Urban pipe network operation intelligent prediction method based on physical information graph neural network

Abstract

The invention relates to the field of pipe network operation state prediction, in particular to an intelligent urban pipe network operation prediction method based on a physical information graph neural network, and the method comprises the steps: obtaining monitoring data and topological structure information of a pipe network, and generating time sequence graph data; constructing a prediction model, wherein the prediction model is a mixed model combining a graph neural network and a physical information neural network; performing feature extraction on the time sequence diagram data based on a graph neural network to model dynamic characteristics of a pipe network; embedding the physical constraint into a loss function based on a physical information neural network; generating an operation state prediction result of the pipe network according to the prediction model; and a prediction result is verified, and the model is optimized and adjusted according to verification. The method has the advantages that the pipe network state is accurately predicted, the key parameters are inversed in real time, and reliable support is provided for optimization scheduling and fault diagnosis of the pipe network.

Description

Technical Field

[0001] This application relates to the field of prediction of pipeline network operation status, and particularly to an intelligent prediction method for urban pipeline network operation based on a physics-informed graph neural network. Background Art

[0002] Urban pipeline networks are important infrastructure of modern urban water supply systems, and their operation efficiency directly affects the utilization rate of water resources and the safety of water supply. With the continuous advancement of urbanization, the scale and complexity of water supply pipeline networks are gradually increasing. How to efficiently manage and optimize the operation of pipeline networks has become the core issue of concern in the water service industry. The complexity of pipeline network systems is mainly reflected in two aspects: one is the complex topological structure, with diverse connection relationships between nodes and pipelines, and there may be redundant designs in some areas; the other is the significant dynamic nature of the operation state, and parameters such as node pressure and flow rate are affected by seasonal changes, fluctuations in water demand, and the operating conditions of equipment. This complexity makes it a technical challenge to perform real-time prediction of pipeline network status and invert key parameters (such as pipeline roughness coefficient and leakage point location).

[0003] In the prior art, the modeling and analysis of pipeline networks mainly rely on numerical simulation tools such as EPANET. These tools perform calculations based on classical hydraulics formulas and can provide prediction results of pressure and flow distribution in pipeline networks, which are widely used in the design, operation management, and optimization of water supply systems. However, these simulation tools have significant limitations. First, they have high requirements for the accuracy of input parameters. For example, parameters such as the roughness coefficient of pipelines need to be accurately input to ensure the reliability of simulation results. However, in actual operation, these parameters are often difficult to directly measure and usually rely on empirical estimates or experimental data acquisition, which inevitably introduces errors and thus affects the simulation accuracy. Second, these tools have certain bottlenecks in terms of computational efficiency. Especially when the scale of the pipeline network is large or multiple operating conditions need to be considered, the computational complexity increases significantly, making it difficult to meet the requirements of real-time analysis. This computational limitation makes it difficult for simulation tools to provide a quick response in emergencies (such as pipeline leakage or burst), affecting the timeliness and accuracy of decision-making.

[0004] In recent years, with the development of artificial intelligence technology, data-driven methods have gradually been introduced into the prediction of pipeline network status and parameter inversion. Machine learning models can achieve the prediction of pipeline network operation status through learning historical operation data, which alleviates the dependence on accurate input parameters to a certain extent in traditional methods.

[0005] However, there are also many problems with data-driven methods themselves. First of all, machine learning models are usually based on the principle of empirical risk minimization, and their prediction ability highly depends on the quality and distribution of training data. In actual working conditions, the monitoring data may be noisy, incomplete or have distribution shifts, resulting in limited generalization ability of the model and difficulty in providing reliable prediction results under unseen working conditions. Secondly, existing machine learning methods often lack the constraints of physical laws, which may lead to the phenomenon that the prediction results of the model deviate from reality, such as the predicted values violating physical constraints (such as node flow imbalance or pressure exceeding the actual possible range).

