Urban drainage system multi-target prediction method based on SWMM and graph convolutional neural network

Abstract

The invention discloses an urban drainage system multi-target prediction method based on an SWMM and a graph convolutional neural network, and the method comprises the steps: obtaining the actual measurement data of a target urban drainage system pipeline and an inspection well, constructing an SWMM model, and outputting the data; generating a space-time diagram sequence deep learning network GraphSAGE-GRU model on the basis of a graph convolutional neural network GraphSAGE and a gated cycle unit network GRU; and taking the preprocessed data as input, inputting real-time or predicted rainfall data by utilizing the trained model, and synchronously outputting node water head and pipeline load prediction results of the target urban drainage system. According to the method, the problem that the topological structure of an urban drainage system is not considered in a traditional method can be solved, and the hydraulic attributes of the inspection well and the pipeline can be comprehensively output. The problem that a traditional agent model is too black and lacks structural information is solved, and more simulation result output is provided by setting a space-time diagram structure.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrological models, and particularly relates to a multi-objective prediction method for urban drainage systems based on SWMM and graph convolutional neural networks. Background Technique

[0002] The operation process of urban drainage systems is comprehensively affected by various factors such as rainfall intensity, terrain features, pipe network structure, and land use types, making its modeling and prediction a complex non-linear problem. Traditional urban drainage system modeling methods usually rely on physically-driven hydro-hydraulic models, such as SWMM (Storm Water Management Model), etc. These models can relatively accurately describe the dynamic behavior of drainage systems by simulating processes such as rainfall runoff, pipe flow, and pump station operation. However, the construction and operation of physical models require a large amount of parameter calibration and high-precision input data. At the same time, when facing complex urban drainage networks, the computational efficiency is low, making it difficult to meet the needs of real-time prediction and scheduling.

[0003] In recent years, with the rapid development of data-driven methods, machine learning-based surrogate models have gradually become an important direction for urban drainage system modeling. These surrogate models can efficiently replace traditional physical models for prediction and simulation by using historical data and machine learning algorithms. For example, methods such as support vector machine (SVM), extreme learning machine (ELM), and long short-term memory network (LSTM) have been widely applied to the modeling of rainfall-drainage processes. These methods have to a certain extent solved the problem of high computational complexity of traditional physical models. However, to build an efficient surrogate model for urban drainage systems, it is necessary to fully consider the multi-time scale dynamic characteristics and spatial heterogeneity of drainage systems, and at the same time combine the prior knowledge of physically-driven models to improve the prediction accuracy and generalization ability of the model. This research direction can not only provide an efficient tool for the real-time simulation and scheduling of urban drainage systems, but also provide a scientific basis for urban flood control and disaster reduction and water resource management.

[0004] At present, deep learning methods have achieved extremely successful applications in many fields such as image recognition, speech recognition, and driverless driving. At the same time, they are also widely used in the construction of surrogate models for urban drainage systems. In 2019, She et al. "A Dynamic Flow Forecast Model for Urban Drainage Using the Coupled Artificial Neural Network" [J]. Water Resources Management, 2019, 33(9): 3143 - 3153. developed a coupled neural network called the RBF-NARX prediction model (RNFM) for predicting urban drainage outflows; in 2021, Seyedashraf et al. "Disaggregation-Emulation Approach for Optimization of Large Urban Drainage Systems" [J]. Water Resources Research, 2021, 57(8):

[0005] e2020WR029098. Simulated a part of the drainage network while representing the rest through a surrogate model to reduce the optimization cost; in 2021, Lu et al. "Surrogate Global Optimization for Identifying Cost-Effective Green Infrastructure for Urban Flood Control With a Computationally Expensive Inundation Model" [J]. Water Resources Research, 2022, 58(4):

