A method and system for managing rainwater in high-altitude areas of sponge cities

Abstract

The present invention discloses a method and system for managing rainwater in high-altitude sponge cities, which specifically relates to the technical field of rainwater prediction and management, and is used to solve the problem of poor scheduling of rainwater prediction and management in high-altitude areas. The present invention targets small watersheds in high altitude areas, first collects and preprocesses data such as elevation, permeability, soil thickness, and vegetation coverage, generates a continuous physical field, maps the facility network, and then divides the grid units. A multivariate predictor is constructed based on the central attributes of the grid units, and after cloud-based training, incremental fine-tuning is sent to edge nodes, and the model is optimized through scene matching and federated learning. After a rainstorm, the observation and prediction residuals are automatically extracted to incrementally train a neural network and generate confidence intervals. The edge nodes divide risks according to the confidence intervals, flood control depth, and minimum infiltration thresholds, and output gate opening and pump group power recommendations, thereby significantly improving the prediction accuracy and response time of extreme rainstorm peaks in small watersheds in high altitude areas, and realizing refined risk perception scheduling and rainwater resource utilization.

Description

Technical Field

[0001] The present invention relates to the technical field of rainwater prediction and management, and more specifically, to a method and system for managing rainwater in high-altitude areas of a sponge city. Background Art

[0002] Sponge City is an ecological urban development model that simulates the natural hydrological cycle to enhance 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 stored water resources during droughts, thereby alleviating problems such as urban waterlogging, drought, and water pollution, and achieving sustainable management of urban water resources.

[0003] Deficiencies in existing technologies: In the management of extreme rainstorms in high-altitude small watersheds in sponge cities, most of the methods rely on single statistical rainfall and runoff models or pure data-driven methods, ignoring the amplifying effect of the sharp fluctuations in high terrain on the timing and intensity of rainstorm runoff peaks, and failing to incorporate hydrodynamic conservation relationships into model constraints. As a result, systematic prediction errors such as late peaks or low intensity often occur under conditions of short-term heavy rainfall. For example, a small mountain watershed suddenly received 90 mm of rainfall within one hour, but the traditional model only predicted a peak of 60 mm with a delay of 20 minutes, resulting in the failure of the regulating reservoir to open the gates to discharge floodwaters in time, seriously restricting the refined effects of waterlogging prevention and control and groundwater recharge in high-altitude areas, and affecting the city's water cycle scheduling and management. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for managing high-altitude rainwater in a sponge city, so as to solve the problem of poor prediction, management and scheduling of high-altitude rainwater in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for managing rainwater in high-altitude areas of a sponge city, comprising the following steps:

[0007] Data collection and preprocessing are performed on high-relief areas. The data is interpolated using the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage. Anomalies are removed to generate a continuous physical field. Facility and pipeline information is mapped to an elevation coordinate system. After topological repair, a facility information layer and a facility attribute list are output, and grid cells are divided.

[0008] A predictor is built based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The primary predictor is dynamically matched based on the scenario performance matrix. A global update package is generated through federated learning aggregation and then the local model prediction is synchronously updated.

[0009] After each rainstorm event, edge nodes automatically extract the latest observations and forecast residuals and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to statistically analyze the central tendency and fluctuation range of the predicted values and generate confidence intervals for the model.

[0010] The edge node divides high-risk, low-risk and safe units according to the prediction confidence interval, flood control depth threshold and minimum infiltration threshold, and generates gate opening and pump group power recommendations.

[0011] In a preferred embodiment, data collection and preprocessing are performed on high-relief areas, and the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage are interpolated to remove anomalies and generate a continuous physical field. The specific process is as follows:

[0012] Obtain the original elevation data of the high-relief area, perform projection transformation, unify it into the same coordinate system, apply Gaussian smoothing filtering to the projected elevation data to remove measurement noise and form smoothed elevation information, then perform polynomial error correction on the smoothed elevation information based on the ground control points and output the corrected elevation information;

[0013] Soil and vegetation parameters are interpolated for high-relief areas, soil permeability coefficient data and soil layer thickness data are obtained from borehole sampling, and vegetation coverage data are obtained from remote sensing classification;

[0014] Kriging interpolation is applied to the soil permeability coefficient, soil thickness and vegetation coverage data respectively to generate continuous permeability coefficient field, soil thickness field and vegetation coverage field;

[0015] The outliers in the interpolated permeability coefficient field, soil thickness field and vegetation coverage field are removed, and the gaps are filled by interpolation again.

[0016] In a preferred embodiment, the facility and pipeline information is mapped to the elevation coordinate system, and after topology repair, the facility information layer and facility attribute list are output and the grid units are divided. The specific process is as follows:

[0017] Extract facility and pipeline information from regional design drawings, including the spatial location and structural parameters of infiltration tanks, storage tanks, and main drainage pipelines;

[0018] 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 broken or overlapping pipe segments in the facility information layer and output a complete facility attribute list.

[0019] Based on the corrected elevation information, the local slope and curvature are calculated for each location in a GIS environment or a dedicated script, and the Delaunay triangulation algorithm is called to automatically generate a series of triangular mesh units with the edge length function as a constraint;

[0020] The center elevation, permeability, soil thickness and vegetation coverage of each grid cell are used as initial conditions, and the real-time rainfall intensity and boundary water level observations are used as boundary inputs. The finite volume iteration method is used to calculate the water depth change and flow velocity distribution of each cell in each time step.

