Sponge city high terrain rainwater management method and system

Abstract

The invention discloses a sponge city high-terrain rainwater management method and system, particularly relates to the technical field of rainwater prediction management, and is used for solving the problem of poor high-terrain rainwater prediction management scheduling. For a high terrain small watershed, data such as elevation, permeability coefficient, soil layer thickness and vegetation coverage are collected and preprocessed, a continuous physical field is generated, a facility pipe network is mapped, and then grid units are divided; a multivariate predictor is constructed based on grid cell center attributes, edge node increment fine adjustment is issued after cloud training, and a model is optimized through scene matching and federal learning; after rainstorm, automatically extracting observation and prediction residual increment to train a neural network and generating a confidence interval; and the edge nodes divide risks according to the confidence interval, the flood control water depth and the minimum infiltration threshold value and output gate opening and pump set power suggestions, so that the high-terrain small-watershed extreme rainstorm peak prediction precision and response time efficiency are remarkably improved, and refined risk perception scheduling and rainwater resource utilization are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of rainwater prediction management. More specifically, the present invention relates to a rainwater management method and system for high terrains in sponge cities. Background Art

[0002] A sponge city is an ecological urban development model that simulates the natural hydrological cycle and improves the city's ability to absorb, store, purify, and reuse rainwater. Its core goal is to make the city act like a "sponge", infiltrating, retaining, storing, and purifying rainwater during rainfall, and releasing and utilizing the stored water resources during droughts, thereby alleviating problems such as waterlogging, drought, and water pollution, and achieving sustainable management of urban water resources.

[0003] Deficiencies of the prior art: In the extreme rainstorm management of small watersheds with high terrains in sponge cities, it mostly relies on single statistical rainfall and runoff models or pure data-driven methods, ignoring the amplification effect of the sharp undulations of high terrains on the timing and intensity of rainstorm runoff peaks, and not incorporating the hydrodynamic conservation relationship into the model constraints. Therefore, systematic prediction errors such as late peaks or low intensities often occur under short-term heavy rainfall conditions. For example, in a mountainous small watershed, 90 millimeters of rainfall suddenly occurs within 1 hour, and the traditional model only predicts a peak of 60 millimeters with a delay of 20 minutes, resulting in the failure of the storage tank to open the sluice for flood discharge in time, severely restricting the refined effect of waterlogging prevention and groundwater recharge in high terrain areas, and affecting the urban water cycle scheduling management. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, the present invention provides a rainwater management method and system for high terrains in sponge cities to solve the problem of poor rainwater prediction management and scheduling in the high terrain as described in the above background art.

[0005] To achieve the above object, the present invention provides the following technical solutions: A rainwater management method for high terrains in sponge cities, comprising the following steps: Collect and preprocess data for high terrain areas, and combine borehole permeability coefficients, soil layer thicknesses, and remote sensing vegetation coverage interpolation. After removing anomalies, generate a continuous physical field, map facility and pipeline information to the elevation coordinate system, and output a facility information layer and a facility attribute list after topological repair, and divide grid cells; Construct a predictor based on the attributes of the center of the grid cell, train it in the cloud and then send it to the edge node for local incremental fine-tuning, dynamically match the primary predictor according to the scenario performance matrix, and generate a global update package through federated learning to synchronously update the local model prediction; After each rainstorm event, the edge nodes automatically extract the latest observed and predicted residuals and use them as training samples to perform incremental iteration on the neural network. After fine-tuning, a random masking mechanism is used for forward inference, and the central tendency and fluctuation range of the predicted values are statistically analyzed to generate the confidence interval of the model. Based on the predicted confidence interval, flood control water depth threshold, and minimum infiltration threshold, the edge nodes divide the high-risk, low-risk, and safe units and generate suggestions for the gate opening and pump group power.

[0006] In a preferred embodiment, data collection and preprocessing are carried out on the high-elevation area, and combined with the borehole permeability coefficient, soil layer thickness, and remote sensing vegetation coverage interpolation, and after removing anomalies, a continuous physical field is generated. The specific process is as follows: Obtain the original elevation data of the high-elevation area, perform projection transformation to unify it to the same coordinate system, apply Gaussian smoothing filter to the projected elevation data to remove measurement noise, form smooth elevation information, and then perform polynomial error correction on the smoothed elevation information according to the ground control points, and output the corrected elevation information; Interpolate the soil and vegetation parameters of the high-elevation area, obtain the soil permeability coefficient data and soil layer thickness data from borehole sampling, and obtain the vegetation coverage data from remote sensing classification; Apply the Kriging method to interpolate the soil permeability coefficient, soil layer thickness, and vegetation coverage data respectively to generate continuous permeability coefficient fields, soil layer thickness fields, and vegetation coverage fields; Remove outliers from the interpolated permeability coefficient field, soil layer thickness field, and vegetation coverage field, and interpolate again to fill the holes.

[0007] In a preferred embodiment, map the facility and pipeline information to the elevation coordinate system, and after topological repair, output the facility information layer and the facility attribute list, and divide the grid cells. The specific process is as follows: Extract the facility and pipeline information from the regional design drawing, including the spatial positions and structural parameters of the infiltration pond, storage pond, and main drainage pipeline; Map the extracted facility and pipeline information to the coordinate system of the corrected elevation information to generate the facility information layer, perform topological repair on the broken or overlapping pipe segments in the facility information layer, and output the complete facility attribute list; Based on the corrected elevation information, calculate the local slope and curvature at each location in the GIS environment or a dedicated script, call the Delaunay triangulation algorithm, and automatically generate a series of triangular grid cells with the side length function as the constraint; Take the central elevation, permeability coefficient, soil layer thickness, and vegetation coverage of each grid cell as the initial conditions, take the real-time rainfall intensity and boundary water level observation values as the boundary inputs, and calculate the water depth change and flow velocity distribution of each cell at each time step through the finite volume iteration method.

