to obtain the system baseline state before the event, and 3 to 5 hours backward to fully capture the entire process of the event's occurrence, development, and dissipation. Simultaneously, based on the geographical information of the pipeline network, the system applies clustering algorithms such as DBSCAN to group physically adjacent monitoring points, such as those within a 500-meter range upstream and downstream of the same main pipeline, into the same spatial neighborhood, thereby generating a neighborhood grouping set. This set organizes the scattered monitoring point data into several data clusters related to the specific hydraulic event.

[0052] The sampling delay and clock drift of the neighborhood group set are corrected, and interpolation is performed on the intermittent missing segments to generate a complete and aligned set;

[0053] Specifically, the system performs time correction and data interpolation on the neighborhood group set to generate a complete alignment set, eliminating the inherent sampling delay and clock drift of different sensors, and filling in intermittent missing segments generated during data transmission. Within each group, the system selects a monitoring point with the highest signal quality or the most stable sampling frequency as a reference, and then accurately estimates the sampling delay and clock drift of each sequence by calculating the cross-correlation function between the reference sequence and other sequences. The calculation process of this delay can be expressed by the following formula:

[0054] ,

[0055] In this formula, M is the time delay that maximizes the cross-correlation value, St is the sequence to be corrected, Sr is the reference sequence within the group, i represents the time point index in the sequence, and d represents the attempted time shift. Based on the calculated delay M, the system shifts the timestamps of non-reference sequences to achieve precise temporal alignment of all data streams within the group. After alignment, the system then processes intermittent missing segments in the data. For short-term missing segments with a length of 5 sampling periods or less, the system uses cubic spline interpolation to fill them; for longer data gaps, it uses data from other sensors in the aligned neighborhood and interpolates using methods such as spatiotemporal kriging interpolation to generate a complete and time-synchronized aligned set.

[0056] The event alignment sequence is generated by fusing the event boundaries and spatial adjacency relationships of the completed alignment set.

[0057] Specifically, the system integrates the event boundaries and spatial adjacency relationships of the complete alignment set to generate the final event alignment sequence. The processed data is then solidified into a structured standard sequence containing rich contextual information. The system converts the timestamps of all data points in the complete alignment set from absolute time to relative time relative to the event trigger moment, thus unifying all grouped data into a relative time coordinate system centered on the event. Simultaneously, the system attaches the start and end information of the associated time window as event boundary metadata and the sensor adjacency relationships determined during grouping as spatial adjacency metadata to the time series. The final generated event alignment sequence is a high-dimensional data structure that not only contains complete and synchronized time-series data from all relevant monitoring points but also explicitly encodes the event background to which these data belong and their topological relationships in physical space, providing ideal input for subsequent construction of neural network models capable of understanding the dynamic behavior of the pipeline network.

[0058] Optionally, the step of constructing a reachable topology graph with directionality and connectivity based on the event alignment sequence and by fusing pipeline connection relationships and flow direction rules includes:

[0059] Obtain the node set and pipeline connection information of the event alignment sequence, and perform connectivity component analysis to generate a connectivity graph;

[0060] Specifically, discrete pipeline connection information is integrated with hydraulic rules to construct a directed graph structure that accurately represents the actual direction of water flow and the overall network connectivity, providing a physical topological foundation for subsequent model understanding of system dynamics. First, the system performs connectivity component analysis to generate a connectivity graph. This operation aims to identify physically isolated subsystems within the pipeline network. The system extracts all monitoring points, manholes, and confluence points as a node set from the event alignment sequence generated in the previous step, and obtains pipeline connection information, i.e., a list of pipe segments connecting each node, from the pipeline network's geographical structure data. Based on these nodes and edges, the system constructs an undirected graph and executes connectivity component analysis algorithms such as depth-first search or breadth-first search to divide the entire pipeline network into several disconnected subgraphs. Each subgraph is a connected component, and the set of these subgraphs constitutes the connectivity graph. This step ensures that subsequent flow direction determination is performed only within physically connected regions.