[0021] In a preferred embodiment, a predictor is constructed 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:

[0022] Four complementary predictors are constructed based on the central elevation, hydraulic conductivity, soil thickness, vegetation coverage, historical rainfall time series, and corresponding water depth observations of each grid cell.

[0023] The predictor includes a simulation error correction predictor trained using a decision tree regression algorithm with the difference between the physical simulation results and the sensor observation as the target;

[0024] Using the physical equation consistency constraint and water depth observation as dual losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the hydrodynamic conservation.

[0025] Taking the multidimensional attributes of grid cells and historical runoff as input, an attribute association predictor based on an ensemble tree algorithm is trained to exploit the nonlinear interactions between soil, vegetation, and topography.

[0026] Taking the time-dependent characteristics of rainfall and runoff as input, a long short-term memory network is trained to capture the fluctuation trend after a sudden change in rainfall.

[0027] After each predictor completes batch training for the first time in the cloud, it is sent to the edge nodes of each watershed 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.

[0028] In a preferred embodiment, the main predictor is dynamically matched according to the scenario performance matrix, and the local model prediction is updated synchronously after a global update package is generated through federated learning aggregation. The specific process is as follows:

[0029] Based on the historical backtesting database, short-duration heavy rainfall, long-duration weak rainfall, and typical rainstorm events with uneven spatial distribution are divided into different scenario categories. The peak arrival time, peak height, and fluctuation trend performance indicators of each predictor in each category are statistically analyzed to form a scenario performance matrix.

[0030] Based on the current rainfall time series and terrain segmentation, the most suitable scenario category is automatically matched, and the best predictor for that scenario category is extracted from the matrix as the primary model. If the water depth curve output by the primary model deviates from the real-time observation data by more than a preset threshold, the backup model call mechanism is triggered, switching between the suboptimal models in sequence until the prediction accuracy is restored;

[0031] The rainfall time series includes the inflection point of the cumulative rainfall curve and the time of sudden change of rainfall intensity;

[0032] 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.

[0033] For the same network layer or decision tree branch structure, a majority improvement rejection strategy is adopted: for parameter directions that have improved in most nodes, the update is retained; for parameters that have changed in some nodes, they are considered as noise suppression and not adopted;

[0034] 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;

[0035] 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.

[0036] If the error indicators of a watershed rebound or fall 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.

[0037] In a preferred embodiment, after each rainstorm event, the edge node automatically extracts the latest observation and prediction residuals and uses them as training samples to perform incremental iterations on the neural network. The specific process is as follows:

[0038] After each round of heavy rain, the difference between the actual water depth series recorded by local sensors and the predicted water depth series output by the integrated predictor is automatically summarized to construct a residual dataset containing timestamps and corresponding residuals. For the latest period of data in the residual dataset, the most recent 50 residuals are selected in descending order as fine-tuning samples.

[0039] After the sample preparation is completed, the physical information enhanced 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. 10 rounds of backpropagation optimization are performed with parameters of 0.005 learning rate and 10 batch size.

[0040] After fine-tuning is completed, the updated model parameter snapshot is saved together with the latest residual statistics, and the old model is replaced and immediately used for water depth prediction in the next period.

[0041] In a preferred embodiment, after fine-tuning is completed, a random masking mechanism is used for forward reasoning to calculate the central tendency and fluctuation range of the predicted value and generate the confidence interval of the model. The specific process is as follows:

[0042] After fine-tuning, the system performs a self-check on the new model at the edge node: using the latest 50 observations, the predicted output is recalculated and compared with the true value. If the root mean square error decreases by 10% or more compared to before fine-tuning, the online iteration is considered valid. Otherwise, the system returns to the last valid parameter state and reports to the cloud.

[0043] After completing online fine-tuning, the uncertainty quantification phase begins. For the input data at the same moment, some connected nodes in the network are randomly blocked and forward reasoning is performed to simulate the output fluctuations of the model under different internal states.

[0044] Each inference generates a set of water depth prediction values. After summarizing the prediction results, the arithmetic mean and standard deviation are calculated. Based on the mean and standard deviation, the upper and lower limits of the confidence interval are derived.

[0045] After obtaining the confidence interval, compare the confidence interval with the preset flood control and infiltration thresholds:

[0046] When the confidence upper limit approaches or exceeds the flood control threshold, more storage space is automatically reserved and the emergency diversion strategy is activated in advance;

[0047] When the confidence lower limit is far lower than the infiltration requirement, the operating power of the pump unit is reduced and natural infiltration is carried out;

[0048] If the confidence interval as a whole falls within the safe interval, normal scheduling is performed.

