Urban inland inundation water recession process numerical simulation method and system based on physical-data fusion

Abstract

The invention discloses an urban inland inundation water recession process numerical simulation method and system based on physical-data fusion, and belongs to the technical field of urban flood control and intelligent water affairs. A hydraulic model and a data driving model are fused, a prediction result is dynamically weighted through gating attention, and real-time correction is performed by using Kalman filtering; the method comprises the steps of multi-source data fusion and enhancement, wherein data collection and processing are carried out, space-time alignment is achieved, and feature construction is carried out; mechanism and mathematical double-engine collaborative prediction: collaborative prediction is carried out based on a physical model engine and a data driving engine, and adaptive fusion is realized; online reasoning and dynamic correction, including Kalman filtering dynamic correction and extreme scene emergency processing; and performing model evaluation and iteration. According to the method, the problem of high-precision prediction of the recession time and the ponding range in a complex urban environment is solved, the generalization ability of the model in non-seen scenes such as pipeline blockage and rainstorm extrapolation is improved, and a reliable decision basis is provided for urban inland inundation emergency scheduling and drainage facility optimization.

Description

Technical Field

[0001] The present invention relates to the technical field of urban flood control and intelligent water services, and specifically to a numerical simulation method and system for the urban waterlogging recession process based on physical-data fusion. Background Art

[0002] With the acceleration of global climate change and urbanization, extreme rainfall events occur frequently, leading to increasingly serious urban waterlogging disasters. Accurately predicting the recession process is the core prerequisite for formulating emergency drainage plans. The current mainstream methods mainly rely on two technical routes: physical models based on hydraulic equations and data-driven models, but both have significant limitations.

[0003] In terms of physical models, drainage network dynamic models represented by SWMM (Storm Water Management Model) simulate the coupling process of surface runoff and pipe flow by solving the Saint-Venant equations. Although such models have clear physical meanings, their prediction accuracy highly depends on the accurate input of pipe network geometric parameters (such as pipe diameter, slope). In actual engineering, problems such as pipe aging and deformation, sediment blockage, and illegal connection are common, resulting in systematic deviations between the theoretical model parameters and the real system. The literature (Zhang et al., Water Research 2021) shows that when the effective cross-sectional area of the pipe decreases by 20%, the error in the recession time predicted by SWMM can reach 41%. In addition, traditional models are difficult to express the dynamic interaction mechanism between surface depressions and underground pipe networks. For example, the flow capacity of the curb opening on the road will significantly affect the local ponding recession rate, and these microstructures are often not included in the urban GIS database.

[0004] On the other hand, pure data-driven methods represented by deep learning predict by mining statistical laws in historical inundation data. Although such models can capture complex nonlinear relationships, they produce prediction results that violate the law of mass conservation due to the lack of physical constraints. For example, in the review experiment of the extremely heavy rainstorm in 2023, the recession time of a certain area predicted by an LSTM model was 2.7 hours faster than the actual time because it did not consider the backwater effect caused by the full flow of the upstream pipe. More seriously, when the rainfall intensity exceeds the historical record (such as a once-in-a-century rainstorm), the prediction error of the data-driven model increases exponentially (the extrapolation error reported in the literature Liu et al., Nature Hazards 2023 exceeds 200%), which poses a major risk to disaster prevention decision-making.

[0005] In recent years, although there have been research attempts to integrate physical models with machine learning, such as splicing pipe network data and rainfall sequences and inputting them into a fully connected network, it only achieves the superposition of shallow features and fails to model the spatio-temporal constraints of the pipe network topology on water flow propagation. In an actual drainage system, the connection relationships between nodes will form a cascading effect - congestion in downstream pipes may trigger chain overload of multiple upstream nodes, and this non-local interaction cannot be effectively expressed by traditional convolutional or recurrent neural networks. In addition, existing methods generally ignore the dynamic error distribution characteristics between the output of physical models and measured data. For example, the accuracy of physical models is relatively high when the drainage capacity of the pipe network is sufficient at the initial stage of a rainstorm, while the error increases sharply after the pipes are full during continuous heavy rainfall, which requires the fusion model to have the ability to adaptively adjust weights. Summary of the Invention

[0006] The technical task of the present invention is to address the above deficiencies and provide a numerical simulation method and system for the urban waterlogging recession process based on physical-data fusion, which can solve the problem of high-precision prediction of the recession time and waterlogging range in complex urban environments, and at the same time improve the generalization ability of the model in unseen scenarios such as pipe blockages and rainstorm extrapolation, providing a reliable decision-making basis for urban waterlogging emergency dispatch and drainage facility optimization.

[0007] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0008] A numerical simulation method for the urban waterlogging recession process based on physical-data fusion, which integrates a hydraulic model and a data-driven model, dynamically weights the prediction results through gated attention, and uses a Kalman filter for real-time correction;

[0009] The implementation of this method includes the following steps:

[0010] 1) Multi-source data fusion and enhancement: Collect and process data, achieve spatio-temporal alignment, and construct features;

[0011] 2) Mechanism + mathematical double-engine collaborative prediction: Collaboratively predict based on a physical model engine and a data-driven engine, and achieve adaptive fusion;

[0012] 3) Online inference and dynamic correction: Include Kalman filter dynamic correction and extreme scenario emergency handling;

[0013] 4) Model evaluation and iteration.

[0014] Furthermore, the data processing transforms the pipe network topology into an attribute graph structure and extracts features such as depression depth, impermeability rate, and drainage hierarchy;

[0015] The multi-source data fusion and enhancement specifically include:

[0016] Data integration: Spatially and temporally align the scattered pipe network topology data (pipe connection relationships, pipe diameters and slopes), dynamic rainfall data (radar gridded rainfall), and historical inundation data (water level monitoring points, camera images), and uniformly map them to a 500-meter gridded geographic coordinate system with a time resolution accurate to 10 minutes; construct a bidirectional topology map for the pipe network system, where nodes represent inspection wells / pumping stations and edges represent pipe attributes (including flow capacity, blockage risk labels).

[0017] Abnormal data processing: For sensor noise, use a spatio-temporal sliding window filter (3 previous and 3 subsequent time steps + 3 adjacent grids) to remove outliers; for missing pipe network parameters (such as unknown pipe diameters), infer approximate values based on a graph neural network and verify the reliability in combination with historical drainage records.

[0018] For the above-mentioned spatio-temporal alignment, spatial alignment: Use QGIS to convert the pipe network node coordinates to the WGS84 coordinate system, overlay them with the radar rainfall grid, perform Kriging interpolation (selecting the Gaussian model for the semi-variogram) on the discrete water level monitoring points to generate a 500m×500m gridded historical inundation field; temporal alignment: Take rainstorm events as units, intercept the complete temporal data from 1 hour before the start of rainfall to after the recession of floodwater, and for non-uniformly sampled data (such as 15-minute rainfall + 5-minute water level), use cubic spline interpolation to unify it to a 10-minute interval; Key feature construction includes: Pipe network topology features: Calculate the "drainage level" (the shortest number of pipes from the outlet) and "catchment capacity" (the total cross-sectional area of downstream pipes) for each node; Surface features: Extract the "depression depth" (the height difference between the local terrain lowest point and the surroundings) and "impervious area ratio" (based on satellite image segmentation) for each grid.