[0006] In addition, the existing methods have insufficient adaptability in dynamic environments. The operating state of the pipe network has significant time dependence, and parameters such as pressure and flow will be dynamically adjusted with changes in external conditions. However, most traditional methods and existing machine learning models assume that the input data is static or quasi-static, and fail to fully capture the dynamic characteristics in the time series, resulting in limited prediction ability of the model in complex environments. For example, when there is a sudden change in water demand, the existing model may not be able to quickly adjust parameters, resulting in a prediction lag behind the actual change.

[0007] Based on the above analysis, the current pipe network modeling and analysis technologies have deficiencies in terms of parameter dependence, real-time performance, physical consistency, and dynamic adaptability. First of all, traditional numerical models rely on high-precision input parameters, which are often difficult to obtain in practice, affecting the accuracy and reliability of the simulation. Secondly, the limitations of existing methods in terms of real-time performance make it impossible to meet the rapid response requirements for emergencies and difficult to adapt to the dynamic changes of complex working conditions. In addition, the lack of embedding of physical laws in data-driven methods leads to inconsistent prediction results with the actual situation in some key scenarios. Finally, the neglect of dynamic characteristics makes the existing models perform poorly in processing time series data and unable to meet the needs of intelligent operation of modern urban pipe networks. Therefore, how to combine the physical laws of the pipe network with data-driven models has become an urgent problem to be solved. Summary of the Invention

[0008] In order to solve the above technical problem of how to combine the physical laws of the pipe network with data-driven models, the present application provides an intelligent prediction method for urban pipe network operation based on physical information graph neural network.

[0009] In the first aspect, the present application provides an intelligent prediction method for urban pipe network operation based on physical information graph neural network, adopting the following technical solutions: An intelligent prediction method for urban pipe network operation based on physical information graph neural network, characterized by including the steps of: Obtain the monitoring data and topological structure information of the pipe network and generate time series graph data; Build a prediction model, which is a hybrid model combining a graph neural network and a physics-informed neural network; extract features from time series graph data based on the graph neural network to model the dynamic characteristics of the pipe network; embed physical constraints into the loss function based on the physics-informed neural network; Generate the prediction results of the operation status of the pipe network according to the prediction model; verify the prediction results and optimize and adjust the model according to the verification; Among them, the loss function is set as the sum of the data fitting loss, the product of the continuity equation loss for constraining the flow conservation of the pipe network nodes and the first weight coefficient, and the product of the energy equation loss for correcting the pressure loss of the pipeline and the second weight coefficient; determine the continuity equation loss according to the flow rate at the nodes of the pipe network and the external water supply or water consumption; calculate the energy equation loss according to the pipeline pressure and pipeline physical parameters.

[0010] Optionally, the continuity equation loss is the sum of the absolute differences between the sum of the flow rates from each node to other nodes and the water consumption of that node.

[0011] Optionally, the energy equation loss is the sum of the absolute differences between the inlet pressure, outlet pressure and frictional loss along the way of each pipeline.

[0012] Optionally, the first weight coefficient and the second weight coefficient are custom parameters, and the first weight coefficient and the second weight coefficient are dynamically adjusted during the training process. The expression of the dynamic adjustment is: ; ; represents the first parameter, represents the second parameter, represents the data fitting loss, represents the continuity equation loss for constraining the flow conservation of the pipe network nodes, represents the energy equation loss for correcting the pressure loss of the pipeline; make the first parameter the adjusted first weight coefficient and the second parameter the adjusted second weight coefficient.

[0013] Optionally, the prediction results include the predicted node pressure, the predicted pipeline flow rate and / or the inversion value of the key parameter.

[0014] Optionally, the method for inverting the key parameter is: compare the difference between the predicted value generated by the prediction model and the monitoring data, and iteratively optimize the estimated value of the key parameter.