[0049] In a preferred embodiment, the edge node divides high-risk, low-risk, and safe units based on the prediction confidence interval, flood control depth threshold, and minimum infiltration threshold, and generates gate opening and pump group power recommendations. The specific process is as follows:

[0050] The upper and lower limits of the predicted confidence interval of each triangular grid cell are compared one by one with the preset flood control depth threshold and minimum infiltration threshold, and the cells are divided into three categories: high risk, low risk and safe;

[0051] When the upper limit of the confidence interval exceeds the flood control depth threshold, the grid cell is determined to be high-risk and priority is given to expanding the proportion of storage ponds used;

[0052] When the lower limit of the confidence interval is lower than the minimum infiltration threshold, the grid cell is judged to be low-risk, and the infiltration pond capacity recommendation is prioritized for natural infiltration treatment;

[0053] For grid cells in the safety zone, the default opening degree and power range of the gate and pump group are generated according to the conventional scheduling logic.

[0054] A sponge city high-altitude rainwater management system, used to implement the above-mentioned sponge city high-altitude rainwater management method, comprising:

[0055] The high-relief network partitioning module is used to collect and preprocess data in high-relief areas. It combines the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage interpolation to eliminate anomalies and generate a continuous physical field. It also maps facility and pipeline information to the elevation coordinate system. After topological repair, it outputs a facility information layer and a facility attribute list to divide the grid cells.

[0056] The grid fine-tuning prediction module is used to build a predictor based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The main predictor is dynamically matched according to the scenario performance matrix. After generating a global update package through federated learning aggregation, the local model prediction is synchronously updated.

[0057] The prediction model adjustment module is used to automatically extract the latest observations and prediction residuals from edge nodes after each rainstorm event, and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to statistically analyze the central tendency and fluctuation range of the predicted values and generate the confidence interval of the model.

[0058] The management suggestion generation module is used to divide edge nodes into 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.

[0059] Technical effects and advantages of the present invention:

[0060] The present invention systematically collects and preprocesses data in high-altitude areas, integrates borehole permeability coefficient, soil thickness and remote sensing vegetation cover interpolation to generate a continuous physical field, maps infiltration pool, storage pool 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 federated learning mechanism to dynamically match and synchronize the optimization model; after each rainstorm event, the observation prediction residuals are automatically extracted for online incremental iteration of the neural network, and confidence intervals are generated 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 in high altitude areas, realizing refined risk perception scheduling, and enhancing the effects of rainstorm prevention and control and rainwater resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of a high-altitude rainwater management method for a sponge city according to the present invention.

[0062] Figure 2 This is a structural schematic diagram of a sponge city high-altitude rainwater management system according to the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0064] Example 1: Figure 1 As shown, a method for managing rainwater in high-altitude areas of a sponge city comprises the following steps:

[0065] Data collection and preprocessing are performed on high-relief areas. The data is interpolated using the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage. Anomalies are removed to generate a continuous physical field. Facility and pipeline information is mapped to an elevation coordinate system. After topological repair, a facility information layer and a facility attribute list are output, and grid cells are divided.

[0066] A predictor is built based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The primary predictor is dynamically matched based on the scenario performance matrix. A global update package is generated through federated learning aggregation and then the local model prediction is synchronously updated.

[0067] After each rainstorm event, edge nodes automatically extract the latest observations and forecast residuals and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to statistically analyze the central tendency and fluctuation range of the predicted values and generate confidence intervals for the model.

[0068] The edge node divides high-risk, low-risk and safe units according to the prediction confidence interval, flood control depth threshold and minimum infiltration threshold, and generates gate opening and pump group power recommendations.

[0069] Step 1: Build digital twin data enhanced with physical information in high terrain. The specific steps are as follows:

[0070] Perform data collection and preprocessing to obtain the original elevation data of the high-relief area, perform projection transformation, unify it into the same coordinate system, apply Gaussian smoothing filtering to the projected elevation data to remove measurement noise and form smoothed elevation information. Then, perform polynomial error correction on the smoothed elevation information based on the ground control points and output the corrected elevation information.

[0071] 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;

[0072] Kriging interpolation is applied to the soil permeability coefficient, soil thickness and vegetation coverage data respectively to generate continuous permeability coefficient field, soil thickness field and vegetation coverage field;

[0073] The outliers in the interpolated permeability coefficient field, soil thickness field and vegetation coverage field are removed, and the gaps are filled by interpolation again to ensure that each data field is smooth and has no isolated anomalies;

[0074] Facility and pipeline network information is integrated, and facility and pipeline information is extracted from the regional design map, including the spatial location and structural parameters of the infiltration tank, storage tank, and main drainage pipeline. This extracted facility and pipeline information is mapped to the coordinate system of the corrected elevation information to generate a facility information layer. Broken or overlapping pipe sections in the facility information layer are topologically repaired, and a complete facility attribute list is output.

[0075] Taking a small watershed pilot area in a mountainous area as an example, the calibrated elevation information refers to the digital elevation data (DEM) that has been projected, smoothed, and calibrated with ground control points. It records the altitude value of each location on the surface and is usually obtained from LiDAR or total station measurements.

[0076] For example, the plane coordinates of an infiltration pool in the design drawing are (X=512345.67m, Y=341234.56m), and the design elevation of the base plate is 365.00m. By aligning the coordinates with the grid cell in the corrected DEM, the surface elevation value of the cell can be read (such as 367.25m), and then the buried depth of the infiltration pool base plate relative to the terrain can be calculated (367.25m-365.00m=2.25m).