[0019] Furthermore, for the above-mentioned data integration, use Delaunay-MLS interpolation to map the discrete water level data into a continuous inundation field, and the interpolation weight is based on an exponential decay function.

[0020] Furthermore, for the collaborative prediction of the mechanism + mathematical physics dual engine

[0021] Physical model engine: Based on the SWMM hydraulics kernel, simulate the interaction process between pipe network water flow and surface runoff, and output the theoretical recession curve; introduce a dynamic parameter correction module: Automatically adjust parameters, including Manning coefficient, infiltration rate, etc., according to the real-time rainfall intensity and pipe load status to alleviate the model deviation caused by pipe blockage.

[0022] Data-driven engine: Includes surface inundation field prediction and pipe network status prediction

[0023] Surface inundation field prediction: Use a ConvLSTM network to learn the spatio-temporal propagation law of the historical waterlogging depth field, and capture local factors not considered by traditional models, including road enclosures and temporary drainage facilities.

[0024] Pipe network status prediction: Analyze the topological relationship of the pipe network through a graph attention network (GNN) to predict the flow change trend of key nodes (such as pumping stations and easily blocked points);

[0025] Adaptive fusion decision-making: Through an intelligent weight allocation module, dynamically adjust the contribution ratio of the physical model and the data model according to the real-time scenario: When the rainfall intensity is low (such as <30mm / h), the physical model is mainly used (weight 70%), relying on its stable mechanism; When in a high-load state (such as pipe full flow, heavy rainfall), gradually increase the weight of the data model to 80%, and utilize its ability to capture complex non-linear relationships; Automatically generate a spatially differentiated fusion weight map through a gated neural network. For example, assign a higher physical weight to areas with a dense underground pipe network, while rely on data-driven prediction for low-lying areas lacking pipe network data;

[0026] The gated attention weight is jointly calculated based on the physical prediction value, data gradient, node flow, and rainfall intensity.

[0027] Furthermore, the Manning coefficient of the hydraulic model is dynamically adjusted according to the pipe sedimentation degree, and the sedimentation degree is calculated based on the cumulative flow deviation;

[0028] The physical model optimization process is as follows:

[0029] SWMM parameter dynamic calibration: Add a "sedimentation degree - Manning coefficient" feedback loop to the traditional SWMM model:

[0030] n(t) = n_base × (1 + 0.2 × Clog_level);

[0031] Where, Clog_level = cumulative flow prediction deviation / design flow;

[0032] Automatically calibrate the infiltration parameters every 6 hours, and use the SCE-UA optimization algorithm to minimize the node flow error;

[0033] Boundary condition enhancement: Input the over-flow capacity of the roadside curb as a dynamic boundary condition, and the calculation formula is:

[0034] Q_road = C × L × h^1.5;

[0035] Where, C: curb opening coefficient, L: curb length, h: road surface water depth;

[0036] Data-driven model training includes the following:

[0037] ConvLSTM surface flooding prediction:

[0038] Input: Past 3-hour rainfall field sequence (6×6 grid) + current terrain features (depression depth, slope);

[0039] Training strategy: Adopt Curriculum Learning, first train light rain events, and gradually increase the complexity of heavy rain samples;

[0040] GNN pipeline state prediction:

[0041] Graph construction: Define pipeline nodes as graph nodes, pipes as directed edges, and edge attributes include pipe diameter, slope, and length;

[0042] Message passing: Adopt the GATv2 network to aggregate the flow rate and water depth information of neighbor nodes;

[0043] Output: The probability distribution of the flow rate and fullness degree of each node in the next 1 hour;

[0044] The implementation of the adaptive fusion module includes:

[0045] Gating weight generation network:

[0046] Input: Node flow predicted by the physical model + inundation depth predicted by the data model + real-time rainfall intensity;

[0047] Structure: A 3-layer fully connected network (128→64→1 neuron), and the output is the weight value α ∈ [0,1] activated by Sigmoid;

[0048] Dynamic adjustment rule:

[0049] Rainfall intensity > 50mm / h, the weight α is adjusted to: α ← α × 0.7 + 0.15;

[0050] Pipe fullness > 80% for 30 minutes continuously, the weight α is adjusted to: α ← max(0.3, α - 0.2);

[0051] Sensor data is abnormal 5 times continuously, the weight α is adjusted to: α ← 0.1 (forcing data dominance).

[0052] Furthermore, generate extreme rainfall scenario data through physically constrained GAN, and the generator loss function includes the residual term of the mass conservation equation.

[0053] Furthermore, the online inference and dynamic correction achieve closed-loop feedback optimization, including:

[0054] Real-time data assimilation: Access real-time data streams, including real-time data streams such as manhole cover water level sensors and road cameras, dynamically correct the model output through Kalman filtering, and update the prediction results every fixed time period (such as 5 minutes); when it is detected that the actual water recession speed continuously deviates from the predicted value (such as the error > 15%), automatically trigger the recalibration of model parameters;

[0055] Among them, the Kalman filter dynamic correction process includes:

[0056] State vector: including the flood depth of each grid, node flow, and fusion weight α;

[0057] Observation equation: The water level sensor directly corresponds to the flood depth of the grid, and the pump station flowmeter corresponds to the node flow;

[0058] Update process: In the prediction stage, the state of the next time step is inferred based on the fusion model; in the update stage, when new sensor data arrives, the Kalman gain K is calculated to correct the state estimate; a complete prediction-update cycle is executed every fixed time interval (such as 15 minutes);

[0059] Emergency handling for extreme scenarios includes:

[0060] Full-flow warning for pipelines: If the flow of a certain node exceeds 90% of the design capacity, immediately retrieve the backup model (trained based on historical similar events) for local prediction replacement.

[0061] Extrapolated rainfall compensation: When the rainfall intensity exceeds the historical maximum record, start the physical constraint enhancement mode to force the model to output the upper limit of the recession rate that conforms to mass conservation; The response to sudden heavy rainfall includes: When the radar predicts that the rainfall in the next 1 hour exceeds the historical extreme value, start the "physical constraint enhancement mode" to force the data model to output to satisfy mass conservation Limit the recession rate not to exceed 120% of the maximum drainage capacity of the pipe network; Fault tolerance for equipment failures: If a large area of sensors in a certain area fails, switch to the "pure physical prediction + adjacent area data compensation" mode.