[0008] In a preferred embodiment, a predictor is constructed based on the attributes of the grid unit center, and after training in the cloud, it is sent to the edge node for local incremental fine-tuning. The specific steps are as follows: Four complementary predictors are constructed based on the central elevation, permeability, soil thickness, vegetation coverage, historical rainfall time series and corresponding water depth observation attributes of each grid cell. The predictor includes a simulation error correction predictor trained using a decision tree regression algorithm with the difference between the physical simulation result and the sensor observation as the target; Taking the physical equation consistency constraint and water depth observation as double losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the conservation of hydrodynamics. Taking the multi-dimensional attributes of grid cells and historical runoff as input, the attribute association predictor based on the ensemble tree algorithm is trained to mine the nonlinear interactions between soil, vegetation and terrain. Taking the time-dependent characteristics of rainfall and runoff as input, the long short-term memory network is trained to capture the fluctuation trend after the sudden change of rainfall; After each predictor completes batch training for the first time in the cloud, it is sent to the edge nodes of each basin for local fine-tuning. The fine-tuning process uses the residual between the current rainfall observation and the batch training results as a new sample for iterative update.

[0009] In a preferred embodiment, the main predictor is dynamically matched according to the scenario performance matrix, and the local model prediction is synchronously updated after a global update package is generated through federated learning aggregation. The specific process is as follows: According to the historical backtesting database, short-term heavy rainfall, long-term weak rainfall and typical rainstorm events with uneven spatial distribution are divided into different scenario categories, and the peak arrival time, peak height and fluctuation trend performance indicators of each predictor under each category are statistically analyzed to form a scenario performance matrix; According to the current rainfall time series and terrain segment division, the most suitable scenario category is automatically matched, and the best predictor under the scenario category is extracted from the matrix as the main model. If the water depth curve output by the main model deviates from the real-time observation data beyond the preset threshold, the backup model calling mechanism is triggered, and the suboptimal models are switched in sequence until the prediction accuracy is restored; The rainfall time series includes the inflection point of the cumulative rainfall curve and the time of sudden change of rainfall intensity; After completing local fine-tuning at each watershed edge node, the difference between the local model parameters and the initial model parameters is uploaded to the cloud, which aggregates the parameter differences from each node; For the same network layer or decision tree branch structure, the majority improvement selection strategy is adopted: for the parameter direction that has been improved in most nodes, the update is retained, and for the parameters that have changed in some nodes, they are regarded as noise suppression and are not adopted; After screening, a unified global update package is formed and sent back to each edge node. After receiving the update package, each edge node applies it to the local model for prediction; After the global update deployment is completed, each edge node immediately conducts cross-domain backtesting, running the new generation model in the original typical short-term heavy rainfall, long-term weak rainfall, and uneven spatial distribution scenarios respectively, calculating the root mean square error and maximum absolute error indicators, and comparing them with the results of the historical version. If there is a rise in the error indicator of a basin or it fails to meet the expectation, the online performance feedback reporting mechanism is automatically triggered, and the error distribution of the basin and the real-time observation residuals are uploaded to the cloud together.

[0010] In a preferred implementation, after each rainstorm event, the edge node automatically extracts the latest observation and prediction residuals and uses them as training samples to perform incremental iteration on the neural network. The specific process is as follows: After each rainstorm event ends, the difference between the real water depth sequence recorded by the local sensor and the predicted water depth sequence output by the integrated predictor is automatically summarized to construct a residual data set containing timestamps and corresponding residuals. For the latest period data in the residual data set, the nearest 50 residuals are selected in descending order as the fine-tuning samples; After the sample preparation is completed, the physics-informed neural network model is loaded on the edge node, and an iterative training process with the least squares residual sum as the loss function is set. Ten rounds of backpropagation optimization are performed with a learning rate of 0.005 and a batch size of 10. After fine-tuning, the snapshot of the updated model parameters and the latest residual statistics are saved together, replacing the old model, and immediately used for the water depth prediction in the next period.

[0011] In a preferred implementation, after fine-tuning, a random masking mechanism is used for forward inference, and the central tendency and fluctuation range of the predicted values are statistically analyzed to generate the confidence interval of the model. The specific process is as follows: After fine-tuning, the system conducts self-check on the new model at the edge node: using the latest 50 observation inputs, recalculating the predicted output and comparing it with the real value. If the root mean square error decreases by more than or equal to 10% compared with that before fine-tuning, it is confirmed that this online iteration is effective; otherwise, it returns to the previous effective parameter state and reports to the cloud. After completing the online fine-tuning, enter the uncertainty quantification link. For the input data at the same moment, randomly mask some connection nodes in the network to perform forward inference, simulating the output fluctuations of the model under different internal states. Each inference generates a set of water depth prediction values. After summarizing the prediction results, calculate the arithmetic mean and standard deviation, and based on the mean and standard deviation, deduce the upper and lower limits of the confidence interval. After obtaining the confidence interval, compare the confidence interval with the preset flood control and infiltration thresholds: When the upper confidence limit approaches or exceeds the flood control threshold, automatically reserve more storage space and activate the emergency diversion strategy in advance; When the lower confidence limit is far below the infiltration requirement, reduce the operating power of the pump group and perform natural infiltration; If the entire confidence interval falls within the safe interval, perform regular scheduling.

[0012] In a preferred embodiment, the edge node divides high-risk, low-risk, and safe units according to the predicted confidence interval, the flood control water depth threshold, and the minimum infiltration threshold, and generates suggestions for the gate opening and pump group power. The specific process is as follows: Compare the upper and lower limits of the predicted confidence interval of each triangular grid unit with the preset flood control water depth threshold and the minimum infiltration threshold one by one, and classify the units into high-risk, low-risk, and safe categories; When the upper limit of the confidence interval exceeds the flood control water depth threshold, it is determined as a high-risk grid unit, and the use ratio of the storage pond is preferentially increased; When the lower limit of the confidence interval is lower than the minimum infiltration threshold, it is determined as a low-risk grid unit, and the capacity suggestion of the infiltration pond is preferentially reserved for natural infiltration treatment; For the grid units in the safe interval, generate the default opening degree and power interval of the gate and pump group according to the regular scheduling logic.