[0077] For example, a section of main drainage pipe runs from node A (X=512500.00m, Y=341200.00m, elevation 370.10m) to node B (X=512600.00m, Y=341350.00m, elevation 368.50m), with a pipe diameter of 500mm and a slope of 0.16. After mapping the spatial positions and elevations of these two points to the DEM one by one, a facility information layer containing the three-dimensional position, pipe diameter and slope can be generated in GIS.

[0078] Through this mapping, each facility or pipeline section obtains precise 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.

[0079] It should be noted that raw elevation data refers to digital elevation point clouds or grids that have not been processed in any way, and usually comes from the following sources:

[0080] Airborne LiDAR (laser radar) aerial survey: Over a small, high-lying watershed in the proposed sponge city, a drone or fixed-wing aircraft equipped with a LiDAR sensor will conduct a contour scan of the entire watershed (e.g., a mountainous area within a 10 km² area with an average slope of ≥15 degrees) to obtain the three-dimensional coordinates (X, Y, Z) of each point.

[0081] Satellite remote sensing (such as SRTM, ALOS) downsampled DEM: In the peripheral areas where airborne LiDAR coverage is insufficient, satellite DEM data with a resolution of 30m or 12.5m is used to supplement the boundary areas;

[0082] Ground GNSS static measurement: Deploy dual-frequency GNSS base stations at key control points (GCPs) to obtain absolute elevation and correct aerial survey point clouds with millimeter-level accuracy.

[0083] After the acquisition is completed, the airborne LiDAR point cloud is fused with the satellite DEM and ground GNSS control points, all data are unified into the same projection coordinate system, and the original elevation data covering the entire high-relief small watershed is output.

[0084] Based on the corrected elevation information, the local slope and curvature are calculated for each location in a GIS environment or a dedicated script. Based on this, a nonlinear mesh edge length function is constructed to segment high-lying steep slopes and appropriately coarsen flat areas. The Delaunay triangulation algorithm is then used to automatically generate a series of triangular mesh cells using this edge length function as a constraint. Each cell is a basic calculation unit, and its geometric center, vertices, and adjacency relationships are determined by this triangulation process.

[0085] After the grid is generated, the continuous field data constructed in the preprocessing stage, such as elevation, permeability coefficient, soil thickness, and vegetation coverage, can be immediately extracted at the center coordinates of each triangular unit and spatially superimposed with the facility information layer.

[0086] Equation configuration and numerical solution are performed. Based on the multidimensional properties of the grid cells, the conservation of mass and momentum equations are applied to each triangular cell to describe the rainfall input, surface runoff, and infiltration convergence process. First, the center elevation, permeability coefficient, soil thickness, and vegetation coverage of each grid cell are used as initial conditions. Real-time rainfall intensity and boundary water level observations are used as boundary inputs. Through the finite volume iteration method, the water depth change and flow velocity distribution of each cell within each time step are calculated. During the solution process, the infiltration loss term in the form of Darcy's law is introduced for the infiltration process, so that the infiltration amount is related to the permeability coefficient and soil thickness.

[0087] The numerical solution is executed in parallel on the cloud cluster, and the water depth and flow velocity time series of each unit are output in real time, providing physical simulation results for neural network training and local prediction.

[0088] Taking spatial coordinates and timestamps as input, a deep feedforward neural network is constructed. The number of nodes and layers is set according to the nonlinear characteristics of the watershed, and the water depth prediction value at the corresponding moment is generated at the output of the network.

[0089] During network training, two losses are introduced simultaneously: one is the data fitting loss, which compares the network predictions with the water depth values observed by each sensor during the rainfall event; the other is the physical residual loss, which calculates the degree to which the predicted output is not satisfied by the mass conservation and momentum conservation equations, and uses this residual as a penalty term.

[0090] By adjusting the network weights, the trained network can not only accurately restore the historically observed water depths, but also naturally follow the laws of hydrodynamics during the prediction process.

[0091] It should be noted that batch training with global historical data is used in the initial training phase, followed by small-batch fine-tuning in the cloud for extreme rainstorm events to balance universality and local accuracy.

[0092] The physical numerical solution results are compared with the neural network prediction results. The root mean square error is used as an indicator to evaluate the accuracy of the water depth prediction. 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.

[0093] 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 a rainstorm. The neural network first quickly gives a short-term water depth prediction, which is then corrected by combining local numerical solutions. Finally, the comprehensive prediction results are pushed to the control unit through the message middleware for intelligent scheduling of gates and pump groups, ensuring both high-precision predictions and low-latency responses in high-altitude sponge city rainwater management scenarios.

[0094] Step 2: After completing spatial discrete grid division and attribute mapping, generalize and adapt to the diversity of terrain, soil, and vegetation through the collaboration of multiple predictors and the federated learning mechanism. The specific steps are as follows:

[0095] Four complementary predictors are constructed based on the attributes of each grid cell, including the center elevation, hydraulic conductivity, soil thickness, vegetation coverage, historical rainfall time series, and corresponding water depth observations.

[0096] First, the simulation error correction predictor is trained using a decision tree regression algorithm, targeting the difference between physical simulation results and sensor observations, to capture the systematic deviation of the physical solution in local extreme processes.

[0097] Using the physical equation consistency constraint and water depth observation as dual losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the hydrodynamic conservation.