[0062] Furthermore, the model evaluation and iteration include:

[0063] Offline verification: To comprehensively evaluate the model performance, a spatio-temporal cross-validation method is adopted, and the training set and test set are divided with rainstorm events as independent units to avoid overfitting caused by time correlation; The specific process includes: constructing 10 groups of rainstorm scenarios in different years (such as once in 1 year to once in 50 years) based on historical data, and calculating the recession time error (the absolute difference between the predicted and actual recession times), the overlap degree of the waterlogging area (IoU), and the capture rate of extreme events (hit rate of the top 10 prone-to-waterlogging points) respectively. By comparing the prediction results of the traditional SWMM model, the pure data model, and this method, verify the robustness of the fusion model in common and extreme scenarios; For example, in the "Dragon Boat Water" event in Guangzhou in 2022, the average error of the model for the recession time was reduced from 2.1 hours of the traditional model to 0.7 hours, and the IoU of the waterlogging area increased by 22%. The verification results will be used to optimize the parameters of the gating network and the physical constraint weight coefficients;

[0064] Online Monitoring: The system tracks the running status of the model in real time through multi-dimensional indicators, including: constructing a dynamic monitoring dashboard based on the GIS platform, visually displaying the fusion weight heat map (spatial distribution of α value) of the physical model and the data-driven model, and triggering a yellow warning when the weight value fluctuates by more than ±0.3 within 1 hour; comparing the predicted inundation depth with the measured values of sensors grid by grid to generate an error heat map layer (graded in red / yellow / green), and if the error exceeds 15% for 3 consecutive time steps, it will locate to specific pipe network nodes for problem diagnosis (such as abnormal pipe blockage coefficient); synchronously recording the time series curve of the Kalman filter correction amount, and when the correction amplitude in a certain area continues to expand (such as the daily adjustment amount increases by 20%), automatically generating a pipe network health assessment report to prompt the key areas for inspection (such as the cumulative correction amount of node NJ-1024 has reached 38 cm in the past week); the monitoring data is stored in a time series database, supporting the correlation analysis of the model performance and external events at any time period in the past (such as the prediction deviation caused by the pipe diameter change due to construction).

[0065] Continuous Learning: To cope with the dynamic changes of the urban drainage system (such as pipe aging, newly built storage facilities), the following three-level adaptive update mechanism is designed:

[0066] Incremental Fine-tuning: Extract new data (rainfall-inundation records) every day at dawn, use the historical model as the teacher network, and fine-tune the parameters of the data-driven branch with a cosine annealing learning rate (initial value 1e-5), and constrain the physical branch to be immutable to maintain the mechanism stability;

[0067] Extreme Scenario Enhancement: When a rainfall event exceeding historical records is detected, start the physical constraint GAN to generate adversarial samples. The generator synthesizes an inundation field with a rainfall intensity of 120 - 150 mm / h, and the discriminator enforces the mass conservation equation: Expand the diversity of training data;

[0068] Version Iteration: Conduct A / B tests at the end of each month, compare the core indicators such as the recession time error (MAE) and the capture rate of extrapolation scenarios between the new and old models. Only when the comprehensive score of the new model increases by ≥5% and the false alarm rate in the key area decreases, gradually replace the online version through the gray release strategy to ensure business continuity.

[0069] The present invention also claims to protect a numerical simulation system for the urban waterlogging recession process based on physical-data fusion, including a data preprocessing module, a physical model engine, a data model engine, a dynamic fusion module, and a real-time correction module;

[0070] Among them, the data preprocessing module uses a graph neural network to complete the missing pipe network parameters and imposes a flow continuity constraint;

[0071] This system implements the above method.

[0072] The present invention also claims a computer-readable medium, on which computer instructions are stored, and when the computer instructions are executed by a processor, the above-mentioned method is implemented.

[0073] Compared with the prior art, the numerical simulation method and system for urban waterlogging recession process based on physical-data fusion of the present invention have the following beneficial effects:

[0074] 1. Significantly improved prediction accuracy: The average error of recession time prediction is reduced from 2.1 hours to 0.7 hours, and the error of peak rainstorm water level is reduced by 58%.

[0075] 2. Strong physical interpretability: The decision-making basis of the model is shown through a weight heat map, which supports the reverse tracing of problems such as pipeline network blockage.

[0076] 3. Optimized calculation efficiency: The prediction time for a 10km 2 area is less than 60 seconds for 1 hour, and it supports real-time calculation of 50 concurrent areas.

[0077] 4. Reliable extrapolation for extreme scenarios: The error growth under rainfall exceeding the design value is controlled within 35%, and the data generated by the physically constrained GAN improves the extrapolation ability by 70%.

[0078] 5. Dynamic adaptive adjustment: The fusion weight changes automatically with the rainfall intensity, and the Kalman filter is used for real-time correction to maintain a reliability of more than 85%.

[0079] 6. Continuous evolution ability: Incremental learning reduces the monthly error by 2 - 3%, and the physically constrained GAN enhances the accuracy of unexperienced scenarios by 25%.

[0080] 7. High fault tolerance design: When 50% of the sensors fail, the accuracy is still maintained at 75%, and a lightweight backup model is automatically switched in case of anomalies.

[0081] 8. Multi-source data fusion standard: The Delaunay-MLS interpolation algorithm reduces the discrete data mapping error by 27%, and a pipeline network topology quantization standard is established. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 It is a schematic diagram of a prediction framework for the risk probability of water supply pipeline network burst provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0083] The embodiment of the present invention provides a numerical simulation method for urban waterlogging recession process based on physical-data fusion, which fuses a hydraulic model and a data-driven model, dynamically weights the prediction results through gated attention, and uses the Kalman filter for real-time correction.

[0084] Among them, Delaunay-MLS interpolation is used to map discrete water level data into a continuous inundation field, and the interpolation weight is based on an exponential decay function.

[0085] The gated attention weight is jointly calculated from physical prediction values, data gradients, node flows, and rainfall intensities.

[0086] The Manning coefficient of the hydraulic model is dynamically adjusted according to the pipeline siltation degree, which is calculated based on the cumulative flow deviation.

[0087] Generate extreme rainfall scenario data through physically constrained GAN, and the generator loss function includes the residual term of the mass conservation equation.

[0088] During data processing, convert the pipe network topology into an attribute graph structure, and extract features such as depression depth, impervious rate, and drainage level. Use the discriminator to constrain the recession rate of the samples generated by GAN not to exceed the theoretical extreme value of the hydraulic model.

[0089] This method proposes an urban waterlogging recession prediction scheme that deeply couples physical mechanisms with spatio-temporal big data. By constructing a dynamic association model between the pipe network topology and the surface inundation field, combining the gated attention mechanism to adaptively correct the systematic deviation of the physical model, and embedding the mass conservation equation to constrain the prediction boundary of data-driven methods, finally achieve a refined simulation of the underground-surface hydraulic interaction process, solve the problem of high-precision prediction of recession time and waterlogging range in complex urban environments, and at the same time improve the generalization ability of the model in unseen scenarios such as pipe blockages and rainstorm extrapolation, providing a reliable decision-making basis for urban waterlogging emergency dispatch and drainage facility optimization.

[0090] The implementation of this method includes:

[0091] I. Multi-source data fusion and perception:

[0092] Data integration: Spatially and temporally align the scattered pipe network topology data (pipe connection relationships, pipe diameters and slopes), dynamic rainfall data (radar gridded rainfall), and historical inundation data (water level monitoring points, camera images), and uniformly map them to a 500-meter gridded geographic coordinate system with a time resolution accurate to 10 minutes. Construct a bidirectional topology graph for the pipe network system, where nodes represent inspection wells / pumping stations, and edges represent pipe attributes (including flow capacity, blockage risk labels).

