Urban drainage system randomness control method and system based on LSTM proxy model

Abstract

The invention discloses an urban drainage system randomness control method and system based on an LSTM proxy model. According to the method, a control period node overflow evaluation index sample set is generated by using a random generation control strategy based on an SWMM urban drainage pipe network model and a historical and designed rainstorm sequence; then, an LSTM model is adopted, through rainstorm input, a control strategy learns node overflow evaluation indexes, and an evaluation index agent model of an SWMM model based on machine learning is trained; furthermore, according to the uncertainty of forecast rainfall, a random rainstorm history is generated, and the forecast rainfall probability is reduced according to a back substitution reduction method to serve as a rainfall random input scene of random prediction control. A proxy model is used to quickly optimize rainfall input scenes and control strategy generation evaluation indexes, and real-time optimization decision control considering rainstorm inflow uncertainty is realized based on a rolling optimization framework.

Description

Technical Field

[0001] The present invention relates to the technical field of urban drainage network regulation, and in particular, to a stochastic control method and system for urban drainage systems based on an LSTM surrogate model. Background Art

[0002] In China, the occurrence frequency of extreme heavy precipitation events and the average annual precipitation show an increasing trend. From 1961 to 2022, the average annual precipitation in China increased by 0.8% per decade on average, and the climate risk index continued to rise. At the same time, urban construction has changed the natural hydrological cycle process and exacerbated the impact of rainfall and floods. The real-time control of urban drainage systems aims to maximize the drainage and storage capacity of existing facilities to achieve the purpose of overflow control, providing an effective way to alleviate the current situation of urban flooding. It is difficult for the real-time control of urban drainage systems to balance the fine degree of process simulation and the simulation time consumption. In real-time control, the process simulation of urban drainage is extremely time-consuming. Affected by the uncertainties that generally exist, such as precipitation forecast errors, model structure generalization and parameter optimization, and data observation errors, it is difficult for the real-time control of urban drainage systems to achieve a deterministic optimal control effect. Therefore, efficiently solving the optimal control scheme for urban drainage system simulation and establishing a stochastic control method considering uncertainties are the keys to further understanding the impact of the uncertainties of process simulation and drainage facility states on stochastic predictive control and constructing a real-time control strategy adapted to multi-dimensional uncertainties to improve the robustness of the real-time control of urban drainage systems.

[0003] The real-time control of urban drainage systems is closely related to concepts such as "smart city" and "smart drainage system". A smart city is an emerging concept in which a city develops from a static system to a flexible and dynamic system, and most information within the city can be monitored, transmitted, stored, and ultimately used to promote the real-time and intelligent management of the entire city. Real-time control improves the operating efficiency of existing systems by using real-time monitoring data and numerical modeling, thus endowing the drainage system with "intelligence". Real-time control can effectively reduce the investment demand for increasing the storage capacity of the drainage system, which is an attractive factor in areas with a dense population and limited space for new buildings. Therefore, the feasibility and urgency of the real-time control of urban drainage systems lie in that urban drainage systems are mostly designed according to static rules. However, the entire system operates under dynamic load conditions.

[0004] Random predictive control enables real-time control strategies to adapt to observation or cognitive biases and improve the robustness of real-time control strategies by describing uncertain information or allowing decisions to violate constraints to a certain extent. Similar to risk-based operation in reservoir operation, random predictive control usually requires repeatedly solving process models to evaluate system behavior under different scenarios. Compared with water systems such as rivers and reservoirs, the topological structure and water flow dynamics of urban drainage systems make their process simulation more complex. Introducing uncertainty description in real-time control models of urban drainage systems may lead to overly long optimization time and fail to meet the actual application requirements of real-time control strategies. Machine learning, as an emerging technology, has been applied in the field of stochastic control of joint real-time operation of large reservoir groups. Its ability to efficiently simulate runoff generation and concentration processes through rainfall time series and control strategies has been verified, and it has broad application prospects in the stochastic predictive control of more complex urban drainage pipe networks.