[0098] Taking the multidimensional attributes of grid cells and historical runoff as input, an attribute association predictor based on an ensemble tree algorithm is trained to exploit the nonlinear interactions between soil, vegetation, and topography.

[0099] Finally, the time-dependent characteristics of rainfall and runoff are used as input to train a long short-term memory network to capture the fluctuation trend after rainfall mutations.

[0100] After each predictor completes batch training 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 results as a new sample for iterative update, allowing the model to take into account both global versatility and local scene characteristics.

[0101] A specific example is as follows: During the multivariate predictor construction phase, targeted training is performed for each model type based on the unit center elevation, permeability coefficient, soil thickness, vegetation cover, and measured water depth from historical rainfall;

[0102] For example, for a certain grid cell (center elevation 350 meters, permeability 0.000012 meters per second, soil thickness 2.5 meters, and vegetation coverage 65%), the hourly rainfall sequence during the rainstorm on June 15, 2024, is [20, 30, 40, 10] mm / hour, corresponding to the observed water depth of [0.10, 0.30, 0.40, 0.20] meters; the physical simulation solution outputs [0.08, 0.25, 0.35, 0.15] meters, and the residuals between the two are [0.02, 0.05, 0.05, 0.05] meters. Using this residual sequence and grid cell characteristics as samples, a decision tree regression algorithm (maximum tree depth 5 layers, minimum leaf node samples 10) is used to train the simulation error correction predictor, enabling it to automatically correct the underestimation bias of the physical solution at the rainfall peak under similar rainstorm conditions.

[0103] Based on this, a physical information augmentation network (PIN) was designed, with the dual objectives of matching the physical simulation results and observed water depths for the same unit. The network consists of five layers, 64 neurons per layer, and uses Reinforced Lu (ReLU) as the activation function. Training is performed by simultaneously minimizing the water depth fitting error and the residual error (incorporating the residuals of the mass and momentum conservation equations into the loss function), ensuring that the predictions are consistent with measured data while adhering to the laws of hydrodynamics. Furthermore, an attribute association predictor was trained using an ensemble tree algorithm (100 subtrees, learning rate 0.1), using the unit's multidimensional attributes and accumulated runoff (e.g., the maximum hourly runoff of 0.15 meters corresponding to the aforementioned heavy rain) as input. This model further exploits the nonlinear interactions between topography, soil, and vegetation on runoff. Finally, the rainfall time series and the corresponding runoff series were fed into a long-short-term memory network (50 hidden nodes, 4-hour time step) to capture the dynamic characteristics of water depth fluctuations following sudden rainfall changes, thereby constructing a time-series forecast model for water depth trends within the next 1–2 hours.

[0104] Based on the historical backtesting database, typical rainstorm events such as short-term heavy rainfall, long-term weak rainfall, and spatially uneven rainfall are divided into several scenario categories. In each category, key performance indicators such as peak arrival time, peak height, and fluctuation trend of each predictor are statistically analyzed to form a scenario performance matrix.

[0105] During actual operation, the system first automatically matches the most suitable scenario category based on the current rainfall time series (including the inflection point of the accumulated rainfall curve and the time of rainfall intensity mutation) and the terrain segment division, and extracts the best-performing predictor in this category from the matrix as the main model. If the water depth curve output by the main model deviates from the real-time observation data by more than the preset threshold, the backup model calling mechanism is triggered, and the system switches between the suboptimal models in turn until the prediction accuracy is restored.

[0106] 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 a majority improvement selection strategy for the same network layer or decision tree branch structure: for parameter directions that have improved in most nodes, the update is retained; for parameters that have changed in only a few nodes, they are considered to be noise suppression and 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.

[0107] 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 will be automatically triggered, and the error distribution of the watershed and the real-time observation residuals will be 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.

[0108] 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 threshold filtered to determine the moment when it first breaks the threshold 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 pulsations, it is also required that the rainfall intensity continues to be lower than the threshold for a minimum duration set by experts before the rainstorm can be considered to have ended.

[0109] Step 3: Perform online edge fine-tuning and uncertainty quantification. 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 time-varying environmental characteristics and risk mitigation capabilities for scheduling decisions. The specific steps are as follows:

[0110] Perform online incremental fine-tuning. After each round of heavy rain, the difference between the actual water depth series recorded by local sensors and the predicted water depth series output by the integrated predictor is automatically summarized to construct a residual dataset containing timestamps and corresponding residuals. For the latest data periods in this residual dataset, the most recent 50 residuals are selected in descending order as fine-tuning samples to ensure that the samples used can fully reflect the current soil moisture content, vegetation cover changes, and model deviations caused by adjustments to the sudden drainage strategy.

[0111] After sample preparation, the physical information-augmented neural network model was loaded onto the edge node. An iterative training process with the least squares residual sum as the loss function was set up. Ten rounds of backpropagation optimization were performed with a learning rate of 0.005 and a batch size of 10. During this process, only the parameters of the feature extraction layer and output layer related to time-dependent features and spatial attribute mapping in the network were updated, ensuring fine-tuning speed and avoiding the computational overhead of full retraining.

[0112] After fine-tuning is completed, the updated model parameter snapshot is saved together with the latest residual statistics, and the old model is immediately replaced and used for water depth prediction in the next period to support high-precision water depth prediction in subsequent periods.