[0006] e2021WR030928. Using proxy optimization method to reduce the computational consumption of two-dimensional urban flood inundation model; In 2023, Luo et al. "Machine learning-based surrogate model assisting stochastic model predictive control of urban drainage systems" [J]. Journal of Environmental Management, 2023, 346: 118974. Developed a machine learning-based surrogate model with rainfall inflow and control as inputs and system overflow as output to quickly evaluate system performance. Existing research mainly focuses on methods such as linear models, artificial neural networks, and long short-term memory network (LSTM) deep learning. However, existing research ignores the topological structure of complex urban drainage systems, lacks physical information, and can only proxy partial simulation results. Summary of the Invention

[0007] The purpose of the present invention is to provide a multi-objective prediction method for urban drainage systems based on SWMM and graph convolutional neural network, to solve the problems in the prior art that most research lacks topological structure information of drainage systems and has too little proxy information and lacks comprehensiveness.

[0008] To solve the above technical problems, the present invention is implemented as follows:

[0009] The multi-objective prediction method for urban drainage systems based on SWMM and graph convolutional neural network includes the following steps:

[0010] Step S1: Obtain the measured data of the pipes and inspection wells in the target urban drainage system, use the obtained data to construct the SWMM stormwater management model of the target city, and abstract the drainage system into undirected graph data.

[0011] Abstracting the drainage system into undirected graph data specifically means regarding each node in the drainage system (such as pipe intersections, drainage outlets, etc.) as the nodes of the graph, and the connections between nodes (such as pipe connections) as edges, thereby constructing a graph structure.

[0012] Step S2: Collect the long-time-scale rainfall sequences in the area of the target urban drainage system, input them into the SWMM model for simulation, generate the lateral inflow data, head sequence data, and pipe load sequence data beside all inspection wells in the system, and preprocess the data.

[0013] The head data and pipeline load data reflect the operating status of the urban drainage system during the observation period. The head data reflects the change of water level in the drainage pipeline, and the pipeline load data shows the utilization of the pipeline's carrying capacity. The lateral inflow data is crucial for understanding the water volume change of the entire drainage system.

[0014] Step S3: Construct a spatio-temporal graph sequence deep learning network GraphSAGE-GRU model based on the Graph Convolutional Neural Network (GraphSAGE) and the Gated Recurrent Unit Network (GRU), and process the above data, specifically including:

[0015] Step S3.1: For the head data and pipeline load data after preprocessing the time series during the observation period of the SWMM model in Step S2, and the lateral inflow data during the observation period and prediction period, use the Graph Convolutional Neural Network (GraphSAGE) for the first processing to aggregate the feature attributes between each node and its neighbor nodes at all levels, generate a new representation for each node, and obtain graph sequence data. Specifically, a certain number of nodes are sampled from the neighbor nodes of each node, and then the features of these sampled nodes are aggregated (such as average, sum, or pooling operations) to obtain a more representative feature representation of the node. After this step of processing, graph sequence data is obtained, which integrates the information of the node and its neighbor nodes and can better reflect the relative position and role of the node in the entire drainage system.

[0016] Then, input the graph sequence data into a new layer of the Graph Convolutional Neural Network (GraphSAGE) for the second processing to aggregate the feature attributes of nodes at a greater graph distance and obtain three-dimensional tensor data; that is, considering the broader neighbor information of the nodes, further explore the potential features and relationships in the data.

[0017] Step S3.2: Flatten the three-dimensional tensor data in Step S3.1 and input it into the Gated Recurrent Unit Network (GRU) for training to simulate the node head and pipeline load values and obtain two-dimensional tensor data.

[0018] Step S3.3: Input the two-dimensional tensor data into the fully connected layer and re-stack it into graph sequence data to adapt to the input format for subsequent training or prediction.

[0019] Step S4: Randomly select multiple rainfall sequences collected in Step S2 to form a dataset, and divide the dataset into a training set and a validation set according to a ratio of 7:3. Each rainfall time period in the training set is not less than 2 hours; use the training set to train the spatio-temporal graph sequence deep learning network GraphSAGE-GRU model, and use the validation set to verify the trained model; use the mean square loss function for target optimization and add the physical constraint of the head height to improve the physical rationality of the model.