[0061] Calculate the elevation difference and edge direction indication in the connectivity graph, and perform reachability derivation to generate a set of reachability relations;

[0062] Specifically, the system derives reachability based on elevation difference and edge direction indicators to generate a set of reachability relationships, assigning a direction conforming to the physical laws of gravity flow or pumped flow to each pipeline in the connectivity graph. The system traverses each edge in the connectivity graph, connecting node u and node v, and queries the pipe bottom elevation between the two nodes from the geographic structure data. The system determines the flow direction indicator Fuv through the following conditional logic:

[0063] If Hu minus Hv is greater than Ht, then Fuv equals 1;

[0064] If Hv minus Hu is greater than Ht, then Fuv equals -1;

[0065] If the absolute value of Hu minus Hv is less than or equal to Ht, then Fuv equals Puv;

[0066] Here, Fuv represents the flow direction indicator from node u to v, where 1 indicates flow from u to v and -1 indicates flow from v to u. Hu and Hv are the pipe bottom elevations of nodes u and v, respectively, obtained from the pipeline network geographic data. Ht is a preset elevation difference threshold used to determine whether the pipeline is approximately horizontal; its value is typically between 0.05 meters and 0.1 meters to filter out minor elevation changes caused by measurement errors. If the elevation difference is insufficient to clearly determine the direction of gravity flow, the system will use Puv, a predefined edge direction indicator from the pipeline network data. This indicator is typically used to mark the pumping direction of pumping stations or special pipe sections with unclear design flow directions. By performing this determination on all edges, the system generates a dataset consisting of ordered node pairs, thus clarifying the reachability set of all potential flow paths.

[0067] The reachability set and pipeline connection information are combined to set directionality and connectivity, generating a reachable topology graph.

[0068] Specifically, the system combines the reachability relation set with the original pipeline connection information to generate the final reachable topology graph. The derived logical flow is then solidified into the graph structure of the physical pipeline network. The system creates a new directed graph data structure. It traverses each ordered node pair in the reachability relation set, for example, from u to v, and then creates a directed edge from node u to node v in the new directed graph. Simultaneously, the system queries the physical pipe segment attributes connecting u and v from the original pipeline connection information, such as pipe diameter, length, slope, and material, and appends these attributes to the newly created directed edge. After completing the creation of all edges and attribute appending, a reachable topology graph containing all nodes, directed edges, and their physical attributes, exhibiting both directionality and connectivity, is constructed. This provides a structured input with physical implications for subsequent spatiotemporal feature extraction.

[0069] Optionally, the method further includes:

[0070] Extract pipeline connection information and perform noise filtering to generate a clean set;

[0071] Specifically, the system utilizes dynamic observation data to verify and correct the reachability set constructed from static geographic information in the previous steps, thereby generating a more accurate pipeline network topology that reflects real hydraulic behavior. First, the system performs noise edge filtering to generate a cleaned edge set. The goal of this operation is to eliminate non-physical connections introduced into the pipeline network geographic structure data due to digitization errors or modeling simplifications. The system iterates through all pipeline connection information and filters according to preset rules. For example, it removes redundant short pipe segments shorter than a certain threshold, such as 0.5 meters, or deletes suspended edges that connect only one node; these edges do not exist in the actual pipeline network and are considered topological noise. After this processing, the resulting cleaned edge set represents physically valid and necessary pipeline connections.

[0072] The edge attributes of the edge purification set are combined with the flow change rate of the event alignment sequence, and edge direction normalization is performed to generate a direction normalization matrix.

[0073] Specifically, the system integrates the edge attributes of the edge purification set with the flow rate of change of the event alignment sequence, performs edge direction normalization to generate a direction normalization matrix, and infers the flow direction from the actual operating data. For each edge in the edge purification set connecting two nodes i and j, the system extracts their corresponding flow time series data from the event alignment sequence and calculates their first difference, i.e., the flow rate of change. Subsequently, the system determines the dominant flow direction by calculating the time-delay cross-correlation between these two rate of change sequences, and its directionality score can be quantified by the following formula:

[0074] ,

[0075] in This represents the directional score from node i to node j in the directional normalization matrix. and τ represents the flow change rate at nodes i and j, respectively, obtained directly from the event alignment sequence. t represents the time step, and τ is the time lag between the two sequences. The system searches for τ within a preset physically feasible range, such as 0 minutes to 15 minutes, to find the τ value that maximizes the sum of cross-correlation. The SGN function takes the sign of the τ value corresponding to this maximum correlation. If it is positive, it indicates that the change at i precedes that at j, the flow direction is from i to j, and Wij is 1; if it is negative, the flow direction is from j to i. The value is -1. The system performs this calculation on all connected node pairs, ultimately generating an N-by-N directional normalized matrix with elements of 1, -1, or 0, where N is the total number of nodes.

[0076] The consistency between the direction normalization matrix and the reachability relation set is verified, and the reachability relation set is corrected based on the verification result.

