Urban rainwater pipe network and river network hydrodynamic accelerated calculation method based on data-driven substitute model
By constructing the Deep-Drain model, the problems of high computational complexity and complex parameter calibration of traditional models are solved, enabling rapid and high-precision simulation of urban stormwater pipe networks and river flow, thus meeting the rapid response needs of urban flood risk management.
Patent Information
- Application Number
- CN202610482390.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-13
- Publication Date
- 2026-06-23
AI Technical Summary
Traditional one-dimensional hydrodynamic models have high computational complexity and complex parameter calibration processes. Existing data-driven methods are difficult to effectively capture the spatiotemporal dynamic characteristics of urban stormwater pipe networks and river networks. Hyperparameter optimization is inefficient and cannot meet the rapid response requirements of urban flood risk management.
Deep-Drain, a data-driven alternative model based on bidirectional long short-term memory network, graph attention network and residual connection, is constructed. By combining Bayesian optimization method to find the optimal hyperparameter combination, it can partially replace the traditional mechanism model.
It significantly shortens simulation calculation time, maintains high accuracy, and can quickly respond to urban flooding risk assessment and emergency dispatch decisions, possessing good generalization ability and engineering application value.
Smart Images
Figure CN122263671A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of hydrological and water resources and urban hydrodynamic simulation technology, and specifically relates to a method for accelerating the calculation of urban stormwater pipe network and river network hydrodynamics based on a data-driven alternative model. Background Technology
[0002] Urban stormwater drainage network and river network flow simulation is a key component of urban flood risk management, playing a crucial role in urban flooding early warning, drainage system optimization design, and emergency dispatch decision-making. Traditional one-dimensional hydrodynamic models (such as InfoWorks ICM and SWMM) accurately describe pipeline and river structures, boundary conditions, and water transport processes by solving partial differential equations such as the Saint-Venant equations, and can reflect the physical mechanisms of hydraulic systems relatively well. However, these mechanistic models face many challenges in practical applications: First, traditional mechanistic models suffer from high computational complexity and long simulation times. For large-scale urban stormwater drainage and river network coupled systems, the models need to handle tens of thousands of nodes and pipe segments, resulting in enormous computational demands that make it difficult to meet the needs of real-time or near-real-time urban flooding risk assessment and emergency dispatch decision-making. Especially when conducting multi-scenario simulations under extreme rainfall conditions, the computation time increases exponentially, severely limiting the practicality of the models.
[0003] Secondly, mechanistic models have high requirements for the quality of input data, and the parameter calibration process is complex. The model needs precise parameters such as pipeline topology, pipe diameter, slope, and roughness, as well as accurate boundary and initial conditions. In practical engineering, these parameters are often difficult to obtain completely, and the calibration process is time-consuming and labor-intensive, hindering the widespread application of the model.
[0004] In recent years, with the rapid development of machine learning technology, data-driven methods have shown great potential in the field of hydrodynamic simulation and prediction. Data-driven models, by learning patterns from historical observation data or the output data of mechanistic models, can significantly improve computational speed, enabling rapid response. However, existing data-driven methods still have significant shortcomings in the hydrodynamic simulation of urban stormwater drainage networks and river networks: On the one hand, traditional machine learning methods (such as support vector machines and random forests) are difficult to effectively capture the spatiotemporal dynamic characteristics of hydrodynamic processes. Stormwater pipe networks and river network systems have complex topological structures and spatiotemporal dependencies, and simple data-driven models cannot fully express these complex nonlinear relationships.
[0005] On the other hand, existing deep learning methods have limitations when processing graph-structured data. Although recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) can capture time-series features, they have limited ability to handle the spatial topological relationships of pipe networks and river networks; while graph convolutional networks (GCNs) can handle graph structures, they are insufficient in modeling dynamic features over time and lack differentiated processing of the importance between nodes.
[0006] Furthermore, existing data-driven models often employ grid search or random search methods for hyperparameter optimization, which are inefficient, make it difficult to find the optimal model configuration, and hinder the full realization of model performance.