[0005] The existing technologies mainly have the following disadvantages: In existing real-time control technologies for urban drainage pipe networks, physical process models represented by SWMM (Storm Water Management Model) are used to simulate runoff generation and concentration in urban pipe network systems. However, the simulation process of this high-fidelity model is time-consuming, and it is difficult to meet the timeliness requirements of predictive control when facing complex urban pipe network systems or considering random predictions with errors. Another technology has developed a linear surrogate process model by simplifying the non-linear description of the physical process model to reduce computational complexity and improve the timeliness of the stochastic predictive control model. The effectiveness of this method depends on the accurate simulation of the overflow volume in the high-fidelity physical process model by the linear surrogate model. However, when the system exhibits complex non-linear dynamic processes, the applicability and extensibility of the linear surrogate model are difficult to guarantee. Summary of the Invention

[0006] In view of the problem of uncertainty in stormwater inflow in the real-time control of urban drainage systems, the present invention uses random predictive control to seek control strategies that can adapt to different stormwater inflow scenarios. The difficulty lies in how to achieve accurate and fast process simulation that matches the real-time control requirements. To this end, the present invention improves random predictive control by coupling a machine learning surrogate model, and proposes a stochastic control method and system for urban drainage systems based on an LSTM surrogate model.

[0007] According to one aspect of the specification of the present invention, there is provided a stochastic control method for an urban drainage system based on an LSTM surrogate model, including: According to the forecast rainfall data, generate stormwater inflow scenarios by random simulation, and determine representative inflow scenarios and their corresponding probabilities through synchronous back substitution reduction method as the rainfall random input data set for real-time control; Use the trained surrogate model to generate rapid optimization of evaluation indicators, and implement real-time optimization decision control considering the uncertainty of rainstorm inflow based on the rolling optimization framework; Among them, the training of the surrogate model includes: Obtain the topological relationship of the urban drainage system, construct the boundary of the SWMM model of the urban drainage pipe network, and obtain the historical rainstorm sequence of the city; Based on the historical rainstorms and design rainstorm sequences of the city and randomly generated control strategies, run the SWMM model of the urban drainage pipe network to generate a sample set of evaluation indicators; According to the sample set of evaluation indicators, train the LSTM-based surrogate model and output the trained surrogate model.

[0008] As a further technical solution, the method further includes: Verify the effectiveness of the surrogate model based on the predictive control scheme of the SWMM model of the urban drainage pipe network.

[0009] As a further technical solution, the training of the surrogate model further includes: According to the SWMM model, taking rainstorm inflow and gate control strategy as input variables, simulate the overflow volume of the drainage system during the evaluation period; Taking the waterlogging overflow index simulated by the SWMM model as the objective function for optimizing the real-time control model, establish a single-prediction LSTM surrogate model for the waterlogging overflow index.

[0010] As a further technical solution, according to the forecast rainfall data, using stochastic simulation to generate rainstorm inflow scenarios, it further includes: For each rainstorm inflow scenario, assume that its error compared to the real inflow scenario follows a Gaussian distribution with a mean of 0, and at the same time assume that the standard deviation of the error has a linear relationship with the prediction period.

[0011] As a further technical solution, determine the representative inflow scenarios and their corresponding probabilities through the synchronous back substitution reduction method, including: Calculate the probability distance between scenario k and the remaining scenarios; Sort the calculated probability distances and select the pair of scenarios with the smallest probability distance; Delete the scenario with the smallest occurrence probability in this pair of scenarios; Update the occurrence probabilities of the remaining scenarios in this pair of scenarios; Repeat the above process until the size of the new scenario set is equal to the preset number of representative scenarios.

[0012] As a further technical solution, using the trained surrogate model to generate rapid optimization of evaluation indicators, and implementing real-time optimization decision control considering the uncertainty of rainstorm inflow based on the rolling optimization framework, includes: (1) Update the t moment proxy model state using real-time rainfall observation data; (2) Input the forecast rainfall data into the proxy model and use an optimization method to determine the optimal control strategy within the control period t ~ t + m where refers to the number of time steps in the control period; m (3) Execute the control strategy for the ~ t ~ t + 1 time steps. As long as the optimization time required for each control time step is less than the control time step itself, the optimization result can be obtained before the control strategy is executed; (4) Enter the next time step and repeat steps (1) to (3) until the simulation ends.

