A municipal drainage pipe network state prediction method based on a graph and deep learning

Abstract

The application discloses a kind of municipal drainage pipe network state prediction methods based on graph and deep learning, comprising the following steps: collecting target area municipal drainage pipe network topology, and real-time acquisition each monitoring point feature;Using the above information establishes the graph set of space-time sequence;The graph model of space-time sequence established is preprocessed;Graph convolutional neural network and gate recurrent unit are combined to learn the graph data established, and relevant model is trained;The model trained is used to predict the state of drainage pipe network;The prediction result is reconstructed into graph, and each point predicted overflow is visualized and displayed;Finally, the graph model and algorithm prediction result established are used to troubleshoot and prevent flood warning for drainage pipe network system.Application of the present application can reduce the complexity of model establishment, improve the accuracy of drainage pipe network system state prediction, accurately troubleshoot blocked fault pipes and early warning possible disaster areas in the short term, effectively improve the disaster prevention and mitigation capacity of city.

Description

Technical Field

[0001] This invention relates to the field of municipal drainage network condition prediction, and more specifically, to a drainage network condition prediction method based on graph and deep learning. Background Technology

[0002] With the rapid urbanization and continuous advancement of urban infrastructure construction in my country, the impermeable area of ​​urban areas has increased significantly. However, the construction and management of municipal drainage networks, the main flood discharge pathways for cities, have long lagged behind urban development. This has led to increasingly prominent urban flooding and overflow pollution problems in my country. Therefore, there is an urgent need for a method to predict the status of drainage networks, providing early warnings of areas prone to flooding or overflow pollution in the short term, and enabling fault diagnosis of existing drainage network systems to improve urban disaster prevention and mitigation capabilities.

[0003] Current research on methods for predicting the condition of municipal drainage pipe networks is extremely limited. It mainly relies on combining Geographic Information Systems (GIS) and Flood Management Models (SWMM) to use mechanistic modeling for water volume prediction. However, the formation mechanism of stormwater is complex and has a certain degree of randomness, requiring consideration not only of complex and variable climate and meteorological conditions but also of the underlying surface conditions and pipe topology of the target city. Such prediction methods involve complex and cumbersome modeling processes, lack universality, and are difficult to accurately predict the condition of drainage pipe networks, resulting in significant errors. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a graph-based and deep learning-based method for predicting the state of municipal drainage pipe networks. This invention establishes a spatiotemporal graph model of the municipal drainage pipe network in the target area and combines a graph convolutional neural network (GCN) and a gated recurrent unit (GRU) neural network for state prediction. This reduces the parameter information required for model building, improves the accuracy of municipal drainage pipe network state prediction, and analyzes the established graph model and algorithm prediction results to conduct fault diagnosis and flood warning for the drainage pipe network system, effectively improving the city's disaster prevention and mitigation capabilities.

[0005] The objective of this invention is achieved through the following technical solution, including the following steps:

[0006] S1: Collect information on the topology of the municipal drainage network in the target area, and place wirelessly communicating online monitoring devices at each connecting well to obtain real-time information such as liquid level, flow rate, precipitation, and node overflow at the monitoring nodes.

[0007] S2: Using the collected urban drainage network topology information and the sampling data of monitoring nodes at various times, the drainage network system of the target area is modeled as a spatiotemporal sequence graph set G = {G1, G2, ..., G...} t}, where t represents the current time. A graph model for the current time can be established by sampling once at each monitoring node. The graph model structure is shared across different times, but the sampled feature values ​​differ. The graph models from all times are then aggregated to form a graph set.

[0008] Furthermore, taking the current time t as an example, a graphical model G can be established. t ={V,E}, where V = {v1,···,v N} represents the set of graph nodes, i.e., the set of monitoring points, where N is the number of monitoring points, and is represented by a vector. The characteristics of each node are represented by i = 1, ..., N; E represents the set of edges connecting the nodes, that is, the spatial relationship of the pipes connecting the wells, which can be represented as the adjacency matrix of the graph. Where A i,j Representative node v i With v j The quantitative spatial relationship between them.

[0009] Furthermore, the node features when establishing the graph model include, but are not limited to: the current liquid level, flow rate, precipitation, node overflow, and the average liquid level, peak liquid level, liquid level variance, average flow rate, peak flow rate, flow rate variance, average precipitation, peak precipitation, precipitation variance, average node overflow, peak node overflow, node overflow variance, and the start / stop information of drainage pumping stations in the two neighboring node areas near the node. This feature can be represented by 0 and 1. If there are pumping stations in the two neighboring node areas near the node and they are in the open state, then this feature is 1; otherwise, this feature is 0.