[0007] Therefore, there is an urgent need for a new computational method that can balance simulation accuracy and computational efficiency, effectively capture the spatiotemporal characteristics of hydrodynamic processes, and possess efficient hyperparameter optimization capabilities, in order to meet the practical needs of rapid simulation of urban stormwater pipe networks and river networks, and provide efficient and reliable technical support for urban flood risk management. Summary of the Invention
[0008] The technical problem to be solved by this invention is to provide an accelerated calculation method for urban stormwater pipe network and river network hydrodynamics based on a data-driven alternative model. By constructing a data-driven alternative model, this invention effectively replaces the calculation process of traditional mechanism models, significantly shortens the simulation calculation time, and meets the needs of rapid response for urban flood risk assessment and emergency dispatch decision-making.
[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A data-driven alternative model-based method for accelerating the hydrodynamic calculation of urban stormwater drainage networks and river networks, comprising the following steps: A data-driven alternative model, Deep-Drain, is constructed based on a bidirectional long short-term memory network, a graph attention network, and residual connections to simulate a partially simplified process of urban stormwater drainage networks and river flow. We construct a mechanistic model to simulate the one-dimensional hydrodynamic process under measured rainfall and design rainfall with different return periods, and process and construct a dataset of stormwater pipe network and river network flow process, while dividing it into training set, validation set and test set; The optimal combination of hyperparameters for Deep-Drain is found using Bayesian optimization methods. Quantitative analysis of the efficiency improvement effect of Deep-Drain using typhoon and rainstorm events.
[0010] Preferably, the Deep-Drain understands the spatiotemporal dynamic characteristics of rainwater pipe network and river network flow by combining graph attention network and bidirectional long short-term memory network. The graph attention network aggregates information in the graph structure to capture the spatial dependencies between rainwater wells and river cross-sections in the rainwater pipe network and river network, while the bidirectional long short-term memory network captures the temporal dependencies.
[0011] Preferably, the bidirectional long short-term memory network employs two independent long short-term memory networks, which process sequences in ascending and descending order respectively, and merge them through connection or weighted summation to generate an information representation containing complete sequential relationships; The graph attention network identifies the topology and node feature information of the graph, and aggregates the information of neighboring nodes by introducing learnable attention weighting between nodes. The residual connection is used to alleviate the gradient vanishing problem in the stacking of multiple Deep-Drain modules, promote the effective flow of information, and improve the expressive power, convergence speed and generalization ability of Deep-Drain.
[0012] Preferably, the rainwater pipe network and river network flow process dataset includes the inflow of rainwater wells and river cross-sections, the water level of rainwater wells and river cross-sections, the overflow of rainwater wells and river cross-sections, and the return flow of rainwater wells and river cross-sections.
[0013] Preferably, the Bayesian optimization method includes the following steps: A probabilistic proxy model between hyperparameters and network performance is constructed using Gaussian processes. The next observation location is determined by improving the acquisition function, and the optimal observation characteristics that the location can provide are predicted. In each optimization iteration, the hyperparameter combination most likely to bring the greatest improvement is selected from the current Gaussian process model, the performance of Deep-Drain under this combination is evaluated, and the new hyperparameter combination and performance pair are added to the training set as a new pair of observation data to update the Gaussian process model. Optimization stops when the maximum number of iterations is reached.
[0014] Preferably, the specific process of constructing a probabilistic proxy model between hyperparameters and network performance through a Gaussian process is as follows: A probabilistic proxy model between hyperparameters and network performance is constructed using Gaussian processes, and an objective function is then built. f ( x () represents Deep-Drain in hyperparameter combination x The performance of the following f ( x It follows a Gaussian process: ; in,m ( x ) is the mean function, representing the combination of hyperparameters. x Expected performance; k ( x , x ') is the covariance function, describing the combination of hyperparameters. x and x Similarity between '; GP represents a Gaussian process.