[0013] According to one aspect of the specification of the present invention, there is provided a stochastic control system for an urban drainage system based on an LSTM proxy model, including: A first main module for generating rainstorm inflow scenarios by stochastic simulation based on forecast rainfall data, and determining representative inflow scenarios and their corresponding probabilities through synchronous back substitution reduction method as a rainfall random input data set for real-time control; A second main module for using the trained proxy model to generate rapid optimization of evaluation indicators and implementing real-time optimization decision control considering the uncertainty of rainstorm inflow based on a rolling optimization framework; Wherein, the training of the proxy model includes: Obtain the topological relationship of the urban drainage system, construct the boundary of the SWMM model of the urban drainage pipe network, and obtain the historical rainstorm sequence of the city; Based on the historical rainstorms and design rainstorm sequences of the city and randomly generated control strategies, run the SWMM model of the urban drainage pipe network to generate a sample set of evaluation indicators; Train the LSTM-based proxy model according to the sample set of evaluation indicators and output the trained proxy model.

[0014] According to one aspect of the specification of the present invention, there is provided a stochastic control device for an urban drainage system based on an LSTM proxy model, including a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the stochastic control method for an urban drainage system based on an LSTM proxy model as described above.

[0015] According to one aspect of the specification of the present invention, there is provided a non-transitory computer-readable storage medium storing computer instructions that cause the computer to execute the stochastic control method for an urban drainage system based on an LSTM proxy model as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The machine learning surrogate model adopted by the present invention can accurately simulate the input-output relationship of the high-fidelity physical process model, and significantly improve the timeliness of process simulation while basically not reducing the real-time control effect.

[0017] (2) The stochastic prediction control model based on the machine learning surrogate model of the present invention can significantly improve the robustness and reliability of the real-time control strategy by systematically considering the influence of the uncertainty of stormwater inflow. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0019] Figure 1 FIG. is a schematic flow chart of a stochastic control method for an urban drainage system based on an LSTM surrogate model provided by an embodiment of the present invention.

[0020] Figure 2 FIG. is a schematic flow chart of the training of an LSTM surrogate model provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention. In addition, the technical features in the various embodiments or individual embodiments provided by the present invention can be combined with each other arbitrarily to form a new technical solution. This combination is not restricted by the order of steps and / or the structural composition mode, but must be based on what can be achieved by those of ordinary skill in the art. When the combination of technical solutions results in contradictions or cannot be achieved, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0022] An embodiment of the present invention provides a stochastic control method for an urban drainage system based on an LSTM surrogate model, as Figure 1As shown in the figure, first, according to the predicted rainfall data, random simulation is used to generate rainstorm inflow scenarios, and the representative inflow scenarios and their corresponding probabilities are determined through synchronous back-substitution reduction method, serving as the rainfall random input data set for real-time control. Then, the trained surrogate model is used to quickly optimize the evaluation index, and based on the rolling optimization framework, real-time optimization decision-making control considering the uncertainty of rainstorm inflow is realized.

[0023] Among them, as Figure 2 shown, the training of the surrogate model includes: S1: Obtain the topological relationship of the urban drainage system, construct the boundary of the SWMM model of the urban drainage pipe network, and obtain the historical rainstorm sequence of the city.

[0024] Verify that the SWMM model of the urban drainage pipe network can be opened using EPA SWMM 5.2. According to the actual scenario, set the gate control time step and the gate opening step, and set its forecast period duration according to the rainfall forecast lead time. When the sample size of historical rainstorm events is insufficient, the Chicago rainfall pattern needs to be introduced, and the return period and rainfall duration of the appropriate rainfall pattern are selected according to the control effect of the surrogate model. The rainstorm sequence should include the date, time, and rainfall amount.

[0025] S2: Based on the historical rainstorms and design rainstorm sequences of the city and the randomly generated control methods, run the SWMM model of the urban drainage pipe network to generate a sample set of evaluation indexes.