[0093] Abnormal data processing: For sensor noise, use spatio-temporal sliding window filtering (the previous and next 3 time steps + adjacent 3 grids) to remove outliers. For missing pipe network parameters (such as unknown pipe diameters), infer their approximate values based on graph neural networks and verify their reliability in combination with historical drainage records.

[0094] II. Mechanism + mathematical double-engine collaborative prediction:

[0095] Physical model engine: Based on the SWMM hydraulics kernel, it simulates the interaction process between pipe network water flow and surface runoff, and outputs the theoretical recession curve. Introduce a dynamic parameter correction module: According to the real-time rainfall intensity and pipeline load status, automatically adjust parameters such as Manning coefficient and infiltration rate to alleviate the model deviation caused by pipeline blockage.

[0096] Data-driven engine: Prediction of surface inundation field: Use the ConvLSTM network to learn the spatio-temporal propagation law of the historical water accumulation depth field, and capture local factors not considered by traditional models (such as road enclosures and temporary drainage facilities). Prediction of pipe network status: Analyze the topological relationship of the pipe network through the graph attention network (GNN) to predict the flow change trend of key nodes (such as pumping stations and easy-to-block points).

[0097] Adaptive fusion decision-making: Design an intelligent weight allocation module to dynamically adjust the contribution ratio of the physical model and the data model according to the real-time scenario: When the rainfall intensity is low (such as <30mm / h), the physical model is mainly used (weight 70%), relying on its stable mechanism; When in a high-load state (such as pipe full flow and heavy rainfall), gradually increase the weight of the data model to 80%, and utilize its ability to capture complex non-linear relationships; Automatically generate a spatially differentiated fusion weight map through a gated neural network. For example, assign a higher physical weight to areas with dense underground pipe networks, while rely on data-driven prediction for low-lying areas lacking pipe network data.

[0098] III. Closed-loop feedback optimization:

[0099] Real-time data assimilation: Connect real-time data streams such as manhole cover water level sensors and road cameras, and dynamically correct the model output through Kalman filtering, updating the prediction results every 5 minutes. When it is detected that the actual recession speed continuously deviates from the predicted value (such as the error > 15%), automatically trigger the recalibration of model parameters.

[0100] Emergency handling for extreme scenarios: Early warning of pipe full flow: If the flow rate at a certain node exceeds 90% of the design capacity, immediately retrieve the backup model (trained based on historical similar events) for local prediction replacement. Extrapolated rainfall compensation: When the rainfall intensity exceeds the historical maximum record, activate the physical constraint enhancement mode to force the model to output the upper limit of the recession rate that conforms to mass conservation.

[0101] The specific implementation steps of this method are as follows:

[0102] 1. Multi-source data fusion and enhancement.

[0103] (1) Data collection and cleaning.

[0104] Data sources: Underground pipeline network data: Drainage pipeline CAD drawings are obtained from the urban water department, including attributes such as pipe diameter, slope, material, and inspection well coordinates. Surface data: A DEM terrain model with a resolution of 5 cm is generated through drone oblique photography, and key surface features such as roadside curbs and green space boundaries are marked. Dynamic data: Radar rainfall grid data from the meteorological bureau (5 km × 5 km, updated every 12 hours) and Internet of Things water level sensors (deployed at flood-prone points, sampled every 1 minute) are accessed.

[0105] Outlier processing: For water level sensor data, sliding window Z-Score detection is used (window size 30 minutes, threshold ±3σ) to remove instantaneous pulse noise. For missing pipe diameter attributes in the pipeline network CAD data, they are filled based on the weighted average of the upstream and downstream pipe diameters (weight is the reciprocal of the pipe length).

[0106] (2) Spatiotemporal alignment and feature engineering.

[0107] Spatial alignment: Use QGIS to convert the pipeline network node coordinates to the WGS84 coordinate system and overlay them with the radar rainfall grid. Kriging interpolation (Gaussian model is selected for the semivariogram) is performed on discrete water level monitoring points to generate a 500m × 500m gridded historical inundation field. Temporal alignment: Taking rainstorm events as units, complete temporal data from 1 hour before the start of rainfall to after the recession of water is intercepted. For non-uniformly sampled data (such as 15 minutes of rainfall + 5 minutes of water level), cubic spline interpolation is used to unify it to a 10-minute interval. Key feature construction: Pipeline network topological features: Calculate the "drainage level" (the shortest number of pipes from the outlet) and "water collection capacity" (total cross-sectional area of downstream pipes) for each node. Surface features: Extract the "depression depth" (height difference between the local terrain lowest point and the surroundings) and "proportion of impervious area" (based on satellite image segmentation) for each grid.

[0108] 2. Dual-engine collaborative prediction of physical model + data model.

[0109] (1) Physical model optimization:

[0110] Dynamic calibration of SWMM parameters: Add a "deposition degree - Manning coefficient" feedback loop to the traditional SWMM model:

[0111] N(t) = n_base × (1 + 0.2 × Clog_level);

[0112] where Clog_level = cumulative flow prediction deviation / design flow.

[0113] The infiltration parameters are automatically calibrated every 6 hours, and the SCE-UA optimization algorithm is used to minimize the node flow error.

[0114] Enhanced boundary conditions: The road curb flow capacity is input as a dynamic boundary condition. The calculation formula is as follows:

[0115] Q_road = C × L × h^1.5;

[0116] Where, C is the curb opening coefficient, L is the curb length, and h is the depth of road surface water.

[0117] (2) Data-driven model training:

[0118] ConvLSTM surface inundation prediction: Input: Rainfall sequence for the past 3 hours (6×6 grid) + current terrain features (depression depth, slope). Training strategy: Curriculum learning is used, training first for light rain events and gradually increasing the complexity of heavy rain samples.

[0119] GNN Pipeline Network Status Prediction: Graph Construction: Defines pipeline network nodes as graph nodes, pipes as directed edges, and edge attributes including diameter, slope, and length. Message Passing: Utilizes the GATv2 network to aggregate flow and water depth information from neighboring nodes. Output: 1-hour probability distribution of flow and fullness for each node.

[0120] (3) Implementation of adaptive fusion module:

[0121] Gated Weight Generation Network: Input: Node flow predicted by the physical model + Inundation depth predicted by the data model + Real-time rainfall intensity. Structure: 3-layer fully connected network (128→64→1 neurons), output Sigmoid activated weight value α∈[0,1]. Dynamic adjustment rules:

[0122] Condition Weight α adjustment Rainfall intensity > 50 mm / h α←α×0.7+0.15 Pipe fullness > 80% for 30 minutes α ← max(0.3, α - 0.2) Sensor data is abnormal 5 times in a row α ← 0.1 (force data dominance)

[0123] 3. Online reasoning and dynamic correction.

[0124] (1) Kalman filter dynamic correction: State vector: Contains the flooding depth of each grid, the node flow rate, and the fusion weight α. Observation equation: The water level sensor directly corresponds to the grid flooding depth, and the pump station flow meter corresponds to the node flow rate. Update process: Prediction phase: The next temporal state is estimated based on the fusion model. Update phase: When new sensor data arrives, the Kalman gain K is calculated and the state estimate is corrected. A complete prediction-update cycle is performed every 15 minutes.