[0013] A high-elevation rainwater management system for a sponge city, used to implement the above-mentioned high-elevation rainwater management method for a sponge city, including: A high-elevation network division module, used to collect and preprocess data in the high-elevation area, interpolate the borehole permeability coefficient, soil layer thickness, and remote sensing vegetation coverage, generate a continuous physical field after removing anomalies, map the facility and pipeline information to the elevation coordinate system, and output the facility information layer and the facility attribute list after topological repair, and divide the grid units; A grid fine-tuning prediction module, used to build a predictor based on the attributes of the grid unit center, perform incremental fine-tuning on the edge node locally after training in the cloud, dynamically match the primary predictor according to the scenario performance matrix, and generate a global update package through federated learning to synchronously update the local model prediction; A prediction model adjustment module, used to automatically extract the latest observations and prediction residuals by the edge node after each rainstorm event, and use them as training samples to perform incremental iteration on the neural network. After fine-tuning, use the random masking mechanism for forward inference, statistically analyze the central tendency and fluctuation range of the predicted values, and generate the confidence interval of the model; The management suggestion generation module is used for edge nodes to divide high-risk, low-risk and safe units according to the prediction confidence interval, flood control depth threshold and minimum infiltration threshold, and generate gate opening and pump group power recommendations.

[0014] Technical effects and advantages of the present invention: The present invention systematically collects and preprocesses data in high-altitude areas, integrates borehole permeability coefficient, soil thickness and remote sensing vegetation coverage interpolation to generate a continuous physical field, maps infiltration pools, storage pools and pipeline information to form a facility information layer and then divides the grid units; constructs a multivariate predictor based on the center elevation and physical properties of the grid unit, sends local incremental fine-tuning to the edge node after cloud training, and relies on the scene performance matrix and the federated learning mechanism to dynamically match and synchronize the optimization model; automatically extracts the observation prediction residuals after each rainstorm event for online incremental iteration of the neural network, and generates confidence intervals through random masked forward reasoning; the edge node divides the risk level according to the confidence interval, flood control depth and minimum infiltration threshold, and outputs gate opening and pump group power recommendations, thereby significantly improving the prediction accuracy and response speed of extreme rainstorm peaks in small watersheds with high terrain, realizing refined risk perception scheduling, and enhancing the effects of rainstorm prevention and control and rainwater resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The present invention is a flow chart of a method for managing rainwater in high terrain of a sponge city.

[0016] Figure 2 This is a structural schematic diagram of a high-altitude rainwater management system for a sponge city according to the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Example 1: Figure 1 As shown, a method for managing rainwater in high terrain of a sponge city comprises the following steps: Data collection and preprocessing are performed on high-relief areas. The data are interpolated based on the borehole permeability coefficient, soil thickness and remote sensing vegetation coverage. After removing anomalies, a continuous physical field is generated. The facility and pipeline information is mapped to the elevation coordinate system. After topological repair, the facility information layer and facility attribute list are output and the grid units are divided. Build a predictor based on the attributes at the center of grid cells, train it in the cloud and then send it to the local edge node for incremental fine-tuning. Dynamically match the primary predictor according to the scenario performance matrix, and generate a global update package through federated learning to synchronously update the local model prediction; After each rainstorm event, the edge node automatically extracts the latest observation and prediction residuals and uses them as training samples to perform incremental iteration on the neural network. After fine-tuning, a random masking mechanism is used for forward inference, and the central tendency and fluctuation range of the predicted values are statistically calculated to generate the confidence interval of the model; The edge node divides high-risk, low-risk, and safe units according to the prediction confidence interval, flood control water depth threshold, and minimum infiltration threshold, and generates suggestions for gate opening and pump group power.

[0019] Step 1: Build digital twin data with enhanced physical information in high terrain areas. The specific steps are as follows: Perform data collection and preprocessing, obtain the original elevation data of the high terrain area, perform projection transformation to unify it to the same coordinate system, apply Gaussian smoothing filter to the projected elevation data to remove measurement noise, form smooth elevation information, and then perform polynomial error correction on the smoothed elevation information according to the ground control points to output the corrected elevation information; Interpolate soil and vegetation parameters. Obtain soil permeability coefficient data and soil layer thickness data from borehole sampling, and obtain vegetation coverage data from remote sensing classification; Apply Kriging interpolation to the soil permeability coefficient, soil layer thickness, and vegetation coverage data respectively to generate continuous permeability coefficient fields, soil layer thickness fields, and vegetation coverage fields; Remove outliers from the interpolated permeability coefficient field, soil layer thickness field, and vegetation coverage field, and interpolate again to fill the holes to ensure that each data field is smooth and has no isolated anomalies; Perform facility and pipeline information fusion. Extract facility and pipeline information from the regional design drawings, including the spatial positions and structural parameters of infiltration ponds, storage ponds, and main drainage pipelines. Map the extracted facility and pipeline information to the coordinate system of the corrected elevation information to generate a facility information layer, and perform topological repair on the broken or overlapping pipe segments in the facility information layer to output a complete list of facility attributes.

[0020] Taking a pilot area of a mountainous small watershed as an example, the corrected elevation information refers to the digital elevation data (DEM) after projection transformation, filtering and smoothing, and ground control point correction, which records the altitude values of each position on the ground surface and usually comes from lidar or total station measurement; For example, the planar coordinates of a certain infiltration pond in the design drawing are (X = 512345.67 m, Y = 341234.56 m), and the designed elevation of the bottom plate is 365.00 m. Align this coordinate with the grid cell in the DEM after calibration, and the surface elevation value of this cell (such as 367.25 m) can be read, and then the burial depth of the infiltration pond bottom plate relative to the terrain can be calculated (367.25 m - 365.00 m = 2.25 m).

[0021] Another example is that a main drainage pipeline runs from node A (X = 512500.00 m, Y = 341200.00 m, elevation 370.10 m) to node B (X = 512600.00 m, Y = 341350.00 m, elevation 368.50 m), with a pipe diameter of 500 mm and a slope of 0.16. After mapping the spatial positions and elevations of these two points to the DEM one by one, an infrastructure information layer containing three-dimensional positions, pipe diameters, and slopes can be generated in the GIS.

[0022] Through this mapping, each infrastructure or pipe section obtains accurate three-dimensional coordinates and physical parameters, providing reliable basic data based on high terrain undulations for subsequent hydraulic simulation, network topology verification, and intelligent scheduling.