[0026] The long short-term memory model (LSTM) is used as the surrogate model of the high-fidelity urban drainage process model to learn the relationship between the rainstorm, gate control input, and overflow index output of the process model. A large number of input-output sample data of the process model are required. Among them, the high-fidelity urban drainage process model refers to a physical process model based on physical principles and mathematical equations, considering factors such as urban terrain, buildings, roads, sewer networks, etc., to describe the flow and discharge process of rainwater in the city. Such models usually use numerical methods to solve continuous equations and discrete equations to simulate the flow of water and the change of water volume. In the embodiment of the present invention, the SWMM high-fidelity urban drainage process model is used as the sample generation model.

[0027] The SWMM model takes the rainstorm inflow and gate control strategy as input variables and simulates the overflow volume of the drainage system during the evaluation period. The expression is as follows:

[0028] In the formula, is the system overflow index during the t-th time period of the evaluation period; P is the number of time periods of past information required to predict the system overflow volume; L is the number of time periods included in the evaluation period; is the rainstorm inflow rate of node i at the t-th time period; This is the control strategy for control point j in the t-th time period.

[0029] The system overflow flow rate in each time period is normalized by the rainstorm inflow rate. The system overflow index ranges from 0 to 1, and its calculation expression is as follows:

[0030] In the formula, represents the waterlogging overflow in the t-th time period; is the rainstorm inflow rate of the system in the t-th time period.

[0031] The present invention uses the waterlogging overflow index simulated by the SWMM model as the objective function for optimizing the real-time control model. A single prediction LSTM model is established for the waterlogging overflow index . In this way, a more accurate and rapid simulation of the waterlogging overflow index can be achieved. By repeatedly generating random control strategies and combining them with the design rainstorm events and historical measured rainstorm data as the input of the SWMM model, data samples for surrogate model training can be obtained. Among them, the historical rainstorm data is not less than 3 events, and the design rainstorms use the Chicago rain type design rainstorms with return periods of 2 years, 3 years, 5 years, 10 years, 20 years and rainfall durations of 2 hours, 3 hours, and 6 hours.

[0032] S3: Through the LSTM machine learning model, use the rainstorm and control sequence set to learn the evaluation index, train the evaluation index surrogate model of the SWMM model based on machine learning, and analyze the simulation accuracy of the machine learning surrogate model.

[0033] By repeatedly generating random control strategies and combining them with the design rainstorm events and historical measured rainstorm data as the input of the SWMM model, data samples for surrogate model training can be obtained. The size of the training data samples can be determined by the relationship between the training data length and the simulation accuracy index of the machine learning surrogate model, that is, if the prediction accuracy of the machine learning surrogate model cannot be significantly improved only by increasing the size of the training data samples, the current sample size is selected. The number of training rounds and the number of running repetitions of the machine learning can be adjusted according to the performance of the machine learning surrogate model.

[0034] After obtaining the trained surrogate model, the process of the present invention embodiment using this surrogate model for random control of the urban drainage system includes: S4: According to the forecast rainfall data, use random simulation to generate rainstorm inflow scenarios, and determine the representative inflow scenarios and their corresponding probabilities through the synchronous back substitution reduction method as the rainfall random input data set for real-time control.

[0035] For each rainstorm inflow scenario, assume the error compared to the true inflow scenario obeys a Gaussian distribution with a mean of 0:

[0036] The standard deviation of the error is related to the length of the prediction period l . In the present invention, assume that the standard deviation of the error is linearly related to the prediction period l :

[0037] where being 0.9, 0.8, and 0.7 respectively indicate that the variation ranges of the rainstorm inflow prediction errors are 10% - 20%, 20% - 30%, and 30% - 40%.

[0038] The synchronous back - substitution reduction method is used to select representative scenarios from the set of original rainstorm inflow scenarios, and calculate the occurrence probability of each representative scenario. The subset of representative scenarios calculated by the synchronous back - substitution reduction method has the smallest probability distance from the original probability distribution. Assume that the probability distance between scenario k and j is , the occurrence probability of scenario k is , and the calculation steps of the synchronous back - substitution reduction method are as follows: (1) Calculate the probability distance between scenario k and the remaining scenarios:

[0039] where is the occurrence probability of scenario k, , are the Euclidean distances between the characteristic vectors representing scenario k and scenario j respectively.