[0015] Preferably, the specific process of determining the next observation location by improving the acquisition function and predicting the optimal observation characteristics that the location can provide is as follows: The next observation location is determined by improving the acquisition function, and the optimal observation characteristics that location can provide are predicted, assuming the current... x This is the optimal value, at this point f ( x When the maximum value is reached, the new query position is calculated based on the expected improvement amount. The EI formula is: ; in, x This represents the current optimal combination of hyperparameters; f ( x ) is a combination of hyperparameters x Corresponding network performance; f ( x EI represents the currently known best network performance; EI represents the expected improvement. EI's goal is to maximize f ( x Relative to f ( x The expected improvement is achieved by using the predicted mean of GP. μ ( x ) and standard deviation σ ( x EI can be expressed as: ; in, ; The cumulative distribution function of the standard normal distribution; It is the probability density function of the standard normal distribution.
[0016] Preferably, the step of quantitatively analyzing the efficiency improvement effect of Deep-Drain based on typhoon and rainstorm events includes: By comparing simulated and measured values, the Nash efficiency was used to quantify the process error of water level in rainwater wells and river sections. The coefficient of determination was used to evaluate the ability of the Deep-Drain learning mechanism model to simulate peak water levels in one-dimensional stormwater pipe networks and river networks. The speedup ratio was used to evaluate the improvement in computational efficiency of Deep-Drain relative to the mechanistic model.
[0017] Preferably, the Nash efficiency equation is: ; in, , They are respectively t Simulated and measured values at any given time; This is the average of the measured values; n The number of moments.
[0018] Preferably, the coefficient of determination R is used. 2 The ability of the Deep-Drain learning mechanism model to simulate peak water levels in one-dimensional stormwater pipe networks and river networks was evaluated, with the coefficient of determination as follows: ; in, To simulate the mechanism model i Peak water levels at individual rainwater wells and river sections, in meters; For Deep-Drain simulation of the first i Peak water levels at individual rainwater wells and river sections, in meters; The mean value (m) of the peak water levels in the storm drain well and river cross-section is used to simulate the mechanism model. n It refers to the number of rainwater wells and river cross-sections.
[0019] Preferably, the speedup ratio is used to evaluate the improvement in computational efficiency of Deep-Drain relative to the mechanistic model, where the speedup ratio is: ; in, T 机理模型 The simulation computation time for the mechanism model; T Deep-Drain This represents the simulation computation time for Deep-Drain.
[0020] Preferably, the ratio of the training set, validation set, and test set is 0.6, 0.2, and 0.2, respectively.
[0021] Preferably, the Bayesian optimization is performed in 200 iterations.
[0022] An accelerated computational device for urban stormwater drainage networks and river networks based on a data-driven alternative model, employing the aforementioned accelerated computational method for urban stormwater drainage networks and river networks based on a data-driven alternative model, includes: The model building module is used to build a data-driven alternative model, Deep-Drain, which is built based on a bidirectional long short-term memory network, a graph attention network, and residual connections. It is used to simulate a partially simplified process of urban stormwater pipe network and river network flow. The data processing module is used to build a mechanism model to simulate the one-dimensional hydrodynamic process under measured rainfall and design rainfall with different return periods, and to process and build a dataset of stormwater pipe network and river network flow process, while dividing it into training set, validation set and test set; The parameter optimization module is used to find the optimal combination of hyperparameters for the Deep-Drain using Bayesian optimization methods. The performance evaluation module is used to quantitatively analyze the efficiency improvement effect of the Deep-Drain based on typhoon and rainstorm events.
[0023] A computer device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the steps of the data-driven alternative model-based accelerated calculation method for urban stormwater drainage networks and river networks.
[0024] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the data-driven alternative model-based accelerated calculation method for urban stormwater drainage networks and river networks hydrodynamics.
[0025] The present invention can achieve the following beneficial effects: 1. This invention replaces some computational steps of traditional mechanistic models with a data-driven alternative model, achieving rapid calculation of urban stormwater drainage networks and river flow while maintaining simulation accuracy. The alternative model can be trained based on historical observation data or using the output of the mechanistic model, and can be continuously optimized through incremental learning, avoiding the dependence of traditional models on the calibration of a large number of parameters. This invention balances the accuracy and efficiency of one-dimensional hydrodynamic simulation, providing efficient and reliable technical support for urban drainage scheduling and flood risk management.
[0026] 2. This invention, by constructing a data-driven alternative model, effectively replaces the calculation process of traditional mechanistic models, significantly shortens the simulation calculation time, and meets the needs of rapid response for urban flood risk assessment and emergency dispatch decision-making.