[0077] Specifically, the system verifies the consistency between the direction normalization matrix and the reachability relation set, and corrects the reachability relation set based on the verification results. It uses data-driven conclusions to correct inferences based on static rules, especially for pipe sections with complex hydraulic phenomena such as pressure flow and siphoning. The system traverses each directed edge in the reachability relation set, for example, from node u to v, where the direction is determined by the elevation difference. Then, the system queries the corresponding score Wuv in the direction normalization matrix. If the flow direction determined by the elevation is inconsistent with the flow direction inferred from the data (i.e., Fuv multiplied by Wuv is less than zero, and the absolute value of Wuv, representing the confidence level, is greater than a preset confidence threshold T, such as 0.8), the system determines that the static rule inference is incorrect. Under this trigger condition, the system reverses the direction of that edge in the reachability relation set, correcting it from u to v to v to u. Through this series of verification and correction processes, the final reachability relation set can more accurately reflect the water flow organization pattern of the pipe network in real-world operational events.

[0078] Optionally, the generated topological spatiotemporal representation tensor includes:

[0079] Extract time series data from the event alignment sequence to generate an observation attribute matrix, and extract geometric and connectivity metrics from the reachable topology graph to generate a topology attribute matrix;

[0080] Specifically, the system integrates the dynamic operating status and static topology information of the pipeline network to construct a high-dimensional data structure containing rich spatiotemporal context, namely the topological spatiotemporal representation tensor, which can be directly processed by a neural network model. First, the system extracts dynamic observation data and static topology information to generate an observation attribute matrix and a topology attribute matrix. This operation aims to structurally aggregate heterogeneous data sources. The system traverses the event alignment sequence, extracts time-varying parameters such as flow rate, liquid level, and water quality at each time step for all monitoring nodes, and organizes them into an observation attribute matrix. The rows of this matrix represent time steps, and the columns represent different monitoring nodes and their observation indicators. Simultaneously, the system analyzes the constructed reachable topology graph, calculating the static geometry and connectivity metrics for each node and each edge. Node metrics include in-degree and out-degree, while edge metrics include pipe length, pipe diameter, slope, and material roughness coefficient. This static information is organized into a topology attribute matrix, with its rows and columns corresponding to the nodes and edges in the graph and their respective static attributes.

[0081] Map the observation attribute matrix and the topology attribute matrix to the nodes and edges of the reachable topology graph, and perform time window slicing to generate a time slice set;

[0082] Specifically, the system performs attribute mapping and time window slicing to generate a time slice set, combining the separate attribute matrix with the topological graph structure and segmenting the data according to the requirements of time series modeling. The system maps the data at each time step in the observed attribute matrix (i.e., one row of the matrix) to the corresponding node in the reachable topological graph, serving as the dynamic feature of that node at that moment. Simultaneously, the static features in the topological attribute matrix are permanently attached to the corresponding nodes and edges of the graph. After mapping, the system slides a time window of a preset length along the time axis with a fixed step size. The length of this window, for example, 60 minutes, defines the duration of historical information relied upon by the model for a single prediction; the step size, for example, 15 minutes, determines the generation density of training samples. With each slide, the system captures the dynamic feature sequences of all nodes within the window, as well as the static topological structure of the entire graph, forming a graph snapshot containing spatiotemporal information. The collection of all these graph snapshots constitutes the time slice set.

[0083] The temporal dependence and spatial neighborhood relationship of the time slice set are fused to generate a topological spatiotemporal representation tensor.

[0084] Specifically, the system integrates temporal dependencies and spatial neighborhood relationships within the time slice set to generate the final topological spatiotemporal representation tensor. This operation aims to organize discrete time slices into a continuous, high-dimensional array representing the spatiotemporal evolution process. The system stacks the graph snapshots in the time slice set in chronological order. The resulting tensor can be represented as X, typically with dimensions T x N x F. Here, T represents the number of steps within the time window, inherited from the time window slicing operation; N is the total number of nodes in the reachable topology graph, determined by the network size; and F is the total number of features contained in each node, defined by the following formula:

[0085] ,

[0086] Here, Fd represents the number of dynamic features for each node, such as flow rate and liquid level, and this value is determined by the types of monitoring indicators in the event alignment sequence. Fs represents the number of static features associated with each node, such as the node's degree or the attribute of the nearest edge, and this value is determined by the number of columns in the topological attribute matrix. The resulting topological spatiotemporal representation tensor X has each element precisely corresponding to a specific feature value of a specific node at a specific time, and the structure of the entire tensor implicitly contains the spatial adjacency relationships and temporal evolution relationships between nodes, providing a complete and structured input for subsequent graph neural network models.