[0023] It should be noted that the original elevation data refers to the digital elevation point cloud or grid that has not undergone any processing, and usually comes from the following sources: Airborne LiDAR (Light Detection and Ranging) aerial survey: Over the high-terrain small watershed of the sponge city to be built, an unmanned aerial vehicle or a fixed-wing aircraft is equipped with a LiDAR sensor to conduct contour scanning of the entire watershed (such as a mountainous area within a range of 10 km² with an average slope ≥ 15 degrees), and the three-dimensional coordinates (X, Y, Z) of each point are obtained; Satellite remote sensing (such as SRTM, ALOS) downsampled DEM: In the peripheral area where the airborne LiDAR coverage is insufficient, satellite DEM data with a resolution of 30 m or 12.5 m is used to supplement the boundary area; Ground GNSS static measurement: Dual-frequency GNSS reference stations are deployed at key control points (GCPs) to obtain absolute elevations and correct the aerial point cloud with millimeter-level accuracy; After the acquisition is completed, the airborne LiDAR point cloud is fused with the satellite DEM and ground GNSS control points, and all data is unified into the same projection coordinate system, and the original elevation data covering the entire high-terrain small watershed is output.

[0024] Based on the corrected elevation information, calculate the local slope and curvature at each location in a GIS environment or a dedicated script, and construct a non-linear grid side length function accordingly. The high-relief and steep-slope areas are divided, and the flat areas can be appropriately coarsened. Call the Delaunay triangulation algorithm to automatically generate a series of triangular grid cells with this side length function as the constraint. Each cell is a basic calculation unit, and its geometric center, vertices, and adjacency relationships are all determined by this triangulation process; After the grid generation is completed, the continuous field data constructed in the preprocessing stage, such as elevation, permeability coefficient, soil layer thickness, and vegetation coverage, can be immediately extracted at the central coordinates of each triangular cell and spatially overlaid with the facility information layer.

[0025] Perform equation configuration and numerical solution. That is, based on the multi-dimensional attributes of the grid cells, apply the mass conservation equation and the momentum conservation equation to each triangular cell to describe the processes of rainfall input, surface runoff, and infiltration confluence. First, use the central elevation, permeability coefficient, soil layer thickness, and vegetation coverage of each grid cell as the initial conditions, and use the real-time rainfall intensity and boundary water level observation values as the boundary inputs. Through the finite volume iteration method, calculate the water depth change and flow velocity distribution of each cell at each time step. During the solution process, introduce an infiltration loss term in the form of Darcy's law for the infiltration process to relate the infiltration amount to the permeability coefficient and soil layer thickness; The numerical solution is executed in parallel on the cloud cluster, and the water depth time series and flow velocity time series of each cell are output in real time to provide physical simulation results for neural network training and local prediction.

[0026] Construct a deep feedforward neural network with spatial coordinates and timestamps as inputs. Its number of nodes and layers are set according to the non-linear characteristics of the watershed, and the water depth prediction value at the corresponding moment is generated at the output end of the network; In network training, introduce two parts of losses simultaneously: one is the data fitting loss, which compares the network prediction value with the water depth values observed by each sensor during rainfall events; the other is the physical residual loss, which calculates the degree of non-satisfaction of the equation by substituting the predicted output into the mass conservation and momentum conservation equations, and uses this residual as a penalty term; By adjusting the network weights, the trained network can not only accurately restore the historical observed water depth but also naturally follow the hydrodynamic laws during the prediction process; It should be noted that in the initial training stage, batch training is performed using the global historical data, and then small-batch fine-tuning is carried out on the cloud for extreme rainstorm events to balance generality and local accuracy.

[0027] The physical numerical solution results are compared with the neural network prediction results, and the accuracy of the water depth prediction is evaluated using the root mean square error as an indicator. The permeability loss coefficient in the numerical solution and the physical residual weight in the network training are adjusted according to the error trend until the calibration error converges to the preset range; After calibration, the physical solution program and the neural network model are packaged and deployed to the edge computing node. The node receives rainfall intensity, water level and soil moisture data in real time during the rainstorm. The neural network first quickly gives a short-term water depth prediction, which is then corrected by combining the local numerical solution. Finally, the comprehensive prediction result is pushed to the control unit through the message middleware for intelligent scheduling of gates and pump groups, ensuring both high-precision prediction and low-latency response in the high-altitude sponge city rainwater management scenario.

[0028] Step 2: After completing the spatial discrete grid division and attribute mapping, the generalization adaptation to the diversity of terrain, soil and vegetation is achieved through the coordination of multivariate predictors and the federated learning mechanism. The specific steps are as follows: Four complementary predictors are constructed based on the attributes of each grid cell, such as the central elevation, permeability, soil thickness, vegetation coverage, and historical rainfall time series and corresponding water depth observations. Firstly, the simulation error correction predictor is trained with the difference between the physical simulation results and the sensor observations as the target, using the decision tree regression algorithm to capture the systematic deviation of the physical solution in the local extreme process. Taking the physical equation consistency constraint and water depth observation as double losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the conservation of hydrodynamics. Taking the multi-dimensional attributes of grid cells and historical runoff as input, the attribute association predictor based on the ensemble tree algorithm is trained to mine the nonlinear interactions between soil, vegetation and terrain. Finally, the time-dependent characteristics of rainfall and runoff generation are used as input to train the long short-term memory network to capture the fluctuation trend after the sudden change of rainfall; After each predictor completes batch training for the first time in the cloud, it is sent to each basin edge node for local fine-tuning. The fine-tuning process uses the residual between the current rainfall observation and the batch training result as a new sample for iterative update, so that the model can take into account both global versatility and local scene characteristics; A specific example is as follows: in the multivariate predictor construction stage, targeted training is conducted for each model type based on the unit center elevation, permeability coefficient, soil thickness, vegetation coverage, and measured water depth of historical rainfall; For example, for a certain grid cell (with a central elevation of 350 m, a permeability coefficient of 0.000012 m / s, a soil layer thickness of 2.5 m, and a vegetation coverage of 65%), the hourly sequence rainfall during the rainstorm process on June 15, 2024 is [20, 30, 40, 10] mm / h, and the corresponding observed water depths are [0.10, 0.30, 0.40, 0.20] m; the physical simulation solution outputs [0.08, 0.25, 0.35, 0.15] m, and the residuals between the two are [0.02, 0.05, 0.05, 0.05] m. Using this residual sequence and the grid cell characteristics as samples, a decision tree regression algorithm (with a maximum tree depth of 5 layers and a minimum number of samples per leaf node of 10) is used to train a simulation error correction predictor, enabling it to automatically correct the underestimation bias of the physical solution at the rainfall peak under similar rainstorm conditions.