[0015] Optionally, the method for generating time-series graph data is as follows: construct a node feature matrix, where the feature of each node is a high-dimensional vector composed of pressure and flow rate; construct an edge feature matrix, where the feature of a pipeline is a high-dimensional vector composed of diameter, length, and roughness coefficient; construct an adjacency matrix to represent the topological structure of the pipe network; integrate the node feature matrix, edge feature matrix, and adjacency matrix to generate time-series graph data and store it in a tensor format.

[0016] Optionally, the topological structure information includes the connection relationship between nodes and pipelines, the length of the pipelines, the diameter of the pipelines, and the roughness coefficient of the pipelines.

[0017] In a second aspect, the present application provides an intelligent prediction system for urban pipe network operation based on a physics-informed graph neural network, adopting the following technical solution: The intelligent prediction system for urban pipe network operation based on a physics-informed graph neural network includes: a data input module, which acquires the monitoring data and topological structure information of the pipe network and generates time-series graph data; a topological feature extraction module, which uses a graph neural network to extract features from the time-series graph data; a physical constraint calculation module, which takes a physics-informed neural network as the core and realizes the constraints on node pressure and pipeline flow rate by introducing hydraulic equations, and the hydraulic equations include a continuity equation and an energy equation; a prediction output module, which generates a predicted value of the operation state and an inversion result of key parameters based on the high-dimensional feature vector output by the physical constraint calculation module; a result verification module, which conducts a rationality check on the results of the prediction output module through a multi-level verification mechanism and marks and feedbacks abnormal situations; the data input module, topological feature extraction module, physical constraint calculation module, prediction output module, and result verification module are communicatively connected for data transmission.

[0018] In a third aspect, the present application provides a terminal, adopting the following technical solution: The terminal includes: a processor and a memory, and the memory stores computer program instructions, which, when executed by the processor, implement the intelligent prediction method for urban pipe network operation based on the above-mentioned physics-informed graph neural network.

[0019] The present application has the following technical effects: 1. By combining a physics-informed neural network (PINN) with a graph neural network (GNN), it is not only possible to accurately predict the state of the pipe network, but also to invert key parameters in real time, providing reliable support for the optimal scheduling and fault diagnosis of the pipe network.

[0020] 2. By introducing the Physics-Informed Neural Network (PINN), the continuity equation and energy equation of the pipe network are embedded in the calculation process of the model, effectively solving the problem of lack of physical constraints in data-driven methods. This embedding method ensures that the prediction results of the model conform to the actual hydraulic laws in terms of flow conservation and pressure distribution, greatly improving the reliability and consistency of the results, especially in the process of parameter inversion.

[0021] 3. Using the Graph Neural Network (GNN) to extract the topological characteristics of the pipe network can efficiently model the complex relationships between nodes and pipes. GNN captures the dynamic characteristics of nodes in the pipe network through multi-layer graph convolution, reflecting the spatial dependence of the pipe network more comprehensively than numerical simulation methods. This design significantly improves the expression ability of complex pipe network states and enhances the adaptability of the model to the expansion of the pipe network scale or the change of working conditions.

[0022] 4. By combining real-time monitoring data and historical data, not only the dynamic response ability of the system is improved, but also the accurate inversion of key parameters (such as pipe roughness coefficient, leakage point location) is achieved. Compared with the existing technologies, the present application not only makes remarkable progress in real-time performance and prediction accuracy, but also can quickly respond to emergencies, providing more reliable technical support for the optimal dispatching and safe operation of urban pipe networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present application will become readily understood. In the drawings, several embodiments of the present application are shown in an exemplary rather than restrictive manner, and the same or corresponding reference numerals represent the same or corresponding parts.

[0024] Figure 1 is a flowchart of a method for intelligent prediction of urban pipe network operation based on the physics-informed graph neural network according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The embodiments of the present application disclose a method for intelligent prediction of urban pipe network operation based on the physics-informed graph neural network. Referring to Figure 1 , it includes steps S1 - S4, specifically as follows: S1: Obtain the monitoring data and topological structure information of the pipe network and generate time series graph data.