[0027] 3. While ensuring a significant increase in calculation speed, this invention maintains high simulation accuracy and can accurately capture the nonlinear laws of the flow process in rainwater pipe networks and river networks, providing reliable technical support for engineering applications.
[0028] 4. By combining bidirectional long short-term memory networks and graph attention networks, the spatiotemporal dynamic characteristics of rainwater pipe networks and river network flow can be effectively understood, realizing comprehensive modeling of complex hydrodynamic processes, and demonstrating good generalization ability and engineering application value. Attached Figure Description
[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention; Figure 2 A schematic diagram of the research area in Cangshan District, Fuzhou City; Figure 3 The process of typhoon Lupi, including its rainstorm and tidal events; Figure 4 The determining factor for the peak water level in rainwater wells; Figure 5 It is the determining factor for the peak water level at the river cross-section. Detailed Implementation
[0030] Example 1: This embodiment relates to a data-driven alternative model-based method for accelerating the hydrodynamic calculation of urban stormwater drainage networks and river networks, comprising the following steps: S1, based on Bidirectional Long Short-Term Memory (BiLSTM), Graph Attention Network (GAT) and residual connections, constructs a data-driven alternative model Deep-Drain to simulate a partially simplified process of urban stormwater drainage network and river network flow. S2 constructs a mechanism model to simulate the one-dimensional hydrodynamic process under measured rainfall and design rainfall with different return periods, and processes and constructs a dataset of stormwater pipe network and river network flow process, while dividing it into training set, validation set and test set; S3, Bayesian optimization is used to find the optimal combination of hyperparameters for Deep-Drain; S4. Quantitatively analyze the efficiency improvement effect of Deep-Drain using typhoon and rainstorm events.
[0031] In this embodiment, BiLSTM and GAT, which are involved in the data-driven alternative model in S1 above, are both mature deep learning methods. The specific calculation and solution processes are existing technologies and can be implemented through open-source code libraries. The mechanism model in S2 above is based on Infoworks ICM and is an existing model; therefore, the detailed solution process described above is not listed.
[0032] S11, BiLSTM is an extension of LSTM. It uses two independent LSTMs to process the sequence in forward order (from front to back) and in reverse order (from back to front), respectively, and merges the outputs by concatenation or weighted summation to generate an information representation containing the complete sequential relationship. S12, GAT identifies the topological structure and node feature information of a graph. By introducing learnable attention weighting between nodes to aggregate information of neighboring nodes, it can predict nodes, edges or graphs. S13, residual connections effectively alleviate the gradient vanishing problem caused by stacking multiple Deep-Drain modules, promote effective information flow, and improve the expressive power, convergence speed, and generalization ability of Deep-Drain.
[0033] In this embodiment, the stormwater pipe network and river network flow process dataset in S2 includes stormwater well and river cross-section inflow, stormwater well and river cross-section water level, stormwater well and river cross-section overflow, and stormwater well and river cross-section return flow; the formula used to calculate the design rainfall for different return periods is: ; in, q The average rainfall intensity is expressed in mm / min. P For the return period, a; t The rainfall duration is in minutes; the rainfall process was calculated using the Chicago rainfall pattern, and the design peak rainfall coefficient was used. r =0.39.