[0020] Step S5: Based on the trained model, input real-time or predicted rainfall data, and synchronously output the predicted results of the node water head and pipeline load of the target city's drainage system to support the operation monitoring and scheduling decision-making of the drainage system.

[0021] For further optimization, in step S1, the inspection well data includes the location and burial depth, and the pipeline data includes the starting connection well number, the ending connection well number, the starting inlet offset, the ending inlet offset, the pipeline length, and the pipeline geometric shape parameters.

[0022] For further optimization, in step S2, the operation of the SWMM model includes:

[0023] Step S2.1: Input the rainfall sequence data, and determine the inflow range received by each inspection well according to the Thiessen polygon method. The specific steps are as follows:

[0024] 1) Obtain the geographical coordinates of each inspection well;

[0025] 2) Use geographic information system (GIS) software (such as ArcGIS) or programming languages (such as the scipy.spatial.Voronoi function in Python) to generate Thiessen polygons based on the location data of the inspection wells. Each Thiessen polygon represents the inflow range received by an inspection well.

[0026] 3) Divide the urban drainage area into multiple sub-basins, and determine the inspection wells corresponding to each sub-basin according to the range of the Thiessen polygons.

[0027] Step S2.2: Run the SWMM model and record the output result sequence file.

[0028] For further optimization, in step S2, the data preprocessing includes normalizing the lateral inflow data and node water head data.

[0029] For further optimization, in step S3, the spatio-temporal graph sequence deep learning network GraphSAGE-GRU model is as follows:

[0030]

[0031] In the formula, is the predicted value of the inspection well water head, is the predicted value of the pipeline load, H is the observed value of the inspection well water head, c is the observed value of the pipeline load, Q is the lateral inflow value of the inspection well, G is the undirected graph of the drainage system topology, o is the duration of the observation period, p is the duration of the prediction period, and f is the spatio-temporal graph sequence deep learning network GraphSAGE-GRU.

[0032] For further optimization, in step S3.1, the characteristic attributes of the aggregation node include the water head value of the node inspection well, the side inflow value of the node inspection well, and the load value of the connecting pipeline of the node inspection well.

[0033] For further optimization, in step S3.1, the aggregation formula of the graph convolutional neural network GraphSAGE is as follows:

[0034]

[0035] In the formula, x i ′ is the aggregated feature value of node i, x i is the original feature value of node i, is the average aggregation value of the j - order neighbor feature attributes of node i, and W1 and W2 are learnable weight matrices.

[0036] For further optimization, in step S3.1, the three dimensions of the three - dimensional tensor include the time - series length, the number of graph nodes, and the graph node feature dimension.

[0037] For further optimization, in step S3.2, the gated recurrent unit network GRU is calculated according to the following formula:

[0038] Reset gate:

[0039] R t = σ(X t W xr + H t-1 W hr + b r )#(3)

[0040] Update gate:

[0041] Z t = σ(X t W xz + H t-1 W hz + b z )#(4)

[0042] Candidate hidden state:

[0043]

[0044] Hidden state:

[0045]

[0046] In the formula, x t is the input, H t is the hidden state, is the candidate hidden state, R t is the reset gate, Zt For the update gate, tanh is the non-linear activation function, ⊙ is the Hadamard product, and b r , b z , b h is the bias parameter, and W xr , W hr , W xz , W hz , W xh , W hh is the weight parameter.

[0047] For further optimization, in step S3.2, the two-dimensional tensor data refers to flattening the three-dimensional tensor data in S3.1, where the time series dimension remains unchanged.