[0113] After fine-tuning, the system performs a self-check on the new model at the edge node: using the latest 50 observation inputs, the predicted output is recalculated and compared with the true value. If the root mean square error decreases by at least 10% compared to before fine-tuning, the online iteration is confirmed to be valid; otherwise, it returns to the last valid parameter state and reports to the cloud.

[0114] Uncertainty quantification: After completing online fine-tuning, the uncertainty quantification phase is immediately initiated. For the input data at the same moment, forward reasoning is performed by randomly blocking some connected nodes in the network multiple times (for example, 30 times) to simulate the output fluctuations of the model under different internal states.

[0115] Each inference generates a set of water depth prediction values. After summarizing these 30 sets of prediction results, their arithmetic mean and standard deviation are calculated to characterize the central tendency and dispersion of the model under the current environmental conditions. Based on the mean and standard deviation, the upper and lower limits of the 95% confidence interval are derived to reflect the uncertainty range of the prediction.

[0116] After obtaining the confidence interval, compare it with the preset flood control and infiltration thresholds:

[0117] When the confidence upper limit approaches or exceeds the flood control threshold, more storage space is automatically reserved and the emergency diversion strategy is activated in advance;

[0118] When the confidence lower limit is much lower than the infiltration requirement, the operating power of the pump unit can be appropriately reduced to promote natural infiltration;

[0119] If the confidence interval as a whole falls within the safe interval, regular scheduling is performed according to the central trend value;

[0120] 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 federated learning performance feedback and model improvement, achieving closed-loop optimization and risk control between the end-edge and cloud.

[0121] It should be noted that flood control thresholds, infiltration requirements and safety intervals are determined based on actual measurements or expert opinions, and can also be numerically adjusted based on actual conditions.

[0122] Step 4: Conduct risk-aware scheduling and collaborate with the cloud, edge, and device. After completing online fine-tuning and uncertainty quantification, enter the scheduling decision phase. Combine the predicted confidence interval with the flood control and infiltration demand thresholds to achieve risk-aware intelligent scheduling. The specific steps are as follows:

[0123] 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 unit with the preset flood control depth threshold and minimum infiltration threshold one by one, and divides the units into three categories: high risk, low risk and safe.

[0124] For grid cells identified as high-risk, where the upper limit of the confidence interval exceeds the flood control depth threshold, priority is given to expanding the proportion of storage pond usage and generating preliminary control instructions for the gate opening range (e.g., 60%–80%) and pump group available power range (e.g., 30%–50%) for that cell;

[0125] For units judged as low risk, where the lower limit of the confidence interval is lower than the minimum infiltration threshold, a recommendation is made to prioritize the preservation of the infiltration pool capacity, with the lower limit of the pump group power set to 0% and the upper limit set to the minimum acceptable operating power (e.g., 10%) to promote natural infiltration;

[0126] For units in the safe zone, the default opening degree and power range of the gates and pump groups are generated according to the conventional scheduling logic. This preliminary scheduling result is formed into an edge scheduling table based on the unit, including each unit's risk level, recommended gate opening range, recommended pump group power range and corresponding trigger conditions. It is generated and stored locally at the edge node in seconds, waiting for global optimization call on the cloud.

[0127] After receiving the edge scheduling table and corresponding confidence interval report uploaded by all edge nodes, the cloud first summarizes the scheduling suggestions of each grid unit and constructs a set of global water balance equations to ensure that the total inflow, total infiltration and total discharge are within the allowable range. At the same time, a multi-scenario simulation method is adopted to simulate the water level response and pump energy consumption performance after the execution of each edge suggestion combination. Through parallel computing, a complete scheduling plan that meets the requirements of urban flood control and energy consumption optimization is selected. The order of use of each storage tank, the opening curve of each gate in the future period (such as every 5 minutes), and the time-sharing power setting of the pump group are clarified, and the risk evolution prediction and emergency switching logic of each unit are attached. After the optimization is completed, the cloud will send a global scheduling package containing the global optimal scheduling timing, risk level spectrum and operation instructions to each edge node to ensure that the edge can execute safely and efficiently according to unified standards after receiving the instructions, forming a closed-loop collaboration between cloud, edge and terminal.

[0128] The operation and maintenance platform overlays a risk level color map on the GIS interface, marking each triangular grid cell with warning red, warning yellow, and safety green as high risk, medium risk, and safety levels, respectively. A floating panel displays the predicted confidence interval, recommended gate opening range, and pump group power range for each grid cell. Control instructions containing the precise opening curve and power setting sequence for the corresponding grid cell in the global scheduling package are pushed to each edge node through a secure channel.

[0129] After execution, each edge node 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. The report includes a residual distribution map, deviation statistics, and a list of units that exceed the tolerance threshold (such as ±5%), which is uploaded to the cloud in real time. After the cloud aggregates the reports of each node, it compares the historical performance with the feedback of this round to form a new performance feedback report, and feeds back the typical patterns of water depth error and high or low energy consumption in the report to the federated learning system as the key improvement target 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 iterations, and dynamic adaptation to sudden climate and on-site disturbances is achieved in actual operation, ensuring the long-term and stable operation of the sponge city high-altitude rainwater management system.