[0125] (2) Emergency Scenario Handling:

[0126] Sudden heavy rainfall response: When the radar predicts that the rainfall in the next hour will exceed the historical extreme value, the "physical constraint enhancement mode" is activated to force the data model output to meet the mass conservation requirement. Limit the rate of water recession not to exceed 120% of the maximum drainage capacity of the pipe network. Equipment failure tolerance: If a large area of sensors in a certain area fails, switch to the "pure physical prediction + neighboring area data compensation" mode.

[0127] 4. Model evaluation and iteration:

[0128] (1) Offline verification: To comprehensively evaluate the model performance, a spatio-temporal cross-validation method is adopted. Taking rainstorm events as independent units, the training set and the test set are divided to avoid overfitting caused by temporal correlation. The specific process includes: constructing 10 groups of rainstorm scenarios in different years (such as once-in-a-year to once-in-50-year) based on historical data, and calculating the recession time error (the absolute difference between the predicted and actual recession moments), the overlap degree of the waterlogging range (IoU), and the capture rate of extreme events (the hit rate of the top 10 flood-prone points) respectively. By comparing the prediction results of the traditional SWMM model, the pure data model and the present invention, the robustness of the fusion model in common and extreme scenarios is verified. For example, in the "Dragon Boat Water" event in Guangzhou in 2022, the average error of the model for the recession time was reduced from 2.1 hours of the traditional model to 0.7 hours, and the IoU of the waterlogging range increased by 22%. The verification results will be used to optimize the parameters of the gated network and the physical constraint weight coefficients.

[0129] (2) Online monitoring: The system tracks the running state of the model in real time through multi-dimensional indicators: constructing a dynamic monitoring dashboard based on the GIS platform to visually display the heat map of the fusion weights (spatial distribution of α values) of the physical model and the data-driven model. When the weight value fluctuates by more than ±0.3 within 1 hour, a yellow warning is triggered; comparing the predicted inundation depth with the measured values of the sensors grid by grid to generate an error heat map layer (graded in red / yellow / green). If the error exceeds 15% for three consecutive time steps, it is located at specific pipe network nodes for problem diagnosis (such as abnormal pipe blockage coefficient); synchronously recording the time series curve of the Kalman filter correction amount. When the correction amplitude in a certain area continues to expand (such as a 20% increase in the daily average adjustment amount), an automatic pipe network health assessment report is generated to prompt the key areas for inspection (such as the cumulative correction amount of node NJ-1024 reaching 38 cm in the past week). The monitoring data is stored in a time series database, supporting the retrospective analysis of the association between the model performance and external events at any time period (such as the prediction deviation caused by the pipe diameter change due to construction).

[0130] (3) Continuous learning: To address the dynamic changes in the urban drainage system (such as pipeline aging and newly built storage facilities), a three-level adaptive update mechanism is designed: 1) Incremental fine-tuning: Extract new data (rainfall-flooding records) every early morning, use the historical model as the teacher network, and fine-tune the parameters of the data-driven branch with a cosine annealing learning rate (initial value 1e-5), while constraining the physical branch to be immutable to maintain mechanism stability; 2) Extreme scenario enhancement: When a rainfall event exceeding historical records is detected, start the physical constraint GAN to generate adversarial samples. The generator synthesizes a flooding field with a rainfall intensity of 120-150 mm / h, and the discriminator enforces the mass conservation equation to expand the diversity of training data; 3) Version iteration: Perform an A / B test at the end of each month, compare the core indicators of the new and old models such as the recession time error (MAE) and the capture rate of extrapolation scenarios. Only when the comprehensive score of the new model increases by ≥5% and the false alarm rate in key areas decreases, gradually replace the online version through the gray release strategy to ensure business continuity.

[0131] This method aims at the problem that the prediction results of traditional hydraulic models are too idealistic, and proposes a spatio-temporal prediction framework that jointly trains the physical model of the drainage network and historical flooding data. By constructing a multi-source data fusion network (rainfall, terrain, pipe network topology), combined with an LSTM-graph neural network hybrid model, dynamic error correction is performed on the output of traditional hydraulic models such as SWMM. It includes: establishing a spatio-temporal correlation mapping model between the state of pipe network nodes and surface flooding; introducing a physical constraint loss function to realize the coupling of data-driven and mechanism models; designing a dynamic weight allocation mechanism to fuse real-time sensor data. This method can reduce the recession time prediction error by more than 40%, significantly improving the urban waterlogging emergency response ability.

[0132] The embodiment of the present invention also provides a numerical simulation system for the urban waterlogging recession process based on physics-data fusion, including a data preprocessing module, a physical model engine, a data model engine, a dynamic fusion and real-time correction module;

[0133] Among them, the data preprocessing module uses a graph neural network to complete missing pipe network parameters and impose a flow continuity constraint.

[0134] This system implements the numerical simulation method for the urban waterlogging recession process based on physics-data fusion described in the above embodiment.

[0135] The data preprocessing module realizes multi-source data fusion perception, including:

[0136] Data integration: Align the scattered pipe network topology data (pipe connection relationships, pipe diameters and slopes), dynamic rainfall data (radar gridded rainfall), and historical inundation data (water level monitoring points, camera images) in space and time, and uniformly map them to a 500-meter gridded geographic coordinate system with a time resolution accurate to 10 minutes. Construct a bidirectional topology graph for the pipe network system, where nodes represent inspection wells / pumping stations, and edges represent pipe attributes (including flow capacity, blockage risk labels).

[0137] Abnormal data processing: For sensor noise, use a spatio-temporal sliding window filter (3 time steps before and after + 3 adjacent grids) to remove outliers. For missing pipe network parameters (such as unknown pipe diameters), infer approximate values based on a graph neural network and verify their reliability in combination with historical drainage records.

[0138] The physical model engine, based on the SWMM hydraulics kernel, simulates the interaction process between pipe network water flow and surface runoff and outputs a theoretical recession curve. Introduce a dynamic parameter correction module: Automatically adjust parameters such as the Manning coefficient and infiltration rate according to the real-time rainfall intensity and pipe load status to mitigate model deviations caused by pipe blockages.

[0139] The data model engine includes: Prediction of the surface inundation field: Use a ConvLSTM network to learn the spatio-temporal propagation law of the historical water accumulation depth field and capture local factors not considered by traditional models (such as road enclosures, temporary drainage facilities). Prediction of the pipe network state: Analyze the pipe network topology relationship through a graph attention network (GNN) to predict the flow change trend of key nodes (such as pumping stations, easily blocked points).

[0140] Dynamic fusion: Design an intelligent weight allocation module to dynamically adjust the contribution ratio of the physical model and the data model according to the real-time scenario: At low rainfall intensities (such as <30 mm / h), mainly rely on the physical model (weight 70%) due to its stable mechanism; at high load states (such as pipe full flow, heavy rainfall), gradually increase the weight of the data model to 80% to utilize its ability to capture complex non-linear relationships; Automatically generate a spatially differentiated fusion weight map through a gated neural network. For example, assign a higher physical weight to areas with dense underground pipe networks, while rely on data-driven prediction for low-lying areas lacking pipe network data.