[0048] For further optimization, in step S4, the calculation formula of the loss function is as follows:

[0049]

[0050] In the formula, Loss is the loss function, and MSE H is the mean square error of the simulated node head, and MSE C is the mean square error of the simulated pipeline load. T is the length of the prediction period, J is the number of nodes, and K is the number of pipelines.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] 1. Through the multi-objective prediction method for urban drainage systems based on SWMM and graph convolutional neural networks, the present invention can not only make up for the problem that traditional methods do not consider the topological structure of urban drainage systems, but also comprehensively output the hydraulic properties of inspection wells and pipelines.

[0053] 2. Compared with traditional methods, the method of the present invention solves the problem that traditional surrogate models are too black-box and lack structural information, and provides more simulation result outputs by setting the spatio-temporal graph structure.

[0054] 3. The method of the present invention can be widely applied to the rapid simulation of urban drainage systems, and can systematically and completely complete the simulation and prediction of inspection well water levels and pipeline loads, providing a basis for scientific decision-making. Description of the Drawings

[0055] Figure 1 is the flow chart of the multi-objective prediction method for urban drainage systems based on SWMM and graph convolutional neural networks according to the present invention;

[0056] Figure 2 is the schematic diagram of the graph neural network GraphSAGE-GRU model according to the present invention;

[0057] Figure 3 It is a result index chart of simulating and checking the well head and pipeline load of the drainage system agency model in a certain area of Yueyang on July 8, 2020. Specific implementation mode

[0058] The technical solution of the present invention will be described in detail below in conjunction with the embodiments, but the protection scope of the present invention is not limited to the described embodiments.

[0059] Embodiment 1:

[0060] The urban drainage system used in the case study is located in the central area of Yueyang, the second largest city in Hunan Province, China. The originally designed recurrence period of the drainage system is 3 years, the coverage area is 204 hectares, including 28 sub-watersheds, 276 pipelines and 272 inspection wells.

[0061] In this embodiment, as Figure 1 shown, the multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network specifically includes the following steps:

[0062] Step 1: Obtain the measured data of the pipelines and inspection wells in the Yueyang regional drainage system. The inspection well data includes the location and burial depth, and the pipeline data includes the serial number of the starting connection well, the serial number of the ending connection well, the offset of the starting entrance, the offset of the ending entrance, the pipeline length and the pipeline geometric shape parameters. According to the measured data of the pipelines and inspection wells, construct a SWMM rainwater management model, and abstract the drainage system data into undirected graph data. Due to space limitations, only part of the data is taken for display, such as the drainage system node attribute table, as shown in Table 1 specifically.

[0063] Table 1 Drainage system node attribute table

[0064]

[0065]

[0066] Step 2: Collect the annual rainfall sequence data of the Yueyang regional drainage system in 2014 with a resolution of 5 minutes, and input it into the SWMM model for simulation to generate the lateral inflow data, water head sequence data and pipeline load sequence data beside all inspection wells in the system, specifically as follows:

[0067] Step 2.1: Input the rainfall sequence data, and determine the inflow range received by each inspection well according to the Thiessen polygon method.

[0068] Due to space limitations, only part of the data is taken for display, such as the rainfall P monitored from 13:00 to 18:00 on June 23, 2014, as shown in Table 2 specifically.

[0069] Table 2 Rainfall monitored from 13:00 to 15:00 on June 23, 2014 in the implementation area of Yueyang

[0070]

[0071]

[0072] Step 2.2: Run the SWMM model and record the output result sequence file.

[0073] Due to space limitations, only partial data is presented, such as the lateral inflow, node head, and total inflow data beside node J224 from 13:00 to 21:00 on June 23, 2014, as shown in Table 3 specifically.

[0074] Table 3 Lateral Inflow, Node Head, and Total Inflow beside Node J224 from 13:00 to 21:00 on June 23, 2014 in the Yueyang Example Area

[0075]

[0076]

[0077]

[0078] Generate the lateral inflow data, head sequence data, and pipeline load sequence data for all inspection wells in the system, and perform normalization processing on the lateral inflow data and node head data.