[0040] (2) Sort the probability distances, and then select the pair of scenarios with the smallest probability distance, denoted as scenario r and scenario c:

[0041] (3) Delete the scenario with the smallest occurrence probability among scenario r and scenario c.

[0042] (4) Update the occurrence probabilities of the remaining scenarios in scenario r and scenario c. Taking scenario r as an example:

[0043] (5) Repeat steps (1) - (4) until the size of the new scenario set is equal to the preset number of representative scenarios.

[0044] S5: Use the surrogate model to generate rapid optimization of evaluation indicators, and implement real-time optimization decision control considering the uncertainty of rainstorm inflow based on the rolling optimization framework.

[0045] An optimization algorithm is used to search for the control strategy that makes the system perform best during the control period, and this process is repeated continuously to obtain the control strategy for each control period. Its specific process is divided into four steps: (1) Update the t state of the process model at each moment using real-time rainfall observation data.

[0046] (2) Input the precipitation forecast into the process model, and use the optimization method to determine the optimal control strategy during the control period ( t ~ t + m , m indicating the number of control period segments).

[0047] (3) Execute the control strategy for the t ~ t + 1 period. As long as the optimization time required for each control period is less than the control time step itself, the optimization result can be obtained before the control strategy is executed. In particular, in the first period, the gate openings of all control nodes are set to 1.

[0048] (4) Enter the next period, and repeat steps (1) to (3) until the simulation ends.

[0049] Preferably, the embodiment of the present invention further includes the following steps: S6: Take the predictive control scheme based on the SWMM model of the urban drainage network as a benchmark to verify the effectiveness of the machine learning surrogate model.

[0050] By comparing the overflow volume and operation time of the nodes of the urban drainage network simulated by the SWMM model of the urban drainage network and the machine learning surrogate model through the above steps, the effectiveness of the surrogate model is verified.

[0051] In summary, in the embodiments of the present invention, first, based on the SWMM urban drainage network model and historical and design rainstorm sequences, a random generation control strategy is used to generate a sample set of node overflow evaluation indexes for the control period. Then, an LSTM model is adopted to learn the node overflow evaluation indexes through rainstorm input and control strategies, and an evaluation index surrogate model of the SWMM model based on machine learning is trained. Further, according to the uncertainty of the predicted rainfall, a random rainstorm process is generated, and the probability of the predicted rainfall is reduced according to the back substitution reduction method as the random input scenario of the random predictive control for rainfall. The surrogate model is used to quickly optimize the evaluation indexes for the rainfall input scenario and the control strategy, and a real-time optimization decision-making control considering the uncertainty of rainstorm inflow is realized based on the rolling optimization framework. The embodiments of the present invention aim at the contradiction between the fine degree of process simulation and the simulation time consumption, and use a random predictive control model based on a machine learning surrogate model to achieve the goal of improving the robustness of the real-time control strategy under the condition of uncertain rainstorm inflow.

[0052] The implementation basis of each embodiment of the present invention is realized through programmed processing by a device with a processor function. Therefore, in engineering practice, the technical solutions and their functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, an embodiment of the present invention provides a stochastic control system for an urban drainage system based on an LSTM surrogate model, and this device is used to execute a stochastic control method for an urban drainage system based on an LSTM surrogate model in the above method embodiments.

[0053] The system includes: a first main module, which is used to generate a rainstorm inflow scenario by random simulation according to the predicted rainfall data, and determine the representative inflow scenario and its corresponding probability through the synchronous back substitution reduction method as the random input data set of rainfall for real-time control; a second main module, which is used to quickly optimize the evaluation indexes by using the trained surrogate model and realize real-time optimization decision-making control considering the uncertainty of rainstorm inflow based on the rolling optimization framework; wherein, the training of the surrogate model includes: obtaining the topological relationship of the urban drainage system, constructing the boundary of the urban drainage network SWMM model, and obtaining the urban historical rainstorm sequence; based on the urban historical rainstorm and design rainstorm sequences and the random generation control strategy, running the urban drainage network SWMM model to generate a sample set of evaluation indexes; and training an LSTM-based surrogate model according to the sample set of evaluation indexes to output the trained surrogate model.