[0010] S3: Preprocess each graphical model that is created.

[0011] Furthermore, the feature matrix of the graphical model at time t can be represented as: Where d represents the dimension of each node's feature, a graph normalization method is used for preprocessing, that is, the features of the graph model at each time step are normalized separately. The preprocessing method at time t can be expressed as: in μ represents the normalized eigenvalue matrix. t and σ t Then they represent the characteristic matrix X respectively. t The mean vector and standard deviation vector.

[0012] S4: Combine graph convolutional neural networks and gated recurrent neural networks to learn the preprocessed graph model and train the GCN-GRU model.

[0013] It is worth noting that, since the condition of urban drainage pipe networks is easily affected by seasonal climate, the GCN-GRU model needs to be retrained and iterated every certain period of time to ensure the accuracy of predictions.

[0014] Furthermore, step S4 also includes:

[0015] S401: The preprocessed graph model is input into the constructed GCN to extract spatial information at different time points. The established graph convolutional neural network contains two graph filtering layers and a nonlinear activation function, which can be represented as: H t =h2(A,a1(h1(A,X) t ))), where H t Let represent the node feature matrix containing spatial information at time t, h2(·) and h1(·) represent two graph filtering layers, and a1(·) represent the ReLU activation function;

[0016] S402: The spatial information output by the GCN is constructed into a time series and input into a gated recurrent neural network. This recurrent neural network can predict the overflow of each node in the next T' time steps based on information from the past T time steps. Specifically, it can be expressed as follows: in, Let f(·) represent the node overflow vector predicted at time t+T', and let f(·) represent the gated recurrent neural network model.

[0017] S403: The mean square error (Cost) of each node at each time step is calculated and used as an evaluation metric for the GCN-GRU model. It can be expressed as: Among them, Y i,j and Let $C$ represent the actual overflow and the predicted overflow of node $i$ at time $j$, respectively. During training, the $C$ value should be minimized to obtain the optimal model.

[0018] S5: Use the trained GCN-GRU model to predict the status of the drainage network.

[0019] S6: Calculate the prediction results at each time step. The reconstructed graph visualizes the predicted overflow at each monitoring node at future time points. Furthermore, the predicted reconstructed graph can be represented as follows: The spatial relationship of the predicted reconstructed graph is the same as G, that is, the graph model structure remains unchanged, but the features of each node are replaced with the predicted overflow of the corresponding node at the prediction time.

[0020] S7: Based on the established graphical model and the predicted overflow of each node, troubleshooting and flood prevention early warning are performed on the drainage pipe network system. Furthermore, the initial blockage degree d0 and the pipe blockage degree threshold d are set for each pipe. bIf the average liquid level of a monitoring node changes compared to the previous period under different precipitation conditions in the current cycle, the corresponding output pipe blockage degree d of that monitoring node is updated. This update method can be expressed as: In the formula, d and d' represent the pipe blockage degree of the current cycle and the previous cycle, respectively, K represents the number of precipitation segments, and l' k With Δl k These represent the liquid level at the monitoring node in the previous cycle and the change in liquid level between the current cycle and the previous cycle, respectively, under the precipitation conditions of the k-th segment. When the blockage degree d of a certain pipeline exceeds d... b At the same time, troubleshooting work is carried out on the pipeline. Simultaneously, a node predicted overflow threshold 'b' is set. When the overflow of a node exceeds 'b' at a future time, a timely warning is issued, and the corresponding time, node, and pipeline information associated with the problematic node are output. If the number of nodes in the predicted overflow graph exceeding the threshold is less than the set minimum threshold, troubleshooting is conducted on that node and its associated pipelines. If the number of nodes in the predicted overflow graph exceeding the threshold is greater than the set maximum threshold, flood prevention and drainage work is prepared in advance while troubleshooting is underway.