[0079] Step 3: Construct a deep learning network for empty graph sequences, the GraphSAGE-GRU model, based on the Graph Convolutional Neural Network GraphSAGE and the Gated Recurrent Unit Network GRU:

[0080]

[0081] In the formula, is the predicted value of the inspection well head, is the predicted value of the pipeline load, H is the observed value of the inspection well head, c is the observed value of the pipeline load, Q is the lateral inflow value of the inspection well, G is the undirected graph of the drainage system topology, o is the duration of the observation period, p is the duration of the prediction period, and f is the spatio-temporal graph sequence deep learning network GraphSAGE-GRU.

[0082] Process the above data based on this model, specifically including:

[0083] Step 3.1: For the head and load data after preprocessing the time series during the observation period of the SWMM model, and the lateral inflow data during the observation period plus the prediction period, perform the first processing using the Graph Convolutional Neural Network GraphSAGE to aggregate the characteristic attributes between each node and its neighbors at all levels, including the head value of the node inspection well, the lateral inflow value of the node inspection well, and the load value of the pipeline connected to the node inspection well. The formula for GraphSAGE to aggregate nodes is as follows:

[0084]

[0085] where x i ′ is the eigenvalue of node i after aggregation, and x i is the original eigenvalue of node i, is the average aggregation value of the j-th order neighbor feature attributes of node i, and W1 and W2 are learnable weight matrices.

[0086] Then, the obtained graph sequence data is input into a new layer of the graph convolutional neural network GraphSAGE for the second processing to aggregate the node feature attributes at a greater graph distance, obtaining three-dimensional tensor data. The three dimensions include the time series length, the number of graph nodes, and the graph node feature dimension.

[0087] Step 3.2: Flatten the three-dimensional tensor data in Step S3.1 and input it into the gated recurrent unit network GRU for training to simulate the node head and pipeline load value, obtaining two-dimensional tensor data. The two-dimensional tensor data refers to flattening the three-dimensional tensor data in Step 3.1, where the time series dimension remains unchanged.

[0088] The gated recurrent unit network GRU is calculated according to the following formula:

[0089] Reset gate:

[0090] R t = σ(X t W xr + H t-1 W hr + b r ) #(3)

[0091] Update gate:

[0092] Z t = σ(X t W xz + H t-1 W hz + b z ) #(4)

[0093] Candidate hidden state:

[0094]

[0095] Hidden state:

[0096]

[0097] where X t is the input, H t is the hidden state, is the candidate hidden state, and Rt For the reset gate, Z t For the update gate, tanh is the non - linear activation function, ⊙ is the Hadamard product, b r , b z , b h For the bias parameter, W xr , W hr , W xz , W hz , W xh , W hh are the weight parameters.

[0098] Step 3.3: Input the two - dimensional tensor data in Step S3.2 into the fully - connected neural network and re - stack it into graph sequence data. The specific shape of the deep - learning network formed by the above steps is shown in Table 4.

[0099] Table 4 Input - Output Table of the GraphSAGE - GRU Model of the Deep - Learning Network

[0100]

[0101]

[0102] Through the co - design of graph convolution and LSTM, this architecture realizes the end - to - end joint modeling of spatio - temporal graph data. While maintaining lightweight parameters, it effectively fuses multi - scale node features and edge features, can synchronously output the dynamic predictions of nodes and edges, and takes into account both computational efficiency and multi - task performance.

[0103] Step 4: Randomly select the rainfall in a time period not shorter than 2 hours in Step S2 as the training set until the total duration ratio reaches 70% of the total rainfall sequence data. Use the training set to train the spatio - temporal graph sequence deep - learning network GraphSAGE - GRU model; use 30% of the total rainfall sequence data to verify the trained model. Use the mean - square loss function for optimization and add the physical constraint of the water - head height. The calculation formula of the loss function is as follows:

[0104]

[0105] In the formula, Loss is the loss function, MSE H is the mean - square error of simulating the node water - head, MSE C is the mean - square error of simulating the pipeline load, T is the length of the prediction period, J is the number of nodes, and K is the number of pipelines. The change of the loss function during the training process is shown in Table 5.