[0054] An embodiment of the present invention provides a stochastic control system for an urban drainage system based on an LSTM surrogate model. Facing the contradiction between the fine degree of process simulation and the simulation time consumption, the foregoing several modules are adopted, and a random predictive control model based on a machine learning surrogate model is used to achieve the goal of improving the robustness of the real-time control strategy under the condition of uncertain rainstorm inflow.

[0055] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The difference lies only in setting corresponding functional modules. The principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art, based on the above system embodiments, refer to the specific technical solutions in other method embodiments, obtain corresponding technical means by combining technical features, and the technical solutions composed of these technical means, and on the premise of ensuring the practicability of the technical solutions, improve the modules in the above system embodiments to obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0056] Based on the same inventive concept as the foregoing embodiments, an embodiment of the present invention further provides a randomness control device for an urban drainage system based on an LSTM proxy model, including a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the randomness control method for an urban drainage system based on an LSTM proxy model.

[0057] Based on the same inventive concept as the foregoing embodiments, an embodiment of the present invention further provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the randomness control method for an urban drainage system based on an LSTM proxy model.

[0058] In summary of the above embodiments, first, the present invention uses multiple scenarios and occurrence probabilities to characterize the uncertainty of rainstorm inflow prediction; then, combines stochastic optimization and a machine learning proxy model to establish a stochastic prediction control model based on the machine learning proxy model; finally, verifies the effectiveness of the proposed model and method through numerical examples. The results show that: (1) The machine learning proxy model can accurately simulate the input-output relationship of a high-fidelity physical process model, and while basically not reducing the real-time control effect, significantly improve the timeliness of process simulation. In the simulation of the input-output relationship, the Nash efficiency coefficient of the machine learning proxy model in the training period, verification period, and test period is greater than 0.99, and the mean absolute error is less than 0.02. As a process model, the application of the machine learning proxy model in deterministic optimal control shows a control effect comparable to that of a high-fidelity physical process model, while reducing the single-step optimization duration with a control time step of 1200 seconds from 1150 seconds to 16 seconds. The stochastic prediction control model established based on the machine learning proxy model achieves a computational time consumption (137 seconds) less than the control time step itself at each control time step, which can meet the actual application requirements of real-time control strategies.

[0059] (2)The stochastic predictive control model based on the machine learning surrogate model can significantly improve the robustness and reliability of real-time control strategies by systematically considering the impact of stormwater inflow uncertainty. Compared with the deterministic predictive control based on the physical process model, the stochastic predictive control based on the machine learning surrogate model can increase the average resilience reflecting the magnitude of system service loss by 0.3 - 0.6 and the average resilience reflecting the duration of system service loss by 0.4 - 0.6 under different forecast error levels (sets of stochastic stormwater inflow scenarios with forecast error ranges of 10% - 20%, 20% - 30%, and 30% - 40%). Accordingly, the average combined sewer overflow volume for the stochastic scenarios corresponding to the three forecast error levels can be reduced by approximately 2500 m 3 , 3000 m 3 and 4500 m 3 . As the forecast error of stormwater inflow increases, the uncertainty of stormwater inflow gradually becomes the main factor affecting the robustness of real-time control strategies, and the stochastic predictive control model demonstrates a greater improvement in system performance.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A stochastic control method for urban drainage systems based on an LSTM surrogate model, characterized in that Including: According to the forecast rainfall data, stochastic simulation is used to generate rainstorm inflow scenarios, and the synchronous back substitution reduction method is used to determine the representative inflow scenarios and their corresponding probabilities, which are used as the rainfall stochastic input dataset for real-time control; The trained surrogate model is used to quickly optimize the evaluation index, and based on the rolling optimization framework, real-time optimization decision-making control considering the uncertainty of rainstorm inflow is realized; Among them, the training of the surrogate model includes: Obtain the topological relationship of the urban drainage system, construct the boundary of the SWMM model of the urban drainage pipe network, and obtain the urban historical rainstorm sequence; Based on the urban historical rainstorm, design rainstorm sequence and randomly generated control strategies, run the SWMM model of the urban drainage pipe network to generate a sample set of evaluation indexes; According to the sample set of evaluation indexes, train the LSTM-based surrogate model and output the trained surrogate model.