[0087] Optionally, the generation of physical constraint deviation includes:

[0088] Extract the node inflow and outflow attributes of the topological spatiotemporal representation tensor, and perform node mass conservation calculations to generate a conservation bias spectrum;

[0089] Specifically, the deviation between the topological spatiotemporal representation tensor generated in the previous step and the fundamental physical laws of the drainage system is quantified, forming a constraint signal that can be used to guide the neural network learning. First, the system performs node mass conservation calculations to generate a conservation deviation spectrum. This operation aims to ensure that the water volume change at any node in the graph conforms to the law of conservation of mass. The system traverses each time step and each node i in the topological spatiotemporal representation tensor and extracts the inflow and outflow attributes of that node from the tensor. (Node inflow attributes...) It is the sum of the flows of all edges pointing to node i, while the outflow property is... It is the sum of the flows of all edges originating from node i. Simultaneously, the system also extracts the storage attribute of node i itself, namely the rate of change of water storage:

[0090] ,

[0091] in, Let i be the rate of change of water level. Let be the effective catchment area of ​​node i. Then, the system calculates the mass conservation deviation of node i at that moment using the following formula:

[0092] ,

[0093] in This is the conserved deviation value of node i. Theoretically, for a closed physical system, It should always be equal to zero. The system calculates this deviation value for all nodes at all time steps, forming a set of matrices or vectors corresponding to the dimensions of the topological spatiotemporal representation tensor, i.e., the conserved deviation spectrum, which directly reflects the degree to which the data conforms to the law of conservation of mass.

[0094] Extract the edge-bearing attributes and node accumulation attributes of the topological spatiotemporal representation tensor, and perform feasible region delineation to generate feasible region boundaries;

[0095] Specifically, the system generates feasible region boundaries by defining the feasible region, thus defining the physically permissible range of various state variables in the system, such as flow rate and liquid level. The system determines these boundaries based on the static physical properties of the pipeline network. For the edge-bearing attribute, i.e., the pipeline flow rate, the upper limit of its feasible region is the full-pipe flow rate calculated using the Manning formula or Chezy formula combined with parameters such as pipe diameter and slope. The lower limit is typically zero. For node accumulation attributes, i.e., the water level at inspection wells or confluence points, the upper limit of the feasible region is the wellhead elevation or overflow prevention elevation of that node. The lower limit is the bottom elevation of the pipe. The system calculates a set of upper and lower boundaries for each flow rate and level variable involved in the topological spatiotemporal representation tensor. The set of these boundaries constitutes the feasible region boundary of the entire system.

[0096] By comparing the deviation of the conservation deviation spectrum with the feasible domain boundary in the node and edge dimensions, a physical constraint deviation is generated.

[0097] Specifically, such as Figure 2 As shown, the system generates the final physical constraint deviation by comparing the deviation of the conservation deviation spectrum with the feasible region boundary, unifying two different types of physical constraints, such as equality constraints and inequality constraints, into a single, quantifiable deviation signal. The system first checks whether the flow values ​​of all edges and the accumulation values ​​of all nodes in the topological spatiotemporal representation tensor exceed the feasible region boundary defined in the previous step. For any out-of-bounds value, the system calculates the absolute amount of its deviation as its deviation. Then, the system integrates these deviations with the mass conservation deviation values ​​in the conservation deviation spectrum. This integration process may involve normalizing different types of deviations to eliminate dimensional effects and assigning different weights; for example, the weight of mass conservation is usually set relatively high, such as 0.6 to 0.8. The final physical constraint deviation is a set of values ​​aligned with the topological spatiotemporal representation tensor in the node and edge dimensions. It comprehensively quantifies the inconsistency between the existing data representation and physical reality, serving as a key input to the neural network model to guide the model in generating predictions that are more consistent with physical laws.

[0098] Optionally, the generation of the spatiotemporal prediction field includes:

[0099] Construct a graph representation of the topological spatiotemporal representation tensor and a constraint representation of the physical constraint deviation, and perform normalization processing to generate normalized input pairs;

[0100] Specifically, a neural network model is used to fuse feature data containing spatiotemporal dynamics with quantified physical rule deviations to generate accurate predictions of the future state of the drainage system. First, the system performs normalization to generate normalized input pairs. This operation aims to adjust input data with different dimensions and numerical ranges to a similar scale, improving the training efficiency and stability of the neural network model. The system constructs a graph representation of the topological spatiotemporal representation tensor generated in the previous steps and scales each feature dimension, such as flow rate and liquid level, to a uniform range, typically a distribution of 0 to 1 or a mean of 0 and a variance of 1, using max-min normalization or Z-score normalization. Simultaneously, the system constructs a constraint representation of the physical constraint deviations and normalizes it using the same method. This pair of normalized graph representations and constraint representations constitutes the normalized input pair, which serves as the unified input to the neural network fusion module.