[0130] The present invention systematically collects and preprocesses data in high-altitude areas, integrates borehole permeability coefficient, soil thickness and remote sensing vegetation cover interpolation to generate a continuous physical field, maps infiltration pool, storage pool 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 federated learning mechanism to dynamically match and synchronize the optimization model; after each rainstorm event, the observation prediction residuals are automatically extracted for online incremental iteration of the neural network, and confidence intervals are generated 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 in high altitude areas, realizing refined risk perception scheduling, and enhancing the effects of rainstorm prevention and control and rainwater resource utilization.

[0131] Example 2: A sponge city high-altitude rainwater management system, such as Figure 2 As shown, specifically including:

[0132] The high-relief network partitioning module is used to collect and preprocess data in high-relief areas. It combines the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage interpolation to eliminate anomalies and generate a continuous physical field. It also maps facility and pipeline information to the elevation coordinate system. After topological repair, it outputs a facility information layer and a facility attribute list to divide the grid cells.

[0133] The grid fine-tuning prediction module is used to build a predictor based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The main predictor is dynamically matched according to the scenario performance matrix. After generating a global update package through federated learning aggregation, the local model prediction is synchronously updated.

[0134] The prediction model adjustment module is used to automatically extract the latest observations and prediction residuals from edge nodes after each rainstorm event, and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to statistically analyze the central tendency and fluctuation range of the predicted values and generate the confidence interval of the model.

[0135] The management suggestion generation module is used to divide edge nodes into 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.

[0136] The above formulas are all dimensionless and calculated numerically. Specific dimension removal can be achieved by various means such as standardization, which will not be elaborated here. The formula is a formula obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0137] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other 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 comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. 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 via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, ATA hard drives, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state ATA hard drive.

[0138] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0139] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0140] In the several embodiments provided in this 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 schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as 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 mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0141] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0142] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0143] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for managing rainwater in high-altitude areas of a sponge city, characterized in that: The steps include: Data collection and preprocessing are performed on high-relief areas. The data is interpolated using the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage. Anomalies are removed to generate a continuous physical field. Facility and pipeline information is mapped to an elevation coordinate system. After topological repair, a facility information layer and a facility attribute list are output, and grid cells are divided. A predictor is built based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The primary predictor is dynamically matched based on the scenario performance matrix. A global update package is generated through federated learning aggregation and then the local model prediction is synchronously updated. After each rainstorm event, edge nodes automatically extract the latest observations and forecast residuals and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to statistically analyze the central tendency and fluctuation range of the predicted values and generate confidence intervals for the model. The edge node divides high-risk, low-risk, and safe units based on the predicted confidence interval, flood control depth threshold, and minimum infiltration threshold, and generates gate opening and pump group power recommendations; Build a predictor based on the attributes of the grid cell center, and after training in the cloud, send it 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, hydraulic conductivity, soil thickness, vegetation coverage, historical rainfall time series, and corresponding water depth observations 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 results and the sensor observation as the target; Using the physical equation consistency constraint and water depth observation as dual losses, the physical information enhanced neural network is trained so that the output fits the observation and satisfies the hydrodynamic conservation. Taking the multidimensional attributes of grid cells and historical runoff as input, an attribute association predictor based on an ensemble tree algorithm is trained to exploit the nonlinear interactions between soil, vegetation, and topography. Taking the time-dependent characteristics of rainfall and runoff as input, a long short-term memory network is trained to capture the fluctuation trend after a sudden change in rainfall. After each predictor completes batch training in the cloud for the first time, 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 results as a new sample for iterative update. Dynamically match the primary predictor based on the scenario performance matrix, aggregate and generate a global update package through federated learning, and then synchronously update the local model prediction. The specific process is as follows: Based on the historical backtesting database, short-duration heavy rainfall, long-duration weak rainfall, and typical rainstorm events with uneven spatial distribution are divided into different scenario categories. The peak arrival time, peak height, and fluctuation trend performance indicators of each predictor in each category are statistically analyzed to form a scenario performance matrix. Based on the current rainfall time series and terrain segmentation, the most suitable scenario category is automatically matched, and the best predictor for that scenario category is extracted from the matrix as the primary model. If the water depth curve output by the primary model deviates from the real-time observation data by more than a preset threshold, the backup model call mechanism is triggered, switching between the suboptimal models 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, a 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 it to the local model for prediction; After the global update deployment is complete, 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 rainfall scenarios. The root mean square error and maximum absolute error indicators are calculated and compared with the results of the historical version. If the error indicators of a watershed rebound or fall 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.

2. A method for managing rainwater in high-altitude sponge cities according to claim 1, characterized in that: Data collection and preprocessing are performed on high-relief areas. The data are then combined with the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage interpolation to remove anomalies and generate a continuous physical field. The specific process is as follows: Obtain the original elevation data of the high-relief area, perform projection transformation, unify it into the same coordinate system, apply Gaussian smoothing filtering to the projected elevation data to remove measurement noise and form smoothed elevation information, then perform polynomial error correction on the smoothed elevation information based on the ground control points and output the corrected elevation information; Soil and vegetation parameters are interpolated for high-relief areas, soil permeability coefficient data and soil layer thickness data are obtained from borehole sampling, and vegetation coverage data are obtained from remote sensing classification; Kriging interpolation is applied to the soil permeability coefficient, soil thickness and vegetation coverage data respectively to generate continuous permeability coefficient field, soil thickness field and vegetation coverage field; The outliers in the interpolated permeability coefficient field, soil thickness field and vegetation coverage field are removed, and the gaps are filled by interpolation again.