[0141] The real-time correction module realizes closed-loop feedback optimization:

[0142] Real-time data assimilation: Connect real-time data streams such as manhole cover water level sensors and road cameras, dynamically correct the model output through Kalman filtering, and update the prediction results every 5 minutes. When it is detected that the actual recession speed continuously deviates from the predicted value (such as error > 15%), automatically trigger re-calibration of the model parameters.

[0143] Emergency handling for extreme scenarios: Pipeline full flow warning: If the flow at a node exceeds 90% of the design capacity, a backup model (trained based on historical similar events) is immediately called up for local prediction replacement. Extrapolated rainfall compensation: When rainfall intensity exceeds the historical maximum, the physical constraint enhancement mode is activated, forcing the model to output an upper limit on the water withdrawal rate that complies with mass conservation.

[0144] The process of the system realizing the numerical simulation of urban waterlogging retreat process based on physics-data fusion is as follows:

[0145] 1. Multi-source data fusion and enhancement.

[0146] (1) Data collection and cleaning.

[0147] Data Sources: Underground Pipeline Network Data: CAD drawings of drainage pipe networks were obtained from the city's water department, including attributes such as pipe diameter, slope, material, and manhole coordinates. Surface Data: A 5cm resolution DEM terrain model was generated using drone oblique photography, with key surface features such as road curbs and green space boundaries annotated. Dynamic Data: Data from the Meteorological Bureau's radar rainfall grid (5km x 5km, updated every 12 hours) and IoT water level sensors (deployed at flood-prone areas, sampling every 1 minute) were collected.

[0148] Outlier processing: For water level sensor data, a sliding window Z-Score detection (window size 30 minutes, threshold ±3σ) was used to remove transient impulse noise. Missing pipe diameter attributes in the pipe network CAD data were filled using the weighted average of the upstream and downstream pipe diameters (weighted by the inverse of the pipe length).

[0149] (2) Spatiotemporal alignment and feature engineering.

[0150] Spatial alignment: Use QGIS to convert the coordinates of the pipe network nodes into the WGS84 coordinate system and overlay it with the radar rainfall grid. Perform Kriging interpolation on discrete water level monitoring points (Gaussian model is selected for the semi-variogram function) to generate a 500m×500m gridded historical inundation field. Temporal alignment: Take rainstorm events as units and intercept the complete time series data from 1 hour before the start of rainfall to after the water recedes. For non-uniformly sampled data (such as 15 minutes of rainfall + 5 minutes of water level), use cubic spline interpolation to unify to 10-minute intervals. Key feature construction: Pipeline network topology features: Calculate the "drainage level" (the shortest number of pipes from the outlet) and "water collection capacity" (total cross-sectional area of downstream pipes) of each node. Surface features: Extract the "depression depth" (the height difference between the lowest point of the local terrain and the surrounding area) and "percentage of impervious area" (based on satellite image segmentation) of each grid.

[0151] 2. Collaborative prediction using dual engines of physical model and data model.

[0152] (1) Physical model optimization:

[0153] Dynamic calibration of SWMM parameters: Add an "accumulation degree - Manning coefficient" feedback loop to the traditional SWMM model:

[0154] N(t) = n_base × (1 + 0.2 × Clog_level);

[0155] where Clog_level = cumulative flow prediction deviation / design flow.

[0156] Automatically calibrate the infiltration parameters every 6 hours, and use the SCE-UA optimization algorithm to minimize the node flow error.

[0157] Enhanced boundary conditions: Input the over-flow capacity of the roadside curb as a dynamic boundary condition, and the calculation formula is as follows:

[0158] Q_road = C × L × h^1.5;

[0159] where C: curb opening coefficient, L: curb length, h: road surface water depth.

[0160] (2) Data-driven model training:

[0161] ConvLSTM surface flooding prediction: Input: rainfall field sequence in the past 3 hours (6×6 grid) + current terrain features (depression depth, slope). Training strategy: Adopt Curriculum Learning, first train small rain events, and gradually increase the complexity of heavy rain samples.

[0162] GNN pipe network state prediction: Graph construction: Define pipe network nodes as graph nodes, pipes as directed edges, and edge attributes include pipe diameter, slope, and length. Message passing: Use the GATv2 network to aggregate the flow and water depth information of neighbor nodes. Output: Probability distributions of the flow and fullness of each node in the next 1 hour.

[0163] (3) Implementation of the adaptive fusion module:

[0164] Gating weight generation network: Input: node flow predicted by the physical model + flooding depth predicted by the data model + real-time rainfall intensity. Structure: 3-layer fully connected network (128→64→1 neuron), and output a weight value α ∈ [0,1] activated by Sigmoid. Dynamic adjustment rule:

[0165] Condition Weight α adjustment Rainfall intensity > 50 mm / h α←α×0.7+0.15 Pipe fullness > 80% for 30 minutes α ← max(0.3, α - 0.2) Sensor data is abnormal 5 times in a row α ← 0.1 (force data dominance)

[0166] 3. Online inference and dynamic correction.

[0167] (1) Kalman filter dynamic correction: State vector: Contains the flooding depth of each grid, the node flow rate, and the fusion weight α. Observation equation: The water level sensor directly corresponds to the grid flooding depth, and the pump station flow meter corresponds to the node flow rate. Update process: Prediction phase: The next temporal state is estimated based on the fusion model. Update phase: When new sensor data arrives, the Kalman gain K is calculated and the state estimate is corrected. A complete prediction-update cycle is performed every 15 minutes.

[0168] (2) Emergency Scenario Handling:

[0169] Sudden heavy rainfall response: When the radar predicts that the rainfall in the next hour will exceed the historical extreme value, the "physical constraint enhancement mode" is activated to force the data model output to meet the mass conservation requirement. Limit the water withdrawal rate to no more than 120% of the maximum drainage capacity of the pipe network. Equipment fault tolerance: If a large area of sensors in a certain area fails, it switches to the "pure physical prediction + adjacent area data compensation" mode.

[0170] 4. Model evaluation and iteration:

[0171] (1) Offline validation: In order to comprehensively evaluate the performance of the model, a spatiotemporal cross-validation method is used to divide the training set and the test set into independent units based on rainstorm events to avoid overfitting caused by time correlation. The specific process includes: constructing 10 groups of rainstorm scenarios in different years based on historical data (such as once in 1 year to once in 50 years), and calculating the water withdrawal time error (the absolute difference between the predicted and actual withdrawal time), the overlap of the water accumulation range (IoU) and the extreme event capture rate (the hit rate of the top 10 flood-prone points). By comparing the prediction results of the traditional SWMM model, the pure data model and the present invention, the robustness of the fusion model in common and extreme scenarios is verified. For example, in the "Dragon Boat Flood" event in Guangzhou in 2022, the average error of the model for the water withdrawal time was reduced from 2.1 hours of the traditional model to 0.7 hours, and the water accumulation range IoU increased by 22%. The validation results will be used to optimize the gating network parameters and physical constraint weight coefficients.