[0021] The beneficial effects of this invention are:

[0022] Compared to mechanistic models such as SWMM and existing prediction algorithms, this invention employs a graph method to model municipal drainage networks, reducing modeling complexity and parameter requirements. It extracts spatial information between monitoring nodes and between monitoring nodes and pipelines using a graph convolutional neural network (GCN), and combines this with a gated recurrent unit (GRU) neural network to process temporal information, thereby predicting the overflow of each node. This reduces model building complexity, improves the accuracy of drainage network system state prediction, and can accurately identify blocked or faulty pipelines and provide early warnings of areas prone to short-term disasters. Finally, by analyzing the established graph model and algorithm prediction results, fault diagnosis and flood warnings are conducted for the drainage network system, effectively improving the city's disaster prevention and mitigation capabilities. Attached Figure Description

[0023] Figure 1 This is a flowchart of the drainage network status prediction method of the present invention;

[0024] Figure 2 This is a graph model established at time t in a certain region in the example;

[0025] Figure 3 This is a diagram of the GCN-GRU algorithm framework constructed in the embodiment;

[0026] Figure 4 This is a predicted reconstruction map model for a certain region in the embodiment. Detailed Implementation

[0027] To enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and examples.

[0028] See appendix Figure 1 This invention proposes a method for predicting the state of municipal drainage pipe networks based on graph and deep learning, comprising the following steps:

[0029] S1: Collect information on the topology of the municipal drainage network in the target area, and place wirelessly communicating online monitoring devices at each connecting well to obtain real-time information such as liquid level, flow rate, precipitation, and node overflow at the monitoring nodes.

[0030] S2: Using the collected urban drainage network information and sampling data from monitoring nodes at various times, the drainage network system of the target area is modeled as a spatiotemporal sequence graph set G = {G1, G2, ..., G...} t}, where t represents the current time. Each monitoring node performs a unified sampling each time to establish a graph model for the current time. The graph structure is shared across different times, but the sampling characteristics differ. The graph models from all times are aggregated to form a graph set. Furthermore, taking time t as an example, a graph model G can be established. t ={V,E}, where V = {v1,···,v N} represents the set of graph nodes, i.e., the set of monitoring points, where N is the number of monitoring points, and is represented by a vector. The characteristics of each node are represented by i = 1, ..., N. In this embodiment, the node characteristics include: the current liquid level, flow rate, precipitation, node overflow, and the average liquid level, peak liquid level, liquid level variance, average flow rate, peak flow rate, flow rate variance, average precipitation, peak precipitation, precipitation variance, average node overflow, peak node overflow, node overflow variance, and the start / stop information of drainage pumping stations in the two neighboring node areas. This characteristic can be represented by 0 and 1. If there are pumping stations in the two neighboring node areas near the node and they are in the open state, the characteristic is 1; otherwise, the characteristic is 0. E represents the set of edges connecting the nodes, that is, the spatial relationship of the pipes connecting the wells. This relationship can be represented as the adjacency matrix of the graph. Where A i,j Representative node v i With v j The quantitative spatial relationship between them. (Appendix) Figure 2This embodiment presents a graph model established for a certain region at time t, comprising 38 monitoring nodes. Each node contains 17 related features, and the figure only shows the features at node 15 at time t. Further, taking node 22 as an example, its feature 17, namely the associated region of start / stop information of drainage pumping stations within two nearby node areas, includes nodes 22, 6, 21, 23, 5, 7, 13, 16, and 24. That is, if a drainage pumping station exists and is in operation within the aforementioned node areas, then the 17th feature of node 22 is 1; otherwise, it is 0. Additionally, since municipal drainage pipe network systems typically have a fixed drainage direction, the graph model established in this embodiment is directed.

[0031] S3: Preprocess the established spatiotemporal sequence graphical model. Further, the feature matrix of the graphical model at time t can be expressed as... Where d represents the dimension of each node's feature, and in this embodiment, d = 16. A graph normalization method is used for preprocessing, meaning that the features of the graph model at each time step are normalized individually. At time t, this preprocessing method can be expressed as: in μ represents the normalized eigenvalue matrix. t and σ t Then they represent the characteristic matrix X respectively. t The mean vector and standard deviation vector.

[0032] S4: The established graph data is learned using a graph convolutional neural network and a gated recurrent unit (GCN-GRU) to train the GCN-GRU model. It is worth noting that because the condition of urban drainage networks is easily affected by seasonal climate, the GCN-GRU model needs to be retrained and iteratively updated periodically to ensure the accuracy of predictions. (Appendix) Figure 3 The diagram shows the GCN-GRU algorithm framework constructed in this embodiment. In this embodiment, the training set data spans one month, and the sampling interval is 10 minutes. The constructed GCN-GRU model can predict the overflow of each node at each time point in the next 12 hours based on the data from the past two weeks. In this embodiment, the model is retrained every week.