[0029] On this basis, with the physical simulation results and the observed water depth of the same unit as dual objectives, a physics-informed neural network is designed: the network structure has 5 layers, 64 neurons in each layer, and the ReLU activation function is selected. It is trained by simultaneously minimizing the water depth fitting error and the equation residuals (introducing the residuals of the mass conservation and momentum conservation equations into the loss function) to ensure that the prediction not only conforms to the measured data but also does not violate the hydrodynamic laws. At the same time, using the multi-dimensional attributes of this unit and the cumulative runoff (such as the maximum hourly runoff of 0.15 m corresponding to the above rainstorm) as inputs, an ensemble tree algorithm (100 subtrees, learning rate of 0.1) is used to train an attribute correlation predictor to deeply explore the non-linear interaction of terrain, soil, and vegetation on the runoff; finally, the rainfall time series and the corresponding runoff time series at each moment are input into a long short-term memory network (with 50 hidden nodes and a time step of 4 hours) to capture the dynamic characteristics of the water depth fluctuation after the rainfall mutation, and a time series prediction model for the water depth trend within the next 1-2 hours is constructed.

[0030] According to the historical backtest library, typical rainstorm events such as short-term heavy rainfall, long-term light rainfall, and uneven spatial distribution are divided into several scenario categories, and the key performance indicators of each predictor in each category, such as the peak arrival time, peak height, and fluctuation trend, are statistically analyzed to form a scenario performance matrix; During actual operation, first, according to the current rainfall time series (including the inflection point of the cumulative rainfall curve and the rainfall intensity mutation time) and the terrain section division, the most suitable scenario category is automatically matched, and the predictor with the best performance in this category is extracted from the matrix as the primary model. If the deviation between the water depth curve output by the primary model and the real-time observation data exceeds the preset threshold, the backup model call mechanism is triggered, and the models are sequentially switched among the sub-optimal models until the prediction accuracy is restored.

[0031] After local fine-tuning is completed at each basin edge node, the difference between the local model parameters and the initial model parameters (recorded as parameter difference) is uploaded to the cloud. The cloud aggregates the parameter differences from each node and adopts the majority improvement selection strategy for the same network layer or decision tree branch structure: for the parameter direction that has been improved in most nodes, the update is retained; for the parameters that have changed in only a few nodes, they are regarded as noise suppression and are not adopted. After screening, a unified global update package is formed and sent back to each edge node. After receiving the update package, each edge node applies the difference to the local model to achieve full network synchronization of the model version. In this process, the nodes only exchange parameter differences and do not upload the original observation data. After the global update deployment is completed, each edge node immediately conducts cross-domain backtesting, and runs the new generation model in the original typical short-term heavy rainfall, long-term weak rainfall and spatially uneven distribution scenarios, calculates the root mean square error and maximum absolute error indicators, and compares them with the historical version results. If the error indicator of a certain watershed rebounds or fails to meet expectations, the online performance feedback reporting mechanism is automatically triggered, and the error distribution of the watershed and the real-time observation residuals are uploaded to the cloud. In subsequent federated learning rounds, the cloud will prioritize adjusting the corresponding network layer or decision tree depth based on the feedback error pattern, so as to realize the adaptive evolution of the model to cross-domain changes such as extreme climate, vegetation evolution and construction disturbance. Through continuous iterative federated fine-tuning and cross-domain verification closed loop, the generalization ability and prediction robustness under various working conditions in high-altitude small watersheds are continuously improved.

[0032] It should be noted that the rainstorm time is defined by the continuous period when the real-time rainfall intensity continues to be higher than the preset rainstorm intensity threshold. Specifically, the rainfall intensity time series is first filtered by the threshold, and the moment when it first breaks the threshold is determined as the beginning of the rainstorm; then the moment when the rainfall intensity drops below the threshold and remains stable and no longer rises is monitored as the end of the rainstorm; the interval between the two is the rainstorm time. In order to avoid misjudgment caused by short-term pulsation, it is also required that the rainfall intensity continues to be lower than the threshold for the minimum duration set by experts before the rainstorm can be considered over.

[0033] Step 3: Perform online edge fine-tuning and uncertainty quantification. That is, after completing multi-model integration and federated updates, perform local incremental fine-tuning at the edge node and generate prediction uncertainty to improve the model's rapid response to the time-varying characteristics of the environment and the risk protection capability of scheduling decisions. The specific steps are as follows: Perform online incremental fine-tuning. After each heavy rain event ends, automatically summarize the difference between the true water depth sequence recorded by local sensors and the predicted water depth sequence output by the integrated predictor, and construct a residual dataset containing timestamps and corresponding residuals. For the latest several time period data in this residual dataset, select the nearest 50 residuals in descending order as fine-tuning samples to ensure that the samples used can fully reflect the model offset caused by the current soil moisture content, vegetation cover change, and sudden drainage strategy adjustment; After the sample preparation is completed, load the physics-informed neural network model at the edge node, and set up an iterative training process with the least squares residual sum as the loss function. Use parameters of learning rate 0.005 and batch size 10 for 10 rounds of backpropagation optimization. During this process, only update the parameters of the feature extraction layer and output layer in the network related to time-dependent features and spatial attribute mapping, which ensures the fine-tuning speed and avoids the computational overhead caused by full-scale retraining; After the fine-tuning is completed, save the snapshot of the updated model parameters and the latest residual statistics together, immediately replace the old version of the model, and immediately use it for the water depth prediction in the next time period to support the high-precision water depth prediction in subsequent time periods.

[0034] After the fine-tuning is over, the system conducts self-check on the new model at the edge node: use the latest 50 observation inputs, recalculate the prediction output and compare it with the true value. If the root mean square error drops by at least 10% compared with that before fine-tuning, it is confirmed that this online iteration is effective; otherwise, return to the previous effective parameter state and report it to the cloud.

[0035] Perform uncertainty quantification, that is, immediately enter the uncertainty quantification link after the online fine-tuning is completed. For the input data at the same moment, perform forward inference by randomly masking some connection nodes in the network multiple times (for example, 30 times) to simulate the output fluctuations of the model under different internal states; Each inference generates a set of water depth prediction values. After summarizing these 30 groups of prediction results, calculate their arithmetic mean and standard deviation, which are used to characterize the central tendency and dispersion degree of the model under the current environmental conditions. Based on the mean and standard deviation, derive the upper and lower limits of the 95% confidence interval to reflect the uncertain range of the prediction; After obtaining the confidence interval, compare it with the preset flood control and infiltration thresholds: When the upper confidence limit approaches or exceeds the flood control threshold, automatically reserve more storage space and activate the emergency diversion strategy in advance; When the lower confidence limit is far lower than the infiltration requirement, the operating power of the pump group can be appropriately reduced to promote natural infiltration; If the confidence interval as a whole falls within the safe interval, perform regular scheduling according to the central tendency value; All uncertainty quantification results and their corresponding input and output data will be reported to the cloud in the form of structured messages for subsequent performance feedback and model improvement of federated learning, enabling closed-loop optimization and risk control among the edge, cloud, and end devices.