[0172] (2) Online monitoring: The system tracks the running status of the model in real time through multi-dimensional indicators: constructs a dynamic monitoring dashboard based on the GIS platform, and visually displays the fusion weight heat map (spatial distribution of α value) of the physical model and the data-driven model. When the weight value fluctuates by more than ±0.3 within 1 hour, a yellow warning is triggered; compares the predicted inundation depth with the measured values of sensors grid by grid to generate an error heat map layer (graded in red / yellow / green). If the error exceeds 15% for three consecutive time steps, it is located to specific pipe network nodes for problem diagnosis (such as abnormal pipe blockage coefficient); synchronously records the time series curve of the Kalman filter correction amount. When the correction amplitude in a certain area continues to expand (such as a 20% increase in the daily average adjustment amount), an automatic pipe network health assessment report is generated, indicating the key areas for inspection (such as the cumulative correction amount of node NJ-1024 reaching 38 cm in the past week). The monitoring data is stored in a time series database, supporting the correlation analysis of the model performance and external events at any time period in the past (such as the prediction deviation caused by the pipe diameter change due to construction).

[0173] (3) Continuous learning: To cope with the dynamic changes of the urban drainage system (such as pipe aging, newly built storage facilities), a three-level adaptive update mechanism is designed: 1) Incremental fine-tuning: Extract new data (rainfall-inundation records) every morning, use the historical model as the teacher network, and fine-tune the parameters of the data-driven branch with a cosine annealing learning rate (initial value 1e-5), and constrain the physical branch to be immutable to maintain the mechanism stability; 2) Extreme scenario enhancement: When a rainfall event exceeding the historical record is detected, start the physical constraint GAN to generate adversarial samples. The generator synthesizes an inundation field with a rainfall intensity of 120-150 mm / h, and the discriminator is forced to satisfy the mass conservation equation to expand the diversity of training data; 3) Version iteration: Perform A / B testing at the end of each month, compare the core indicators such as the recession time error (MAE) and the capture rate of extrapolation scenarios between the new and old models. Only when the comprehensive score of the new model increases by ≥5% and the false alarm rate in the key area decreases, gradually replace the online version through the gray release strategy to ensure business continuity.

[0174] The embodiment of the present invention also provides a computer-readable medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the numerical simulation method for the urban waterlogging recession process based on physical-data fusion described in the above embodiment is implemented. Specifically, a system or device equipped with a storage medium can be provided, on which software program codes for implementing the functions of any one of the above embodiments are stored, and the computer (or CPU or MPU) of the system or device is made to read and execute the program codes stored in the storage medium.

[0175] In this case, the program code read from the storage medium itself can implement the functions of any of the above-described embodiments. Therefore, the program code and the storage medium storing the program code constitute a part of the present invention.

[0176] Examples of the storage medium for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as a CD-ROM, a CD-R, a CD-RW, a DVD-ROM, a DVD-RAM, a DVD-RW, a DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program code can be downloaded from a server computer via a communication network.

[0177] In addition, it should be clear that not only can the functions of any of the above-described embodiments be implemented by executing the program code read by the computer, but also by causing an operating system or the like operating on the computer based on the instructions of the program code to complete part or all of the actual operations.

[0178] In addition, it can be understood that the program code read from the storage medium is written into the memory provided in an expansion board inserted into the computer or into the memory provided in an expansion unit connected to the computer, and then based on the instructions of the program code, a CPU or the like installed on the expansion board or the expansion unit is caused to execute part or all of the actual operations, thereby implementing the functions of any of the above-described embodiments.

[0179] The present invention has been described in detail above with reference to the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above-described multiple embodiments, those skilled in the art can know that the code review means in different embodiments can be combined to obtain more embodiments of the present invention, and these embodiments are also within the protection scope of the present invention.

Claims

1. A numerical simulation method for the urban waterlogging recession process based on the integration of physics and data, characterized in that, Fuse the hydraulic model and the data-driven model, dynamically weight the prediction results through gated attention, and use Kalman filtering for real-time correction; The implementation of this method includes the following steps: 1) Multi-source data fusion and enhancement: Collect and process data, achieve spatio-temporal alignment, and construct features; 2) Mechanism + mathematical dual-engine collaborative prediction: Based on the physical model engine and the data-driven engine for collaborative prediction, and achieve adaptive fusion; 3) Online inference and dynamic correction: Include Kalman filtering dynamic correction and extreme scenario emergency handling; 4) Model evaluation and iteration.

2. The numerical simulation method for the urban waterlogging recession process based on physical-data fusion according to claim 1, wherein The data processing transforms the pipe network topology into an attribute graph structure, and extracts features such as depression depth, impervious rate, and drainage level; The multi-source data fusion and enhancement specifically include: Data integration: Align the scattered pipe network topology data, dynamic rainfall data (radar gridded rainfall), and historical inundation data in space and time, and uniformly map them to the gridded geographic coordinate system; Construct a bidirectional topology graph for the pipe network system, where the nodes represent inspection wells / pumping stations, and the edges represent pipe attributes; Abnormal data processing: For sensor noise, use a spatio-temporal sliding window filter to remove outliers; For missing pipe network parameters, infer their approximate values based on a graph neural network, and verify the reliability in combination with historical drainage records; The spatio-temporal alignment, spatial alignment: Use QGIS to convert the pipe network node coordinates to the WGS84 coordinate system, overlay with the radar rainfall grid, perform Kriging interpolation on the discrete water level monitoring points, and generate a gridded historical inundation field; Temporal alignment: Take the rainstorm event as a unit, intercept the complete time series data from 1 hour before the start of rainfall to after the recession of floodwater, and for non-uniformly sampled data, use cubic spline interpolation to unify it to a 10-minute interval; Key feature construction includes: Pipe network topology features: Calculate the "drainage level" and "catchment capacity" of each node; Surface features: Extract the "depression depth" and "proportion of impervious area" of each grid.

3. The numerical simulation method for the urban waterlogging recession process based on physical-data fusion according to claim 1 or 2, characterized in that For the data integration, use Delaunay-MLS interpolation to map the discrete water level data into a continuous inundation field, and the interpolation weight is based on an exponential decay function.

4. The numerical simulation method for the urban waterlogging drainage process based on physical-data fusion according to claim 1, wherein The mechanism + mathematical dual-engine collaborative prediction, Physical model engine: Based on the SWMM hydraulic kernel, simulate the interaction process between the pipe network water flow and surface runoff, and output the theoretical recession curve; Introduce a dynamic parameter correction module: Automatically adjust parameters according to the real-time rainfall intensity and pipe load status, including Manning coefficient and infiltration rate, to alleviate the model deviation caused by pipe blockage; Data-driven engine: Include surface inundation field prediction and pipe network status prediction, Surface inundation field prediction: Use a ConvLSTM network to learn the spatio-temporal propagation law of the historical waterlogging depth field, and capture local factors not considered by traditional models, including road enclosures and temporary drainage facilities; Pipe network status prediction: Analyze the pipe network topology relationship through a graph attention network to predict the flow change trend of key nodes; Adaptive Fusion Decision: Through the intelligent weight allocation module, dynamically adjust the contribution ratio of the physical model and the data model according to the real-time scenario: When the rainfall intensity is low, mainly rely on the physical model and its stable mechanism; When in a high-load state, gradually increase the weight of the data model to 80%, and utilize its ability to capture complex non-linear relationships; Automatically generate a spatially differentiated fusion weight map through a gated neural network; The gated attention weight is jointly calculated based on the physical prediction value, data gradient, node flow, and rainfall intensity.