[0106] Table 5 Table of the Change of the Loss Function with the Training Batches

[0107]

[0108]

[0109]

[0110] Step 5: Use the spatio-temporal graph sequence deep learning network model trained in Step 4 to predict the node water head and pipeline load in the Yueyang area simultaneously.

[0111] The evaluation index of the simulation accuracy of water head or load is the Nash efficiency coefficient NSE, which is calculated according to Equation (8):

[0112]

[0113] where Q obs is the model simulation value, Q sim is the measured value, is the average of the measured values, and N is the sequence length.

[0114] Table 6 and Table 7 respectively show the test set data of the water head of the inspection well and the pipeline load in the Yueyang area simulated by the multi-objective prediction method of the urban drainage system based on SWMM and graph convolutional neural network. Limited by space, only partial data are shown.

[0115] Existing methods ignore the topological information existing in the urban drainage system, over-generalize the model, and can only output partial information. For example, the surrogate model predicts the flow near a single drainage outlet. The method described in the present invention uses a graph neural network to utilize the topological information of the urban drainage system and combines LSTM to utilize the time-direction information, so as to be able to output the hydraulic information of all nodes and pipelines in the urban drainage system simultaneously. It can be seen from the result table that the method described in the present application can simultaneously output the predicted values of the water head and load of 272 nodes and 276 pipelines, with good results, providing more comprehensive information for the monitoring and management of the urban drainage system.

[0116] Table 6 Indexes of the test set data of the inspection well water head surrogate model in the Yueyang area

[0117] Inspection well serial number NSE index value J0 0.998 J1 0.996 J2 0.995 J3 0.995 J4 0.992 J5 0.989 J6 0.994 J7 0.923 J8 0.841 J9 0.974 J10 0.980 J11 0.867 J12 0.995 J13 0.994 J14 0.993 J15 0.995 J16 0.991 J17 0.990 J18 0.990 J19 0.992

[0118] Table 7 Indexes of the test set data of the pipeline load surrogate model in the Yueyang area

[0119] Pipeline serial number NSE index value C0 0.998 C1 0.415 C2 0.681 C3 0.996 C4 0.981 C5 0.988 C6 0.987 C7 0.973 C8 0.850 C9 0.733 C10 0.794 C11 0.990 C12 0.975 C13 0.743 C14 0.915 C15 0.862 C16 0.826 C17 0.925 C18 0.935 C19 0.873

[0120] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network, characterized by: The steps include: Step S1: Obtain measured data of the target city’s drainage system pipes and inspection wells, use the acquired data to build the target city’s SWMM stormwater management model, and abstract the drainage system into undirected graph data; Step S2: collect long-term rainfall sequences in the drainage system area of ​​the target city, input them into the SWMM model of the target city for simulation, generate the side inflow data, head sequence data and pipeline load sequence data of all inspection wells in the drainage system of the target city, and pre-process the data; Step S3: Based on the graph convolutional neural network GraphSAGE and the gated recurrent unit network GRU, a spatiotemporal graph sequence deep learning network GraphSAGE-GRU model is constructed, and the above data is processed, specifically including: Step S3.1: The head data and pipeline load data after time series preprocessing in the observation period of the SWMM model in step S2, as well as the lateral inflow data in the observation period and the prediction period, are processed for the first time using the graph convolutional neural network GraphSAGE to aggregate the characteristic attributes between each node and neighboring nodes at all levels, generate a new representation of each node, and obtain graph sequence data; then the graph sequence data is input into a new layer of graph convolutional neural network GraphSAGE for the second processing, aggregate the characteristic attributes of nodes farther away from the graph, and obtain three-dimensional tensor data; Step S3.2: Flatten the three-dimensional tensor data and input it into the gated recurrent unit network GRU for training to simulate the node water head and pipeline load values ​​to obtain two-dimensional tensor data; Step S3.3: Input the two-dimensional tensor data into the fully connected layer and restack it into graph sequence data to adapt to the input format of subsequent training or prediction; Step S4: randomly select multiple rainfall sequences collected in step S2 to form a data set, and divide the data set into a training set and a validation set in a ratio of 7:3, wherein each rainfall period in the training set is not less than 2 hours; use the training set to train the spatiotemporal graph sequence deep learning network GraphSAGE-GRU model, and use the validation set to verify the trained model; use the mean square loss function to perform target optimization, and add a physical constraint on the water head height; Step S5: Based on the trained model, real-time or predicted rainfall data is input, and the node head and pipeline load prediction results of the target city drainage system are synchronously output to support drainage system operation monitoring and scheduling decisions.

2. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 1 is characterized in that: In step S1, the inspection well data includes the location and burial depth, and the pipeline data includes the starting point connection well number, the end point connection well number, the starting point entrance offset, the end point entrance offset, the pipeline length and the pipeline geometric shape parameters.

3. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 2 is characterized in that: In step S2, the SWMM model operation includes: Step S2.1: Input rainfall sequence data and determine the flow range of each inspection well according to the Thiessen polygon method; Step S2.2: Run the SWMM model and record the output result sequence file.

4. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 3 is characterized in that: In the step S2, data preprocessing includes normalizing the lateral inflow data and the node water head data.

5. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 4 is characterized in that: In step S3, the spatiotemporal graph sequence deep learning network GraphSAGE-GRU model is as follows: In the formula, is the predicted value of the inspection well head, is the predicted value of pipeline load, H is the observed value of manhole head, c is the observed value of pipeline load, Q is the inflow value beside the manhole, G is the undirected graph of drainage system topology, o is the observation period, p is the prediction period, and f is the spatiotemporal graph sequence deep learning network GraphSAGE-GRU.

6. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 5 is characterized in that: In the step S3.1, the characteristic attributes of the aggregated node include the water head value of the node inspection well, the lateral inflow value of the node inspection well, and the load value of the pipeline connected to the node inspection well.

7. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 6 is characterized in that: In step S3.1, the graph convolutional neural network GraphSAGE aggregation formula is as follows: In the formula, x i ′ is the eigenvalue of node i after aggregation, x i is the original eigenvalue of node i, is the average aggregated value of the characteristic attributes of the j-order neighbors of node i, W1 and W2 are learnable weight matrices; The three dimensions of the three-dimensional tensor include time series length, number of graph nodes and graph node feature dimension.

8. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 7 is characterized in that: In step S3.2, the gated recurrent unit network GRU is calculated according to the following formula: Reset the gate: R t =σ(X t W xr +H t-1 W hr +b r )#(3) Update Gate: Z t =σ(X t W xz +H t-1 W hz +b z )#(4) Candidate hidden states: Hidden state: Where, X t is the input, H t is the hidden state, is the candidate hidden state, R t To reset the gate, Z t is the update gate, tanh is the nonlinear activation function, ⊙ is the Hadamard product, b r , b z , b h is the bias parameter, W xr , W hr , W xz , W hz , W xh , W hh is the weight parameter.

9. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 8 is characterized in that: In step S3.2, the two-dimensional tensor data refers to flattening the three-dimensional tensor data in S3.1, wherein the time series dimension remains unchanged.

10. The multi-objective prediction method for urban drainage system based on SWMM and graph convolutional neural network according to claim 9 is characterized in that: In step S4, the loss function calculation formula is as follows: In the formula, Loss is the loss function, MSE is H is the mean square error of the simulated node water head, MSE C is the mean square error of the simulated pipeline load, T is the length of the prediction period, J is the number of nodes, and K is the number of pipelines.

Images