2. The stochastic control method for an urban drainage system based on an LSTM proxy model according to claim 1, characterized in that, The method further includes: Taking the predictive control scheme based on the SWMM model of the urban drainage pipe network as a benchmark, verify the effectiveness of the surrogate model.

3. The randomness control method for an urban drainage system based on the LSTM proxy model according to claim 1, wherein, The training of the surrogate model further includes: According to the SWMM model, taking the rainstorm inflow and gate control strategy as input variables, simulate the overflow volume of the drainage system during the evaluation period; Taking the waterlogging overflow index simulated by the SWMM model as the objective function for optimizing the real-time control model, establish a single prediction LSTM surrogate model for the waterlogging overflow index.

4. The stochastic control method for an urban drainage system based on an LSTM proxy model according to claim 1, characterized in that, According to the forecast rainfall data, stochastic simulation is used to generate rainstorm inflow scenarios, and it also includes: For each rainstorm inflow scenario, assume that its error compared with the real inflow scenario follows a Gaussian distribution with a mean of 0, and at the same time assume that the standard deviation of the error is linearly related to the prediction period.

5. The stochastic control method for an urban drainage system based on an LSTM proxy model according to claim 1, characterized in that, Determining the representative inflow scenarios and their corresponding probabilities by the synchronous back substitution reduction method includes: Calculate the probability distance between scenario k and the other scenarios; Sort the calculated probability distances and select the pair of scenarios with the smallest probability distance; Delete the scenario with the smallest occurrence probability in this pair of scenarios; Update the occurrence probabilities of the remaining scenarios in this pair of scenarios; Repeat the above process until the size of the new scenario set is equal to the preset number of representative scenarios.

6. The stochastic control method for urban drainage system based on LSTM proxy model according to claim 1, characterized in that Using the trained surrogate model to quickly optimize the evaluation index, and based on the rolling optimization framework, realizing real-time optimization decision-making control considering the uncertainty of rainstorm inflow includes: (1) Update the t time surrogate model state using real-time rainfall observation data; (2) Input the predicted rainfall data into the surrogate model and use the optimization method to determine the optimal control strategy within the control period t ~ t + m where the optimal control strategy within the control period m refers to the number of time periods in the control period; (3) Execute t ~ t The control strategy for the +1 period. As long as the optimization time required for each control period is less than the control time step itself, the optimization result can be obtained before the control strategy is executed; (4) Enter the next time period, and repeat steps (1) to (3) until the simulation ends.

7. A stochastic control system for urban drainage systems based on an LSTM surrogate model, characterized in that, Including: The first main module is used to generate rainstorm inflow scenarios by stochastic simulation according to the forecast rainfall data, and determine the representative inflow scenarios and their corresponding probabilities by the synchronous back substitution reduction method, which are used as the rainfall stochastic input dataset for real-time control; The second main module is used to quickly optimize the evaluation index using the trained surrogate model, and realize real-time optimization decision-making control considering the uncertainty of rainstorm inflow based on the rolling optimization framework; Among them, the training of the surrogate model includes: Obtain the topological relationship of the urban drainage system, construct the boundary of the SWMM model of the urban drainage pipe network, and obtain the urban historical rainstorm sequence; Based on the urban historical rainstorm, design rainstorm sequence and randomly generated control strategies, run the SWMM model of the urban drainage pipe network to generate a sample set of evaluation indexes; Train an LSTM-based surrogate model according to the sample set of evaluation indicators, and output the trained surrogate model.

8. A randomness control device for urban drainage systems based on an LSTM proxy model, characterized in that, It includes a memory and a processor. The memory stores program instructions executed by the processor. The processor calls the program instructions to execute a method for random control of an urban drainage system based on an LSTM surrogate model according to any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute a method for random control of an urban drainage system based on an LSTM surrogate model according to any one of claims 1 to 6.

Images