[0101] The normalized input pairs are fused and then time-series decoded to generate a time-series decoded sequence;

[0102] Specifically, the system feeds normalized input pairs into the neural network fusion module and performs temporal decoding to generate a temporal decoded sequence. Deep learning models are used to capture complex spatiotemporal dependencies and deduce future evolution trends from historical data and physical constraints. The neural network fusion module typically employs a graph neural network architecture, such as GraphConvolutionalNetwork or GraphAttentionNetwork, which can directly process graph-structured data. First, information from neighboring nodes is aggregated at each time step through graph convolutional layers, thereby learning spatial interactions. Subsequently, a recurrent neural network unit, such as LSTM or GRU, is applied to the time series of each node to capture temporal evolution patterns. During this process, physical constraint biases are also fed into the model, either as an additional dimension of node features or through a dedicated loss function to influence the model's gradient updates. After model learning is complete, the prediction phase begins. The system feeds normalized input pairs composed of the latest historical data into the trained model. The model encoder extracts spatiotemporal features, and the decoder, typically also a recurrent neural network structure, autoregressively generates future states step by step. The model iteratively predicts a certain step size in the time dimension, such as the next 1 to 3 hours. The resulting series of high-dimensional hidden state vectors constitute the temporal decoding sequence.

[0103] The future state estimates of the temporal decoding sequence in the node and edge dimensions are decoded to generate a spatiotemporal prediction field.

[0104] Specifically, the system decodes the future state estimates of the time-series decoding sequence at the node and edge dimensions to generate the final spatiotemporal prediction field. This operation aims to transform the abstract representation within the model back into prediction results with clear physical meaning. The system maps each hidden state vector in the time-series decoding sequence back to the space of the original physical quantities through a fully connected layer or a linear transformation layer. For example, one part of the vector is decoded as the predicted liquid level of the node at a future time, and another part is decoded as the predicted flow rate of the outflow edge. This decoding process is the inverse of the normalization process, that is, restoring the numerical values ​​in the range of 0 to 1 of the model output to predicted values ​​with actual physical units and dimensions according to the previously recorded scaling parameters. The system performs this decoding operation on all nodes and edges at all prediction time steps, ultimately generating a multidimensional data field containing the predicted values ​​of all key state variables such as liquid level and flow rate of the entire pipeline network over a future period, i.e., a spatiotemporal prediction field, which provides direct data support for subsequent early warning and control.

[0105] Optionally, the step of performing uncertainty quantification and error feedback on the spatiotemporal prediction field to generate correction control parameters, and using the correction control parameters to perform a secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field includes:

[0106] Calculate the temporal deviation between the spatiotemporal prediction field and the event alignment sequence, and perform error source decomposition to generate an error source vector;

[0107] Specifically, a closed-loop feedback mechanism is constructed. By quantifying and analyzing the deviation between the initial prediction results and the actual observations, and tracing the source of errors, a correction signal is dynamically generated. This allows for secondary optimization of the original predictions of the neural network, improving the long-term accuracy and adaptability of the predictions. First, the system performs error source decomposition to generate an error source vector. This operation is triggered after each time step of the model's continuous operation. The system compares the predicted value corresponding to the current moment in the spatiotemporal prediction field generated at the previous moment with the actual observation value at this moment in the event alignment sequence, calculating the prediction error of each node and edge. The system further performs statistical analysis on a sliding time window, such as the historical error sequence within the past 6 to 12 prediction periods, decomposing it into systematic biases representing persistently high or low biases, and random biases reflecting short-term fluctuations. These two components together constitute the error source vector.

[0108] The uncertainty in the error source vector is quantized and mapped, and then stability constraints are applied to generate correction control parameters.

[0109] Specifically, the system performs uncertainty quantification and stability constraints on the error source vector to generate correction control parameters. The quantized error is transformed into a smooth and effective adjustment signal. The system uses the variance of the random deviation in the error source vector as a direct measure of the uncertainty of the prediction result for that node or edge. Then, the system uses systematic deviation as the main correction driving force and applies stability constraints to avoid introducing new oscillations during the correction process. This constraint is implemented through a filter, the update process of which can be expressed by the following equation:

[0110] ,

[0111] Where Ki is the correction control parameter calculated for node or edge i in the current cycle, Kprevi is the correction control parameter in the previous cycle, and Bi is the latest calculated systematic deviation for that node or edge. A and B are filter coefficients, and the sum of A and B is 1. B serves as a gain factor, and its value is usually between 0.1 and 0.4. It is used to control the system's response speed to new deviations. A smaller value of B will produce a smoother correction effect.

[0112] The correction control parameters are applied to the estimation of nodes and edges in the spatiotemporal prediction field to generate an adaptive correction field.