[0033] Furthermore, step S4 also includes:

[0034] S401: The preprocessed graph model is input into the constructed GCN to extract spatial information at different time points. The established graph convolutional neural network contains two graph filtering layers and a nonlinear activation function, which can be represented as: H t =h2(A,a1(h1(A,X) t ))), where H tLet represent the node feature matrix containing spatial information at time t, h2(·) and h1(·) represent two graph filtering layers, and a1(·) represent the ReLU activation function;

[0035] S402: The spatial information output by the GCN is constructed into a time series and input into a gated recurrent neural network. This recurrent neural network can predict the overflow of each node in the next T' time steps based on information from the past T time steps. Specifically, it can be expressed as follows: in, Let f(·) represent the node overflow vector predicted at time t+T', and let f(·) represent the gated recurrent neural network model.

[0036] S403: The mean square error (Cost) of each node at each time step is calculated and used as an evaluation metric for the GCN-GRU model. It can be expressed as: Among them, Y i,j and Let $C$ represent the actual overflow and the predicted overflow of node $i$ at time $j$, respectively. During training, the $C$ value should be minimized to obtain the optimal model.

[0037] S5: Use the trained GCN-GRU model to predict the status of the drainage network.

[0038] S6: Reconstruct the prediction results at each time point into a graph, visually displaying the predicted overflow at each monitoring node at future time points. Further, the predicted reconstruction graph can be represented as follows: The spatial relationships of the reconstructed graph are the same as G, meaning the graph model structure remains unchanged, but the features of each node are replaced with the predicted overflow vector of the corresponding node at the prediction time. (Appendix) Figure 4 This is the prediction reconstruction graph for this embodiment. Each node contains predicted values ​​for the next 72 time points. The graph only shows the predicted features at node 15.

[0039] S7: Based on the established graphical model and the predicted overflow diagrams for each node, fault diagnosis and flood prevention early warning are performed on the drainage pipe network system. Furthermore, the initial blockage degree d0 and the pipe blockage degree threshold d for each pipe are set. b If the average liquid level at a monitoring point changes compared to the previous period under different precipitation conditions in the current cycle, the corresponding output pipe blockage degree d of that monitoring point is updated. This update method can be expressed as: In the formula, d and d' represent the pipe blockage degree of the current cycle and the previous cycle, respectively, K represents the number of precipitation segments, and l' k With Δl k These represent the liquid level at the monitoring point in the previous cycle and the change in liquid level between the current cycle and the previous cycle, respectively, under the precipitation conditions of the k-th segment. When the blockage degree d of a certain pipeline exceeds d... bAt the same time, troubleshooting work is carried out on the pipeline. Simultaneously, a node predicted overflow threshold b is set. When the overflow of a node exceeds b at a certain future time, an early warning is issued and the corresponding time, node, and pipeline information associated with the problematic node are output. If only a few nodes in the predicted overflow diagram have predicted overflows exceeding the threshold, troubleshooting is carried out on the node and its associated pipelines. If a large number of nodes in the predicted overflow diagram have predicted overflows exceeding the threshold, flood prevention and drainage work is prepared in advance while troubleshooting is being carried out. In this example, the initial blockage degree of each pipeline is d0=1, the blockage degree threshold is set to d0=5, one cycle is set to one month, the number of precipitation segments is K=6, and the specific daily precipitation is divided into light rain 0-10mm, moderate rain 10-25mm, heavy rain 25-50mm, rainstorm 50-100mm, heavy rainstorm 100-250mm, and extremely heavy rainstorm 250mm or more according to the precipitation intensity. In addition, in this example, the node overflow threshold b is set to 0, that is, as long as node overflow occurs, relevant fault investigation and flood prevention early warning work will be carried out.

[0040] This invention establishes a spatiotemporal graph model of the municipal drainage network in the target area and combines it with a graph convolutional neural network (GCN) and a gated recurrent unit (GRU) neural network for state prediction. This reduces the parameter information required for model building and improves the accuracy of municipal drainage network state prediction. It can provide early warning of overflow nodes and related pipes. Furthermore, by analyzing the established graph model and algorithm prediction results, it can troubleshoot and provide flood warnings for the drainage network system, effectively improving the city's disaster prevention and mitigation capabilities.