[0034] Accordingly, the Bayesian optimization (BO) in S3 includes the following steps: S31, a probabilistic surrogate model between hyperparameters and network performance is constructed using a Gaussian process (GP). The objective function is then constructed. f ( x () represents Deep-Drain in hyperparameter combination x The performance of the following f ( x Obey GP: ; in, m ( x ) is the mean function, representing the combination of hyperparameters. x Expected performance; k ( x , x ') is the covariance function, describing the combination of hyperparameters. x and x Similarity between '; S32, the next observation location is determined using the Expected Improvement (EI) acquisition function, and the optimal observation characteristics that location can provide are predicted. Assume the current... x This is the optimal value, at this point f ( x Once the maximum value is reached, the new query position is calculated based on the expected improvement. The EI formula is: ; in, x This represents the current optimal combination of hyperparameters; f ( x ) is a combination of hyperparameters x Corresponding network performance; f ( x The current best known network performance is 1 / 3. EI's goal is to maximize f ( x Relative to f ( x Expected improvement. (Based on the predicted mean of GP) μ ( x ) and standard deviation σ ( x EI can be expressed as: ; in, ; The cumulative distribution function of the standard normal distribution; is the probability density function of the standard normal distribution; S33, in each optimization iteration, select the hyperparameter combination most likely to bring the greatest improvement from the current GP model. x n+1 The performance of Deep-Drain under this combination was evaluated, and the performance values were obtained. y n+1 The new hyperparameter combination and performance pair ( x n+1 , y n+1 The observations are added as a new pair to the training set to update the GP model. S34: Stop optimization when the set maximum number of iterations is reached; otherwise, return to S32 and repeat the above process.
[0035] In this embodiment, the evaluation indicators involved in the quantitative analysis of Deep-Drain's efficiency improvement effect using typhoon and rainstorm events in step S4 are as follows: S41. By comparing simulated and measured values, the Nash efficiency (NSE) is used to quantify the process error of water level in rainwater wells and river cross-sections. The Nash efficiency is: ; in, , They are respectively t Simulated and measured values at any given time; This is the average of the measured values; n Number of moments; S42, using the coefficient of determination (R²) 2 The ability of the Deep-Drain learning mechanism model to simulate peak water levels in one-dimensional stormwater pipe networks and river networks was evaluated. The coefficient of determination was: ; in, To simulate the mechanism model i Peak water levels at individual rainwater wells and river sections, in meters; For Deep-Drain simulation of the first i Peak water levels at individual rainwater wells and river sections, in meters; The mean value (m) of the peak water levels in the storm drain well and river cross-section is used to simulate the mechanism model. n It refers to the number of rainwater wells and river cross-sections; S43 uses speedup to evaluate the computational efficiency improvement of Deep-Drain relative to the mechanistic model. The speedup is: ; in, T 机理模型 The simulation computation time for the mechanism model; T Deep-Drain This represents the simulation computation time for Deep-Drain.
[0036] Example 2: The method of the present invention will be further described below using the central part of Cangshan District, Fuzhou City, Fujian Province as an example. This area covers 39.17 km². 2 The catchment areas are relatively independent with complex boundary conditions. A mechanistic model of the study area was constructed using Infoworks ICM to simulate one-dimensional hydrodynamic processes under measured rainfall and design rainfall with different return periods. Data sets of stormwater drainage network and river network flow processes were then processed to create datasets. Observed heavy rainfall events were collected from 90 measured rainfall events in Fuzhou City from January 2020 to January 2025. Design rainfall events were derived based on the heavy rainfall intensity formula for Fuzhou City, and 300 rainfall scenarios with return periods of 1-100 years and durations of 60 min, 120 min, and 180 min were designed. The processed data sets of stormwater drainage network and river network flow processes are shown in Table 1.
[0037] Table 1 Datasets of Rainwater Drainage Network and River Network Flow Processes
[0038] The stormwater drainage network and river network flow process datasets were divided into training, testing, and validation sets according to ratios of 0.6, 0.2, and 0.2, respectively. The number of iterations for BO was set to 200 to find the optimal combination of hyperparameters. The hyperparameters optimized by Deep-Drain and their corresponding search range are shown in Table 2.
[0039] Table 2 Hyperparameters of Deep-Drain
[0040] Furthermore, this example uses the trained and validated Deep-Drain to simulate the hydrodynamic processes under typhoon-induced rainfall. The performance is evaluated using the "Lupi" typhoon-induced rainfall event from 00:00 on August 3rd to 00:00 on August 8th, 2021. The rainfall from "Lupi" was 394.1 mm, and the rainfall and tidal processes are shown in Figure 3. The Deep-Drain calculation results are used as simulated values, and the calculation results from the mechanistic model are used as real values. The flow processes of stormwater drainage networks and river networks are compared to evaluate the ability of the data-driven alternative model to capture hydrological patterns and its computational efficiency.