[0113] Specifically, such as Figure 3 As shown, the system applies correction control parameters to the estimation of nodes and edges in the spatiotemporal prediction field, generating an adaptive correction field through quadratic mapping. This operation aims to apply the calculated adjustment to the model's original future predictions. The system combines each predicted value in the newly generated spatiotemporal prediction field from the neural network fusion module with the corresponding node or edge correction control parameters. This quadratic mapping process is typically a direct additive correction, as shown in the following equation:

[0114] ,

[0115] Here, Yc is the final predicted value in the adaptive correction field, Yo is the original predicted value in the spatiotemporal prediction field, and Ki is the stable correction control parameter generated in the previous step. By performing this operation on all prediction points in the spatiotemporal prediction field, the system generates the final adaptive correction field. This correction field not only includes the neural network's prediction of the future but also incorporates self-correction based on historical performance, enabling the entire prediction system to continuously adapt to the drift and changes in the dynamic characteristics of the pipeline network.

[0116] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides a spatiotemporal prediction system for the state of a drainage system based on a neural network, the system comprising:

[0117] The data collection and annotation module is used to acquire pipeline network geographical structure data and operational observation data, and to perform time annotation on the data to generate a time axis annotation set;

[0118] The event alignment module is used to perform event-driven spatiotemporal alignment and missing completion on the timeline annotation set, and generate an event alignment sequence.

[0119] The topology construction module is used to construct a reachable topology graph with directionality and connectivity based on the event alignment sequence and by integrating pipeline connection relationships and flow direction rules.

[0120] The joint representation module is used to perform joint feature extraction on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor.

[0121] The physical constraint module is used to apply constraints based on mass conservation and capacity limitation to the topological spatiotemporal representation tensor and to calculate and generate physical constraint deviation.

[0122] The fusion prediction module is used to fuse the topological spatiotemporal representation tensor with the physical constraint deviation to generate a spatiotemporal prediction field.

[0123] An adaptive correction module is used to perform uncertainty quantification and error feedback on the spatiotemporal prediction field, generate correction control parameters, and use the correction control parameters to perform secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field.

[0124] To verify the feasibility of this invention in practice, it was applied to a drainage system for spatiotemporal prediction of its status in order to address urban flooding. This drainage system has a complex network covering approximately 50 square kilometers and includes thousands of monitoring nodes. Its operational status is affected by multiple factors, including rainfall, tides, and upstream water flow.

[0125] Before applying this invention, the main approach relied on local alarms based on fixed thresholds and manual experience-based scheduling, which resulted in delayed responses and made it difficult to perform forward-looking and global optimization scheduling in the early stages of rainfall events.

[0126] This embodiment uses historical data for model training and validation. The pipeline network geographic structure data comes from a GIS system, containing approximately 5200 manhole nodes and 6100 pipelines, specifying the X / Y / Z coordinates of the nodes, the start and end points of the pipelines, pipe diameter, material, and other information. Operational observation data comes from a SCADA system, including 850 level sensors and 320 flow sensors, with sampling intervals ranging from 5 to 15 minutes.

[0127] First, the system performs data aggregation and annotation. The acquired geographic and observational data are uniformly converted to Parquet format. The operational data is cleaned, removing abnormal records where the liquid level exceeds the upper or lower limit of the pipe diameter by 5%, and duplicate records are removed based on timestamps and sensor IDs. Then, by calculating the signal variance within a 15-minute sliding window, signals with a variance below 0.0001 are identified as invalid silent signals and filtered out, generating a valid observation set. Finally, the UTC timestamps and sampling intervals of all valid records are parsed to generate a timeline annotation set.

[0128] Next, the system performs event-driven spatiotemporal alignment and missing data completion on the timeline annotation set. A rainfall intensity exceeding 15 mm / hour is defined as the trigger time for a "rainfall event." The associated time window is defined as 1 hour before the event trigger to 4 hours after the event. Based on GIS data, sensors within a 500-meter physical distance are divided into spatial neighborhoods, generating neighborhood grouping sets. Within each group, using the sensor with the most stable sampling as the benchmark, the sampling delay and clock drift of other sensors (up to 3 minutes) are corrected by calculating the cross-correlation function. For intermittent missing segments of less than 4 sampling points (approximately 1 hour), cubic spline interpolation is used; for longer missing segments, spatiotemporal kriging is used to interpolate using synchronous data from other sensors within the neighborhood, generating a complete alignment set. Finally, event boundaries and adjacency relationships are fused to generate the final event alignment sequence.