5. The numerical simulation method for the urban waterlogging recession process based on physical-data fusion according to claim 1 or 4, characterized in that, The Manning coefficient of the hydraulic model is dynamically adjusted according to the pipeline siltation degree, and the siltation degree is calculated based on the cumulative flow deviation; The specific process of optimizing the physical model is as follows: Dynamic calibration of SWMM parameters: Add a "siltation degree - Manning coefficient" feedback loop to the traditional SWMM model: n(t) = n_base × (1 + 0.2 × Clog_level); Where, Clog_level = cumulative flow prediction deviation / design flow; Automatically calibrate the infiltration parameters every 6 hours, and use the SCE-UA optimization algorithm to minimize the node flow error; Enhanced boundary conditions: Input the over-flow capacity of the roadside curb as a dynamic boundary condition, and the calculation formula is: Q_road = C × L × h^1.5; Where, C: curb opening coefficient, L: curb length, h: road surface water depth; The training of the data-driven model includes the following contents: ConvLSTM surface flooding prediction: Input: rainfall field sequence in the past 3 hours + current terrain features; Training strategy: Adopt curriculum learning, first train small rain events, and gradually increase the complexity of heavy rain samples; GNN pipe network state prediction: Graph construction: Define the pipe network nodes as graph nodes, and the pipes as directed edges. The edge attributes include pipe diameter, slope, and length; Message passing: Use the GATv2 network to aggregate the flow and water depth information of neighbor nodes; Output: Probability distribution of the flow and fullness degree of each node in the next 1 hour; The implementation of the adaptive fusion module includes: Gated weight generation network: Input: node flow predicted by the physical model + flooding depth predicted by the data model + real-time rainfall intensity; Structure: 3-layer fully connected network, and the output is a weight value α ∈ [0, 1] activated by Sigmoid; Dynamic adjustment rules: When the rainfall intensity > 50mm / h, the weight α is adjusted to: α ← α × 0.7 + 0.15; When the pipe fullness degree > 80% lasts for 30 minutes, the weight α is adjusted to: α ← max(0.3, α - 0.2); When the sensor data is abnormal continuously for 5 times, the weight α is adjusted to: α ← 0.1 (forcing data dominance).

6. The numerical simulation method for the urban waterlogging recession process based on physical-data fusion according to claim 1, characterized in that Generate extreme rainfall scenario data through a physically constrained GAN, and the generator loss function includes a residual term of the mass conservation equation.

7. The numerical simulation method for the urban waterlogging drainage process based on physical-data fusion according to claim 1, wherein The online inference and dynamic correction mentioned above realize closed-loop feedback optimization, including: Real-time data assimilation: Access real-time data streams, including manhole cover water level sensor and road camera data streams, dynamically correct the model output through Kalman filtering, and update the prediction results every fixed time; When it is detected that the actual water recession speed continuously deviates from the predicted value, automatically trigger the recalibration of model parameters; Among them, the dynamic correction process of Kalman filtering includes: State vector: includes the inundation depth of each grid, node flow, and fusion weight α; Observation equation: the water level sensor directly corresponds to the inundation depth of the grid, and the pump station flowmeter corresponds to the node flow; Update process: In the prediction stage, the state of the next time step is inferred based on the fusion model; in the update stage, when new sensor data arrives, the Kalman gain K is calculated to correct the state estimate; a complete prediction-update cycle is executed every fixed time duration; Emergency handling for extreme scenarios, including: Full-flow warning for pipelines: If the flow of a certain node exceeds 90% of the design capacity, immediately retrieve the backup model for local prediction replacement; Extrapolated rainfall compensation: When the rainfall intensity exceeds the historical maximum record, the physical constraint enhancement mode is activated to force the model to output an upper limit of the recession rate that conforms to mass conservation; The sudden heavy rainfall response includes: When the radar predicts that the rainfall in the next 1 hour exceeds the historical extreme value, the "physical constraint enhancement mode" is activated to force the data model output to satisfy mass conservation: Limit the recession rate not to exceed 120% of the maximum drainage capacity of the pipe network; Equipment failure tolerance: If a large area of sensors in a certain area fails, switch to the "pure physical prediction + adjacent area data compensation" mode.

8. The numerical simulation method for the urban waterlogging recession process based on physical-data fusion according to claim 1 or 7, characterized in that The model evaluation and iteration include: Offline verification: Adopt the spatio-temporal cross-validation method, divide the training set and test set with rainstorm events as independent units to avoid overfitting caused by time correlation; the specific process includes: constructing multiple groups of rainstorm scenarios in different years based on historical data, calculating the recession time error, overlap degree of waterlogging range, and capture rate of extreme events respectively, and verifying the robustness of the fusion model in normal and extreme scenarios by comparing the prediction results of the traditional SWMM model, pure data model, and this method; the verification results will be used to optimize the parameters of the gated network and the weight coefficients of physical constraints; Online monitoring: Real-time track the running state of the model through multi-dimensional indicators, including: constructing a dynamic monitoring dashboard based on the GIS platform, visually displaying the heat map of the fusion weight of the physical model and the data-driven model, triggering a yellow warning when the weight value fluctuates beyond the set threshold within a specified time; comparing the predicted inundation depth with the measured value of the sensor grid by grid to generate an error heat map layer, and if the error exceeds the set range for multiple consecutive time steps, locate to specific pipe network nodes for problem diagnosis; synchronously record the time series curve of the Kalman filter correction amount, and automatically generate a pipe network health assessment report to prompt the key areas for inspection when the correction amplitude in a certain area continues to expand; the monitoring data is stored in a time series database to support the correlation analysis of the model performance and external events at any time period in the past; Continuous learning: To cope with the dynamic changes of the urban drainage system, the following three-level adaptive update mechanism is designed: Incremental fine-tuning: Extract new data every morning, use the historical model as the teacher network, and fine-tune the parameters of the data-driven branch with a cosine annealing learning rate, and constrain the physical branch to be immutable to maintain the mechanism stability; Extreme scenario enhancement: When a rainfall event beyond historical records is detected, start the physical constraint GAN to generate adversarial samples. The generator synthesizes the rainfall intensity inundation field, and the discriminator enforces the mass conservation equation: Expand the diversity of training data; Version iteration: Perform an A / B test at the end of each month, compare the core indicators such as recession time error and capture rate of extrapolation scenarios between the new and old models. Only when the comprehensive score of the new model increases by ≥5% and the false alarm rate in the key area decreases, gradually replace the online version through the gray release strategy to ensure business continuity.

9. A numerical simulation system for the urban waterlogging recession process based on physical-data fusion, characterized in that, Including data preprocessing, physical model engine, data model engine, dynamic fusion, and real-time correction modules; Among them, the data preprocessing module uses a graph neural network to complete the missing pipe network parameters and impose flow continuity constraints; This system implements the method described in any one of claims 1 to 8.

10. A computer-readable medium, characterized in that, Computer instructions are stored on the computer-readable medium, and when the computer instructions are executed by a processor, the method described in any one of claims 1 to 8 is implemented.

Images