[0036] It should be noted that the flood control threshold, infiltration requirement, and safety interval are determined based on actual measurements or expert opinions, and can also be numerically adjusted according to the actual situation.

[0037] Step 4: Perform risk-aware scheduling and edge-cloud-end collaboration. That is, after completing online fine-tuning and uncertainty quantification, enter the scheduling decision-making stage. Combine the predicted confidence interval with the flood control and infiltration requirement thresholds to achieve risk-aware intelligent scheduling. The specific steps are as follows: After completing online fine-tuning and uncertainty quantification, the edge node first compares the upper and lower limits of the predicted confidence interval of each triangular grid cell with the preset floodwater depth threshold and minimum infiltration threshold one by one, and classifies the cells into three categories: high risk, low risk, and safe.

[0038] For the grid cells determined to be at high risk, whose upper confidence interval limit exceeds the floodwater depth threshold, prioritize the recommendation to increase the usage ratio of the storage pond, and generate preliminary control instructions for the opening range of the gate of this cell (such as 60% - 80%) and the available power range of the pump group (such as 30% - 50%); For the cells determined to be at low risk, whose lower confidence interval limit is lower than the minimum infiltration threshold, propose the recommendation to prioritize retaining the capacity of the infiltration pond, set the lower limit of the pump group power to 0% and the upper limit to the lowest acceptable operating power (such as 10%) to promote natural infiltration; For the cells in the safe interval, generate the default opening and power ranges of the gate and pump group according to the conventional scheduling logic. This preliminary scheduling result forms an edge scheduling table in units of cells, including the risk level, recommended gate opening range, recommended pump group power range, and corresponding trigger conditions of each cell, and is generated and locally stored at the edge node in seconds for global optimization and invocation by the cloud; After receiving the edge scheduling tables and the corresponding confidence interval reports uploaded by all edge nodes, the cloud first aggregates the scheduling suggestions for each grid cell to construct a global water balance equation system to ensure that the total inflow, total infiltration, and total discharge are within the allowable range. At the same time, it adopts a multi-scenario simulation method to separately simulate the water level response and pump energy consumption performance after the execution of each edge suggestion combination. Through parallel computing, it screens out a complete scheduling plan that meets the requirements of waterlogging prevention and control and optimal energy consumption, clarifies the usage order of each storage pond, the opening curve of each gate in the future time period (such as every 5 minutes), and the time-segment power setting of the pump group, and includes the risk evolution prediction and emergency switching logic of each unit. After the optimization is completed, the cloud issues a global scheduling package containing the global optimal scheduling time sequence, risk level spectrum, and operation instructions to each edge node, ensuring that the edge can execute according to a unified standard, safely and efficiently after receiving the instruction, forming a closed-loop collaboration among the cloud, edge, and terminal.

[0039] The operation and maintenance platform overlays the risk level spectrum color block diagram in the GIS interface, marks each triangular grid cell with warning red, warning yellow, and safety green for high risk, medium risk, and safety levels respectively, and at the same time displays the predicted confidence interval, recommended gate opening interval, and pump group power interval of the unit in the form of a floating panel above each grid cell. It pushes the control instructions containing the precise opening curve and power setting time sequence of the corresponding grid cell in the global scheduling package to their respective edge nodes through a secure channel; After each edge node finishes the execution, it continuously monitors and records the actual water depth and energy consumption data, compares them with the recommended values in the global scheduling package, calculates the water depth residual and energy consumption deviation, and generates an execution effect report. This report includes the residual distribution diagram, deviation statistical value, and a list of units exceeding the tolerance threshold (such as ±5%). It is uploaded to the cloud in real time. After the cloud aggregates the reports of each node, it compares the historical performance with the current round of feedback to form a new performance feedback report, and feeds back the typical patterns of water depth error, high or low energy consumption in the report to the federated learning system as the key improvement objects in the next round of distributed federated learning collaborative update and cross-domain verification stage. Through the end-edge-cloud closed-loop mechanism, the model generalization ability and scheduling accuracy are continuously improved in several rounds of iteration, and dynamic adaptability to sudden climate and on-site disturbances is realized in actual operation, ensuring the long-term stable operation of the high-elevation rainwater management system in the sponge city.

[0040] The present invention conducts systematic data collection and preprocessing on high-elevation areas, fuses the borehole permeability coefficient, soil layer thickness, and remote sensing vegetation coverage interpolation to generate a continuous physical field, maps the infiltration pond, storage pond, and pipeline information to form a facility information layer, and then divides grid cells; constructs a multi-variable predictor based on the center elevation and physical properties of the grid cells, trains it in the cloud and then sends it to the edge node for local incremental fine-tuning, and dynamically matches and synchronously optimizes the model relying on the scenario performance matrix and the federated learning mechanism; automatically extracts the observed prediction residuals after each rainstorm event for online incremental iteration of the neural network, and generates a confidence interval through random masking forward inference; the edge node divides the risk levels according to the confidence interval, flood control water depth, and minimum infiltration threshold, and outputs suggestions for the gate opening and pump group power, thereby significantly improving the prediction accuracy and response speed of extreme rainstorm peaks in small high-elevation watersheds, realizing refined risk perception scheduling, and enhancing the effects of rainstorm prevention and rainwater resource utilization.