[0129] Building upon this foundation, the system begins constructing a reachable topology map with directionality and connectivity. First, connectivity component analysis is performed based on pipeline connection information, dividing the pipeline network into three main connected subsystems. Then, reachability is derived based on the elevation difference between the pipe bottoms of nodes, setting the elevation difference threshold Ht to 0.08 meters to initially determine the direction of gravity flow. A crucial step is that the system uses flow rate change data from event alignment sequences to verify and correct 112 pipelines with insignificant elevation differences and 28 known pressure-upstream pipelines. Specifically, the data-driven flow direction is determined by calculating the time-delay cross-correlation of flow rate change rates at both ends of the pipeline. When the data flow direction conflicts with the elevation-inferred flow direction and the cross-correlation coefficient is greater than 0.75, the system adopts the data-driven result and corrects the reachability relation set. Finally, a reachable topology map accurately reflecting the actual hydraulic path is generated.

[0130] Subsequently, the system performs joint feature extraction and physical constraint calculation. The system integrates dynamic and static features on the reachable topology graph and slices it using time windows of 12 time steps (60 minutes of history) and 3 time steps (15 minutes), generating a topological spatiotemporal representation tensor. For each node, at each time step, the system calculates its deviation Di according to the law of mass conservation. Simultaneously, the system uses the full-pipe flow rate calculated based on the pipe diameter using the Manning formula as the edge capacity constraint and the ground elevation as the node's accumulation capacity constraint. After normalizing the above conservation deviation and capacity deviation, the physical constraint deviation is generated.

[0131] Finally, the system inputs the topological spatiotemporal representation tensor and the physical constraint deviation into a neural network fusion module based on GCN and GRU to generate a spatiotemporal prediction field. This field contains the predicted liquid level and pipeline flow for all nodes in the entire network over the next 12 time steps, or 60 minutes. Then, the adaptive correction module is activated. The system continuously calculates the error between the predicted and actual values ​​and decomposes it into systematic and random deviations. A stable correction control parameter is generated using an exponential moving average filter with a smoothing coefficient B of 0.25, and this parameter is superimposed onto the original prediction field for the next round to generate the final output adaptive correction field.

[0132] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A spatiotemporal prediction method for the state of a drainage system based on neural networks, characterized in that, The method includes: Acquire pipeline network geographical structure data and operational observation data, and generate a time axis annotation set by time-marking the data; Event-driven spatiotemporal alignment and missing data completion are performed on the timeline annotation set to generate an event alignment sequence; Based on the event alignment sequence, a reachable topology graph with directionality and connectivity is constructed by fusing pipeline connection relationships and flow direction rules. This includes: obtaining the node set and pipeline connection information of the event alignment sequence, and performing connectivity component analysis to generate a connectivity graph; calculating the elevation difference and edge direction indication in the connectivity graph, and performing reachability derivation to generate a reachable relation set; combining the reachable relation set and pipeline connection information to set directionality and connectivity, and generating a reachable topology graph. Joint feature extraction is performed on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor; The physical constraint deviation is calculated by applying constraints based on mass conservation and capacity limitation to the topological spatiotemporal representation tensor. This includes: extracting the node inflow and outflow attributes of the topological spatiotemporal representation tensor and performing node mass conservation calculations to generate a conservation deviation spectrum; extracting the edge carrying and node accumulation attributes of the topological spatiotemporal representation tensor and performing feasible region delineation to generate a feasible region boundary; and comparing the deviation of the conservation deviation spectrum from the feasible region boundary in the node and edge dimensions to generate the physical constraint deviation. The spatiotemporal representation tensor and the physical constraint deviation are fused to generate a spatiotemporal prediction field. This includes: constructing a graph representation of the topological spatiotemporal representation tensor and a constraint representation of the physical constraint deviation, and performing normalization processing to generate normalized input pairs; fusing the normalized input pairs and performing temporal decoding to generate a temporal decoding sequence; and decoding the future state estimates of the temporal decoding sequence in the node and edge dimensions to generate the spatiotemporal prediction field. Uncertainty quantification and error feedback are performed on the spatiotemporal prediction field to generate correction control parameters, and the spatiotemporal prediction field is then mapped twice using the correction control parameters to generate an adaptive correction field. The method further includes: extracting pipeline connection information and performing noise edge filtering to generate an edge purification set; fusing the edge attributes of the edge purification set with the flow change rate of the event alignment sequence, and performing edge direction normalization to generate a direction normalization matrix; verifying the consistency between the direction normalization matrix and the reachability set, and correcting the reachability set based on the verification result.