Claims

1. A municipal drainage network state prediction method based on graph and deep learning, characterized in that, The method comprises the following steps: S1: Collecting the municipal drainage network topology information of the target area, and placing online monitoring devices capable of wireless communication at each connection well to obtain the liquid level, flow, precipitation, and node overflow information in real time; S2: using the collected urban drainage network topology information and the sampling data of each time of the monitoring node, modeling the drainage network system of the target area as a set of space-time sequence graphs wherein denotes the current time; S3: Preprocessing each graph model established; S4: Learning the preprocessed graph model in combination with a graph convolutional neural network (GCN) and a gated recurrent unit (GRU), and training a GCN-GRU model; S5: Using the trained GCN-GRU model to predict the state of the drainage network; S6: Reconstructing a prediction reconstruction graph from the prediction results at each time point to visually display the predicted overflow at each monitoring node at the future time point; S7: Combining the established graph model and the predicted overflow of each node to troubleshoot the drainage network system and issue a flood prevention warning; Setting initial clogging degree of each pipe and pipe clogging degree threshold If the average liquid level of a monitoring node under each precipitation condition changes compared with the previous period, update the output pipe clogging degree of the monitoring node The update method can be represented as: Wherein, and respectively represent the pipe clogging degree of the current period and the previous period, represents the number of precipitation segments, and respectively represent the liquid level of the monitoring node in the previous period and the change in the liquid level of the monitoring node in the current period and the previous period under the kth precipitation condition; when the pipe clogging degree exceeds , troubleshoot the pipe; set the node predicted overflow threshold When the overflow of a node at a future time exceeds , timely warning and output the corresponding time, node and pipe information associated with the problem node; if the number of nodes in the predicted reconstruction graph whose predicted overflow exceeds the threshold is less than the set minimum threshold, troubleshoot the node and associated pipes; if the number of nodes in the predicted reconstruction graph whose predicted overflow exceeds the threshold is greater than the set maximum threshold, prepare for flood control and drainage work in advance while troubleshooting.

2. The graph and deep learning based municipal drainage network state prediction method according to claim 1, characterized in that The method for establishing the time-space sequence graph set in step S2 is specifically represented as follows: Each monitoring node is uniformly sampled once to establish the graph model at the current time point, the graph model structures at each time point are shared but the feature values obtained by sampling are different, and the graph models at each time point are summarized to obtain the graph set; Let the current time be The graph model of , where represents a set of graph nodes, i.e., a set of monitoring points, where is the number of monitoring points, and a vector characterizes the features of each node, where ; represents a set of edges connecting the nodes, i.e., the spatial relationship of the pipes connecting the wells to each other, which can be represented as the adjacency matrix of the graph , where represents the quantitative spatial relationship between nodes and .

3. The graph and deep learning based municipal drainage network state prediction method according to claim 1 or 2, characterized in that The step S3 is specifically implemented as follows: Let the graph model feature matrix at time t be where denotes the dimension of each node feature, and the graph normalization method is used for preprocessing, that is, the features of the graph model at each time are normalized separately, The preprocessing at time t is represented as: where denotes the normalized feature matrix, and denotes the mean vector and the standard deviation vector of the feature matrix , respectively.

4. The graph and deep learning based municipal drainage network state prediction method according to claim 3, characterized in that The step S4 is specifically implemented as follows: S401: input the pretreated graph model into the constructed GCN to extract spatial information at different time points; the established graph convolutional neural network comprises two graph filtering layers and a nonlinear activation function, and the process is represented as: wherein represents the node feature matrix containing spatial information at the time point, and represents two graph filtering layers, then represents the ReLU activation function, represents the adjacency matrix; S402: construct the spatial information output by the GCN as a time sequence and input into a gated recurrent neural network, which can predict the overflow of each node at a future time point through information at a past time point, and can be specifically represented as: wherein, represents the node overflow vector predicted at the time point, and represents the gated recurrent neural network model.​​​​​ S403: Calculate the mean square error of each node at each time of the output As the evaluation index of GCN-GRU model, it can be expressed as: Wherein, And Respectively represent the actual overflow and the predicted overflow of the node At time In the training process, the value of Should be as small as possible to obtain the best model.

5. The graph and deep learning based municipal drainage network state prediction method according to claim 4, characterized in that The step S6 is specifically implemented as follows: The predicted reconstructed graph can be represented as The spatial relationship of the predicted reconstructed graph is the same as That is, the graph model structure is unchanged, but the features of each node are replaced by the predicted overflow of the corresponding node at the prediction time.

Images