[0041] Embodiment 2: A rainwater management system for high-elevation areas in a sponge city, as Figure 2 shown, specifically including: A high-elevation network division module, which is used to collect and preprocess data in high-elevation areas, combine borehole permeability coefficient, soil layer thickness, and remote sensing vegetation coverage interpolation, generate a continuous physical field after removing anomalies, map the facility and pipeline information to the elevation coordinate system, perform topological repair, and then output the facility information layer and the facility attribute list, and divide grid cells; A grid fine-tuning prediction module, which is used to construct a predictor based on the attributes of the center of the grid cell, train it in the cloud and then send it to the edge node for local incremental fine-tuning, dynamically match the primary predictor according to the scenario performance matrix, and synchronously update the local model prediction by aggregating through federated learning to generate a global update package; A prediction model adjustment module, which is used to, after each rainstorm event, the edge node automatically extracts the latest observed and predicted residuals and uses them as training samples to perform incremental iteration on the neural network. After fine-tuning, a random masking mechanism is used for forward inference, and the central tendency and fluctuation range of the predicted values are statistically analyzed to generate a confidence interval of the model; A management suggestion generation module, which is used for the edge node to divide high-risk, low-risk, and safe units according to the prediction confidence interval, flood control water depth threshold, and minimum infiltration threshold, and generate suggestions for the gate opening and pump group power.

[0042] The above formulas are all dimensionless and take their numerical calculations. Specifically, various means such as standardization can be used for dimensionless processing, which will not be elaborated here. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain a formula that is closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0043] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center in a wired or wireless (such as infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that contains one or more collections of available media. The available media can be magnetic media (such as floppy disks, ATA hard disks, magnetic tapes), optical media (such as DVDs), or semiconductor media. The semiconductor media can be a solid-state ATA hard disk.

[0044] It should be understood that in various embodiments of the present application, the sequence numbers of the above processes do not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0045] Those of ordinary skill in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professionals can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0046] In several embodiments provided in the present application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some interfaces. The indirect couplings or communication connections of the devices or units can be in an electrical, mechanical, or other form.

[0047] The unit described as a separation component may or may not be physically separated. The component shown as a unit may or may not be a physical unit, and it may be located in one place or distributed across multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0048] In addition, each functional unit in various embodiments of the present application may be integrated into one processing unit, or each unit may exist physically alone, or two or more units may be integrated into one unit.

[0049] As described above, the above are only specific implementation manners of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A rainwater management method for sponge cities with high terrain, characterized in that, It includes the following steps: Collect and preprocess data for the high terrain area, interpolate with borehole permeability coefficient, soil layer thickness and remote sensing vegetation coverage, generate a continuous physical field after removing anomalies, map the facility and pipeline information to the elevation coordinate system, output the facility information layer and the facility attribute list after topological repair, and divide grid cells; Build a predictor based on the attributes at the center of the grid cells, train it in the cloud and then send it to the edge node for local incremental fine-tuning, dynamically match the primary predictor according to the scenario performance matrix, and generate a global update package through federated learning to synchronously update the local model prediction; After each rainstorm event, the edge node automatically extracts the latest observations and prediction residuals and uses them as training samples to perform incremental iteration on the neural network. After fine-tuning, use the random masking mechanism for forward inference, statistically analyze the central tendency and fluctuation range of the predicted values, and generate the confidence interval of the model; The edge node divides high-risk, low-risk and safe units according to the prediction confidence interval, flood control water depth threshold and minimum infiltration threshold, and generates suggestions for the gate opening and pump group power.

2. The method for managing rainwater in high-lying areas of a sponge city according to claim 1, characterized in that: Collect and preprocess data for the high terrain area, interpolate with borehole permeability coefficient, soil layer thickness and remote sensing vegetation coverage, generate a continuous physical field after removing anomalies. The specific process is as follows: Obtain the original elevation data of the high terrain area, perform projection transformation to unify it to the same coordinate system, apply Gaussian smoothing filter to the projected elevation data to remove measurement noise, form smooth elevation information, and then perform polynomial error correction on the smooth elevation information according to the ground control points, and output the corrected elevation information; Interpolate soil and vegetation parameters for the high terrain area, obtain soil permeability coefficient data and soil layer thickness data from borehole sampling, and obtain vegetation coverage data from remote sensing classification; Apply the Kriging interpolation method to the soil permeability coefficient, soil layer thickness and vegetation coverage data respectively to generate continuous permeability coefficient fields, soil layer thickness fields and vegetation coverage fields; Remove outliers from the interpolated permeability coefficient field, soil layer thickness field and vegetation coverage field, and interpolate again to fill the holes.

3. A method for managing rainwater at high elevations in a sponge city according to claim 2, characterized in that: Map the facility and pipeline information to the elevation coordinate system, output the facility information layer and the facility attribute list after topological repair, and divide grid cells. The specific process is as follows: Extract facility and pipeline information from the regional design drawing, including the spatial positions and structural parameters of infiltration ponds, storage ponds and main drainage pipelines; Map the extracted facility and pipeline information to the coordinate system of the corrected elevation information to generate a facility information layer, perform topological repair on the broken or overlapping pipe segments in the facility information layer, and output a complete facility attribute list; Based on the corrected elevation information, calculate the local slope and curvature at each location in the GIS environment, call the Delaunay triangulation algorithm, and automatically generate a series of triangular grid cells with the side length function as the constraint; Take the central elevation, permeability coefficient, soil layer thickness and vegetation coverage of each grid cell as the initial conditions, take the real-time rainfall intensity and boundary water level observation values as the boundary inputs, and calculate the water depth change and flow velocity distribution of each unit at each time step through the finite volume iteration method.

4. The sponge city high-land rainwater management method according to claim 3, characterized in that: The predictor is built based on the attributes of the grid cell center, and after training in the cloud, it is sent to the edge node for local incremental fine-tuning. The specific steps are as follows: Four complementary predictors are constructed based on the central elevation, permeability, soil thickness, vegetation coverage, historical rainfall time series and corresponding water depth observation attributes of each grid cell. The predictor includes a simulation error correction predictor trained using a decision tree regression algorithm with the difference between the physical simulation result and the sensor observation as the target; Taking the physical equation consistency constraint and water depth observation as double losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the conservation of hydrodynamics. Taking the multi-dimensional attributes of grid cells and historical runoff as input, the attribute association predictor based on the ensemble tree algorithm is trained to mine the nonlinear interactions between soil, vegetation and terrain. Taking the time-dependent characteristics of rainfall and runoff as input, the long short-term memory network is trained to capture the fluctuation trend after the sudden change of rainfall; After each predictor completes batch training for the first time in the cloud, it is sent to the edge nodes of each basin for local fine-tuning. The fine-tuning process uses the residual between the current rainfall observation and the batch training results as a new sample for iterative update.