2. The spatiotemporal prediction method for the state of a drainage system based on a neural network according to claim 1, characterized in that, The generated timeline annotation set includes: Acquire pipeline network geographical structure data and operational observation data, and perform data format transformation to generate a unified format set; Clean the abnormal and duplicate records in the unified format set, and perform signal validity screening to generate a valid observation set; The timestamps and sampling intervals of the valid observation set are analyzed to generate a time axis label set.

3. The spatiotemporal prediction method for the state of a drainage system based on a neural network according to claim 2, characterized in that, The generated event alignment sequence includes: Identify the event trigger times and associated time windows in the timeline annotation set, and generate a neighborhood group set by spatial neighborhood grouping; The sampling delay and clock drift of the neighborhood group set are corrected, and interpolation is performed on the intermittent missing segments to generate a complete and aligned set; The event alignment sequence is generated by fusing the event boundaries and spatial adjacency relationships of the completed alignment set.

4. The spatiotemporal prediction method for the state of a drainage system based on a neural network according to claim 1, characterized in that, The generated topological spatiotemporal representation tensor includes: Extract time series data from the event alignment sequence to generate an observation attribute matrix, and extract geometric and connectivity metrics from the reachable topology graph to generate a topology attribute matrix; Map the observation attribute matrix and the topology attribute matrix to the nodes and edges of the reachable topology graph, and perform time window slicing to generate a time slice set; The temporal dependence and spatial neighborhood relationship of the time slice set are fused to generate a topological spatiotemporal representation tensor.

5. The spatiotemporal prediction method for the state of a drainage system based on a neural network according to claim 4, characterized in that, The step of performing uncertainty quantification and error feedback on the spatiotemporal prediction field to generate correction control parameters, and using the correction control parameters to perform a secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field includes: Calculate the temporal deviation between the spatiotemporal prediction field and the event alignment sequence, and perform error source decomposition to generate an error source vector; The uncertainty in the error source vector is quantized and mapped, and then stability constraints are applied to generate correction control parameters. The correction control parameters are applied to the estimation of nodes and edges in the spatiotemporal prediction field to generate an adaptive correction field.

6. A spatiotemporal prediction system for the state of a drainage system based on neural networks, characterized in that, The system includes: The data collection and annotation module is used to acquire pipeline network geographical structure data and operational observation data, and to perform time annotation on the data to generate a time axis annotation set; The event alignment module is used to perform event-driven spatiotemporal alignment and missing completion on the timeline annotation set, and generate an event alignment sequence. A topology construction module is used to construct a reachable topology graph with directionality and connectivity based on the event alignment sequence and by integrating pipeline connection relationships and flow direction rules. This includes: acquiring the node set and pipeline connection information of the event alignment sequence, and performing connectivity component analysis to generate a connectivity graph; calculating the elevation difference and edge direction indicators in the connectivity graph, and performing reachability derivation to generate a reachable relation set; combining the reachable relation set and pipeline connection information to set directionality and connectivity, and generating a reachable topology graph. The joint representation module is used to perform joint feature extraction on the event alignment sequence and the reachable topology graph to generate a topological spatiotemporal representation tensor. The physical constraint module is used to apply constraints based on mass conservation and capacity limitation to the topological spatiotemporal representation tensor and calculate and generate physical constraint deviations. This includes: extracting the node inflow and outflow attributes of the topological spatiotemporal representation tensor and performing node mass conservation calculations to generate a conservation deviation spectrum; extracting the edge carrying and node accumulation attributes of the topological spatiotemporal representation tensor and performing feasible region delineation to generate feasible region boundaries; and comparing the deviations of the conservation deviation spectrum and the feasible region boundaries in the node and edge dimensions to generate physical constraint deviations. The fusion prediction module is used to fuse the topological spatiotemporal representation tensor with the physical constraint deviation to generate a spatiotemporal prediction field. This includes: constructing a graph representation of the topological spatiotemporal representation tensor and a constraint representation of the physical constraint deviation, and performing normalization processing to generate normalized input pairs; fusing the normalized input pairs and performing temporal decoding to generate a temporal decoding sequence; and decoding the temporal decoding sequence to estimate the future states of the nodes and edges, thereby generating the spatiotemporal prediction field. An adaptive correction module is used to perform uncertainty quantification and error feedback on the spatiotemporal prediction field, generate correction control parameters, and use the correction control parameters to perform a secondary mapping on the spatiotemporal prediction field to generate an adaptive correction field. The topology construction module further includes: extracting pipeline connection information and performing noise edge filtering to generate an edge purification set; fusing the edge attributes of the edge purification set with the flow change rate of the event alignment sequence and performing edge direction normalization to generate a direction normalization matrix; verifying the consistency between the direction normalization matrix and the reachability set, and correcting the reachability set based on the verification result.

Images