3. A method for managing rainwater in high-altitude areas of a sponge city according to claim 2, characterized in that: 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. The specific process is as follows: Extract facility and pipeline information from regional design drawings, including the spatial location and structural parameters of infiltration tanks, storage tanks, 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 broken or overlapping pipe segments in the facility information layer and output a complete facility attribute list. Based on the corrected elevation information, the local slope and curvature are calculated for each location in the GIS environment, and the Delaunay triangulation algorithm is called to automatically generate a series of triangular mesh units with the edge length function as the constraint; The center elevation, permeability, soil thickness and vegetation coverage of each grid cell are used as initial conditions, and the real-time rainfall intensity and boundary water level observations are used as boundary inputs. The finite volume iteration method is used to calculate the water depth change and flow velocity distribution of each cell in each time step.

4. A method for managing rainwater in high-altitude sponge cities according to claim 3, characterized in that: After each rainstorm event, the edge nodes automatically extract the latest observation and forecast residuals and use them as training samples to perform incremental iterations of the neural network. The specific process is as follows: After each round of heavy rain, the difference between the actual water depth series recorded by local sensors and the predicted water depth series output by the integrated predictor is automatically summarized to construct a residual dataset containing timestamps and corresponding residuals. For the latest period of data in the residual dataset, the most recent 50 residuals are selected in descending order as fine-tuning samples. After the sample preparation is completed, the physical information enhanced 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. 10 rounds of backpropagation optimization are performed with parameters of 0.005 learning rate and 10 batch size. After fine-tuning is completed, the updated model parameter snapshot is saved together with the latest residual statistics, and the old model is replaced and immediately used for water depth prediction in the next period.

5. A method for managing rainwater in high-altitude areas of a sponge city according to claim 4, characterized in that: After fine-tuning is completed, a random masking mechanism is used for forward inference to 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 a self-check on the new model at the edge node: using the latest 50 observations, the predicted output is recalculated and compared with the true value. If the root mean square error decreases by 10% or more compared to before fine-tuning, the online iteration is considered valid. Otherwise, the system returns to the last valid parameter state and reports to the cloud. After completing online fine-tuning, the uncertainty quantification phase begins. For the input data at the same moment, some connected nodes in the network are randomly blocked and forward reasoning is performed to simulate 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, the arithmetic mean and standard deviation are calculated. Based on the mean and standard deviation, the upper and lower limits of the confidence interval are derived. After obtaining the confidence interval, compare the confidence interval with the preset flood control and infiltration thresholds: When the confidence upper limit approaches or exceeds the flood control threshold, more storage space is automatically reserved and the emergency diversion strategy is activated in advance; When the confidence lower limit is far lower than the infiltration requirement, the operating power of the pump unit is reduced and natural infiltration is carried out; If the confidence interval as a whole falls within the safe interval, normal scheduling is performed.

6. A method for managing rainwater in high-altitude areas of a sponge city according to claim 5, characterized in that: The edge node divides high-risk, low-risk, and safe units based on the predicted confidence interval, flood control depth threshold, and minimum infiltration threshold, and generates gate opening and pump group power recommendations. The specific process is as follows: The upper and lower limits of the predicted confidence interval of each triangular grid cell are compared one by one with the preset flood control depth threshold and minimum infiltration threshold, and the cells are divided into three categories: high risk, low risk and safe; When the upper limit of the confidence interval exceeds the flood control depth threshold, the grid cell is determined to be high-risk and priority is given to expanding the proportion of storage ponds used; When the lower limit of the confidence interval is lower than the minimum infiltration threshold, the grid cell is judged to be low-risk, and the infiltration pond capacity recommendation is prioritized for natural infiltration treatment; For grid cells in the safety zone, the default opening degree and power range of the gate and pump group are generated according to the conventional scheduling logic.

7. A sponge city high-altitude rainwater management system, used to implement a sponge city high-altitude rainwater management method according to any one of claims 1 to 6, characterized in that: include: The high-relief network partitioning module is used to collect and preprocess data in high-relief areas. It combines the borehole permeability coefficient, soil thickness, and remote sensing vegetation coverage interpolation to eliminate anomalies and generate a continuous physical field. It also maps facility and pipeline information to the elevation coordinate system. After topological repair, it outputs a facility information layer and a facility attribute list to divide the grid cells. The grid fine-tuning prediction module is used to build a predictor based on the attributes of the grid cell center. After training in the cloud, it is sent to the edge node for local incremental fine-tuning. The main predictor is dynamically matched according to the scenario performance matrix. After generating a global update package through federated learning aggregation, the local model prediction is synchronously updated. The prediction model adjustment module is used to automatically extract the latest observations and prediction residuals from edge nodes after each rainstorm event, and use them as training samples to perform incremental iterations of the neural network. After fine-tuning, a random masking mechanism is used for forward reasoning to 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 to divide edge nodes into 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.

Images