5. A method for managing rainwater in high-lying areas of sponge cities according to claim 4, characterized in that: Dynamically match the main predictor according to the scenario performance matrix, and generate a global update package through federated learning aggregation to synchronously update the local model prediction. The specific process is as follows: According to the historical backtesting database, short-term heavy rainfall, long-term weak rainfall and typical rainstorm events with uneven spatial distribution are divided into different scenario categories, and the peak arrival time, peak height and fluctuation trend performance indicators of each predictor under each category are statistically analyzed to form a scenario performance matrix; According to the current rainfall time series and terrain segment division, the most suitable scenario category is automatically matched, and the best predictor under the scenario category is extracted from the matrix as the main model. If the water depth curve output by the main model deviates from the real-time observation data beyond the preset threshold, the backup model calling mechanism is triggered, and the suboptimal models are switched in sequence until the prediction accuracy is restored; The rainfall time series includes the inflection point of the cumulative rainfall curve and the time of sudden change of rainfall intensity; After completing local fine-tuning at each watershed edge node, the difference between the local model parameters and the initial model parameters is uploaded to the cloud, which aggregates the parameter differences from each node; For the same network layer or decision tree branch structure, the majority improvement selection strategy is adopted; After screening, a unified global update package is formed and sent back to each edge node. After receiving the update package, each edge node applies the local model for prediction; After the global update deployment is completed, each edge node immediately conducts cross-domain backtesting, running the new generation model in the original typical short-term heavy rainfall, long-term weak rainfall and spatially uneven distribution scenarios, calculating the root mean square error and maximum absolute error indicators, and comparing them with the historical version results; If the error index of a watershed rises or falls short of expectations, the online performance feedback reporting mechanism will be automatically triggered, and the error distribution of the watershed and the real-time observation residuals will be uploaded to the cloud.

6. The sponge city high-land rainwater management method according to claim 5, characterized in that: After each rainstorm event, the edge nodes automatically extract the latest observation and prediction residuals and use them as training samples to perform incremental iterations on the neural network. The specific process is as follows: After each rainstorm event ends, automatically summarize the difference between the true water depth sequence recorded by local sensors and the predicted water depth sequence output by the integrated predictor, construct a residual data set containing timestamps and corresponding residuals, and for the data in the latest time period in the residual data set, select the 50 nearest residuals in descending order as fine-tuning samples; After the sample preparation is completed, load the physics-informed neural network model on the edge node, and set up an iterative training process with the least squares residual sum as the loss function. Perform 10 rounds of backpropagation optimization with parameters of a learning rate of 0.005 and a batch size of 10; After fine-tuning, save the updated model parameter snapshot and the latest residual statistics together, replace the old version of the model, and immediately use it for the water depth prediction in the next time period.

7. A method for managing rainwater in high-lying areas of a sponge city according to claim 6, characterized in that: After fine-tuning, use a random masking mechanism for forward inference, statistically analyze the central tendency and fluctuation range of the predicted values, and generate the confidence interval of the model. The specific process is as follows: After fine-tuning, the system performs self-check on the new model at the edge node: use the latest 50 observation inputs, recalculate the predicted output and compare it with the true value. If the root mean square error drops by more than or equal to 10% compared with before fine-tuning, confirm that this online iteration is effective; otherwise, return to the previous valid parameter state and report to the cloud; After completing online fine-tuning, enter the uncertainty quantification link. For the input data at the same moment, randomly mask some connection nodes in the network to perform forward inference to simulate the output fluctuations of the model under different internal states; Each inference generates a set of water depth predicted values. After summarizing the prediction results, calculate the arithmetic mean and standard deviation, and based on the mean and standard deviation, derive the upper and lower limits of the confidence interval; After obtaining the confidence interval, compare the confidence interval with the preset flood control and infiltration thresholds: When the upper confidence limit approaches or exceeds the flood control threshold, automatically reserve more storage space and activate the emergency diversion strategy in advance; When the lower confidence limit is much lower than the infiltration requirement, reduce the operating power of the pump group for natural infiltration; If the confidence interval as a whole falls within the safe interval, perform normal scheduling.

8. A method for managing rainwater at high elevations in a sponge city according to claim 7, characterized in that: The edge node divides high-risk, low-risk, and safe units according to the prediction confidence interval, flood control water depth threshold, and minimum infiltration threshold, and generates suggestions for the gate opening and pump group power. The specific process is as follows: Compare the upper and lower limits of the prediction confidence interval of each triangular grid unit with the preset flood control water depth threshold and minimum infiltration threshold one by one, and classify the units into high-risk, low-risk, and safe categories; When the upper limit of the confidence interval exceeds the flood control water depth threshold, it is determined as a high-risk grid unit, and the use ratio of the storage tank is preferentially increased; When the lower limit of the confidence interval is lower than the minimum infiltration threshold, it is determined as a low-risk grid unit, and the capacity suggestion of the infiltration pond is preferentially reserved for natural infiltration treatment; For the grid units in the safe interval, generate the default opening degree and power interval of the gate and pump group according to the normal scheduling logic.

9. A sponge city high-elevation rainwater management system for implementing the sponge city high-elevation rainwater management method described in any one of claims 1-8, characterized in that, Including: The high terrain network division module is used to collect and preprocess data in high terrain areas, interpolate the borehole permeability coefficient, soil layer thickness, and remote sensing vegetation coverage, generate a continuous physical field after removing anomalies, map the facility and pipeline information to the elevation coordinate system, output the facility information layer and the facility attribute list after topological repair, and divide the grid cells; The grid fine-tuning prediction module is used to construct a predictor based on the attributes at the center of the grid cells, perform incremental fine-tuning on the edge nodes locally after training in the cloud, dynamically match the primary predictor according to the scenario performance matrix, and synchronously update the local model prediction by aggregating to generate a global update package through federated learning; The prediction model adjustment module is used to automatically extract the latest observations and prediction residuals at the edge nodes after each rainstorm event and use them as training samples to perform incremental iteration on the neural network. After fine-tuning, a random masking mechanism is used for forward inference, and the central tendency and fluctuation range of the predicted values are statistically calculated to generate the confidence interval of the model; The management suggestion generation module is used for the edge nodes to divide high-risk, low-risk, and safe units according to the prediction confidence interval, flood control water depth threshold, and minimum infiltration threshold, and generate suggestions for the gate opening and pump group power.

Images