[0041] Deep-Drain achieved average NSE values of 0.91 and 0.95 for simulated stormwater well and river cross-section water level processes under the rainfall of Typhoon Lupi, respectively, demonstrating a good understanding of the nonlinear relationships between stormwater well inflow and river cross-section water level changes. According to Figures 4 and 5, Deep-Drain and the mechanistic model simulated peak stormwater well and river cross-section water levels showed a significant positive correlation, with the simulated peak water levels of both exhibiting high R-values. 2 The average values are 0.99 and 0.99 respectively, indicating that Deep-Drain has a good ability to learn the mechanistic model to simulate peak water levels. Compared with river cross-sections, Deep-Drain has higher accuracy in simulating peak water levels in stormwater wells, although the accuracy decreases slightly under extreme rainfall. In terms of simulation efficiency, Deep-Drain's computation time is 3.85 minutes, while the mechanistic model's computation time is 27.16 minutes, resulting in a speedup ratio of 7.1. Deep-Drain improves the computational efficiency of simulating one-dimensional stormwater pipe networks and river network flow processes by 6.1 times.
[0042] This demonstrates that the data-driven alternative model Deep-Drain effectively learns the simulation capabilities of mechanistic models, can be used to replace the simplification process of simulating urban stormwater pipe networks and river network flow, and can significantly improve computational efficiency.
[0043] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.
Claims
1. A data-driven alternative model-based accelerated calculation method for urban stormwater drainage networks and river networks hydrodynamics, characterized in that, Includes the following steps: A data-driven alternative model, Deep-Drain, is constructed based on a bidirectional long short-term memory network, a graph attention network, and residual connections to simulate a partially simplified process of urban stormwater drainage networks and river flow. We construct a mechanistic model to simulate the one-dimensional hydrodynamic process under measured rainfall and design rainfall with different return periods, and process and construct a dataset of stormwater pipe network and river network flow process, while dividing it into training set, validation set and test set; The optimal combination of hyperparameters for Deep-Drain is found using Bayesian optimization methods. Quantitative analysis of the efficiency improvement effect of Deep-Drain using typhoon and rainstorm events.
2. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The Deep-Drain method combines graph attention networks and bidirectional long short-term memory networks to understand the spatiotemporal dynamic characteristics of stormwater pipe networks and river network flow. The graph attention network aggregates information in the graph structure to capture the spatial dependencies between stormwater wells and river cross-sections in the stormwater pipe network and river network, while the bidirectional long short-term memory network captures the temporal dependencies.
3. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 2, characterized in that, The bidirectional long short-term memory network employs two independent long short-term memory networks, which process sequences in ascending and descending order respectively, and merge the outputs through connection or weighted summation to generate an information representation containing complete sequential relationships. The graph attention network identifies the topology and node feature information of the graph, and aggregates the information of neighboring nodes by introducing learnable attention weighting between nodes. The residual connection is used to alleviate the gradient vanishing problem in the stacking of multiple Deep-Drain modules, promote the effective flow of information, and improve the expressive power, convergence speed and generalization ability of Deep-Drain.
4. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The stormwater pipe network and river network flow process dataset includes stormwater well and river cross-section inflow, stormwater well and river cross-section water level, stormwater well and river cross-section overflow, and stormwater well and river cross-section return flow.
5. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The Bayesian optimization method includes the following steps: A probabilistic proxy model between hyperparameters and network performance is constructed using Gaussian processes. The next observation location is determined by improving the acquisition function, and the optimal observation characteristics that the location can provide are predicted. In each optimization iteration, the hyperparameter combination most likely to bring the greatest improvement is selected from the current Gaussian process model, the performance of Deep-Drain under this combination is evaluated, and the new hyperparameter combination and performance pair are added to the training set as a new pair of observation data to update the Gaussian process model. Optimization stops when the maximum number of iterations is reached.
6. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 5, characterized in that, The specific process of constructing a probabilistic surrogate model between hyperparameters and network performance using a Gaussian process is as follows: A probabilistic proxy model between hyperparameters and network performance is constructed using Gaussian processes, and an objective function is then built. f ( x () represents Deep-Drain in hyperparameter combination x The performance of the following f ( x It follows a Gaussian process: ； in, m ( x ) is the mean function, representing the combination of hyperparameters. x Expected performance; k ( x , x ') is the covariance function, describing the combination of hyperparameters. x and x Similarity between '; GP represents a Gaussian process.
7. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 5, characterized in that, The specific process of determining the next observation location by improving the acquisition function and predicting the optimal observation characteristics that location can provide is as follows: The next observation location is determined by improving the acquisition function, and the optimal observation characteristics that location can provide are predicted, assuming the current... x This is the optimal value, at this point f ( x When the maximum value is reached, the new query position is calculated based on the expected improvement amount. The EI formula is: ； in, x This represents the current optimal combination of hyperparameters; f ( x ) is a combination of hyperparameters x Corresponding network performance; f ( x EI represents the currently known best network performance; EI represents the desired improvement. EI's goal is to maximize f ( x Relative to f ( x The expected improvement is achieved by using the predicted mean of GP. μ ( x ) and standard deviation σ ( x EI can be expressed as: ； in, ; The cumulative distribution function of the standard normal distribution; It is the probability density function of the standard normal distribution.
8. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The steps for quantitatively analyzing the efficiency improvement effect of Deep-Drain based on typhoon and rainstorm events include: By comparing simulated and measured values, the Nash efficiency was used to quantify the process error of water level in rainwater wells and river sections. The coefficient of determination was used to evaluate the ability of the Deep-Drain learning mechanism model to simulate peak water levels in one-dimensional stormwater pipe networks and river networks. The speedup ratio was used to evaluate the improvement in computational efficiency of Deep-Drain relative to the mechanistic model.
9. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 8, characterized in that, The Nash efficiency equation is: ； in, , They are respectively t Simulated and measured values at any given time; This is the average of the measured values; n The number of moments.
10. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 8, characterized in that, Using the coefficient of determination R 2 The ability of the Deep-Drain learning mechanism model to simulate peak water levels in one-dimensional stormwater pipe networks and river networks was evaluated, with the coefficient of determination as follows: ； in, To simulate the mechanism model i Peak water levels at individual rainwater wells and river sections, in meters; For Deep-Drain simulation of the first i Peak water levels at individual rainwater wells and river sections, in meters; The mean value (m) of the peak water level in the storm drain well and river cross-section is used to simulate the mechanism model. n It refers to the number of rainwater wells and river cross-sections.
11. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The speedup ratio (Speedup) is used to evaluate the computational efficiency improvement of Deep-Drain relative to the mechanistic model. The speedup ratio is: ； in, T 机理模型 The simulation computation time for the mechanism model; T Deep-Drain This refers to the simulation computation time for Deep-Drain.
12. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The ratio of the training set, validation set, and test set is 0.6, 0.2, and 0.2, respectively.
13. The accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model according to claim 1, characterized in that, The Bayesian optimization process involves 200 iterations.
14. A data-driven alternative model-based accelerated computational device for urban stormwater drainage networks and river networks, characterized in that, The method for accelerating the hydrodynamic calculation of urban stormwater drainage networks and river networks based on a data-driven alternative model, as described in any one of claims 1-13, includes: The model building module is used to build a data-driven alternative model, Deep-Drain, which is built based on a bidirectional long short-term memory network, a graph attention network, and residual connections. It is used to simulate a partially simplified process of urban stormwater pipe network and river network flow. The data processing module is used to build a mechanism model to simulate the one-dimensional hydrodynamic process under measured rainfall and design rainfall with different return periods, and to process and build a dataset of stormwater pipe network and river network flow process, while dividing it into training set, validation set and test set; The parameter optimization module is used to find the optimal combination of hyperparameters for the Deep-Drain using Bayesian optimization methods. The performance evaluation module is used to quantitatively analyze the efficiency improvement effect of the Deep-Drain based on typhoon and rainstorm events.
15. A computer device, characterized in that, The device includes a memory and a processor, which are communicatively connected to each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the steps of the accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model, as described in any one of claims 1 to 13.
16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the accelerated calculation method for urban stormwater drainage networks and river networks based on a data-driven alternative model, as described in any one of claims 1 to 13.