[0026] In one embodiment, the monitoring data is sourced from sensors deployed in the pipe network, including node pressure gauges, flow meters, and sensing devices for recording pipe characteristics.

[0027] After the data collected by the sensor is uploaded to a preset system through the network interface, it first undergoes standardization processing by the preprocessing unit. After the preprocessing is completed, the monitoring data is combined with the topological structure of the pipe network to generate time series graph data.

[0028] In one implementation, the steps of preprocessing include: noise filtering to identify and remove noise from the monitoring data. For example, a sliding window filtering algorithm is used to smooth the high-frequency noise of the pressure data. After noise filtering, outlier processing is performed. An outlier detection model is established based on the historical data distribution to detect and correct abnormal data that deviates from the normal range. After outlier processing, missing value imputation is carried out. Interpolation methods can be used to impute the missing data in the time series caused by sensor failures or network transmission interruptions. The interpolation methods include linear interpolation and spline interpolation. The prior art will not be elaborated here.

[0029] In one embodiment, the topological structure information includes the connection relationship between nodes and pipes, the length of the pipes, the diameter of the pipes, and the roughness coefficient of the pipes.

[0030] In one embodiment, the method for generating time series graph data is as follows: construct a node feature matrix, where the feature of each node is a high-dimensional vector composed of pressure and flow rate. The historical time series data composed of pressure and flow rate is the main feature at the node, and other secondary features can also be included. Other secondary features include historical water demand and static elevation, etc.; construct an edge feature matrix, where the feature of the pipe is a high-dimensional vector composed of diameter, length, and roughness coefficient; construct an adjacency matrix to represent the topological structure of the pipe network. Exemplarily, represents node and node are directly connected by a pipe; integrate the node feature matrix, edge feature matrix, and adjacency matrix to generate time series graph data and store it in tensor format for use by subsequent modules.

[0031] S2: Construct a prediction model, where the prediction model is a hybrid model combining a graph neural network and a physics-informed neural network; perform feature extraction on the time series graph data based on the graph neural network to model the dynamic characteristics of the pipe network; embed physical constraints into the loss function based on the physics-informed neural network.

[0032] Use a graph neural network (GNN) to perform feature extraction on the time series graph data. The core of the GNN is to use graph convolution operations to capture the complex relationships between nodes and edges and model the dynamic characteristics of the pipe network.

[0033] Specifically, the design of the graph convolutional network includes the following steps: perform graph convolution operations to aggregate the features of nodes through the adjacency matrix. The formula is: ; where is the node feature of the th layer, is the normalized adjacency matrix, is the layer node feature, is the weight matrix, is the activation function. Edge feature aggregation is performed to capture the influence of edge (pipe) features on node characteristics. An edge feature aggregation module is introduced. After the edge feature vectors are weighted by the graph attention mechanism (GAT), they are combined with adjacent node features in a weighted manner to generate new node features. To describe the dynamic changes in the operation state of the pipe network, a time encoder is introduced to embed time series data into node features, so that the node state of each layer contains time change information for time-dependent modeling.

[0034] This application takes the physics-informed neural network (PINN) as the core and realizes the constraints on node pressure and pipe flow by introducing the hydraulic equation. The specific method is as follows: Design a loss function, and through the automatic differentiation function of the deep learning framework, calculate the gradient of the loss function to optimize the prediction model parameters.

[0035] The expression of the loss function is: ; In the formula, represents the loss function, represents the data fitting loss, represents the first weight coefficient, represents the continuity equation loss for constraining the flow conservation of pipe network nodes, represents the second weight coefficient, represents the energy equation loss for correcting the pressure loss of the pipe; the continuity equation loss is determined according to the flow rate at the nodes of the pipe network and the external water supply or water consumption; the energy equation loss is calculated according to the pipe pressure and pipe physical parameters.

[0036] In one embodiment, the mathematical expression of the continuity equation loss is: ; In the formula, represents the continuity equation loss for constraining the flow conservation of pipe network nodes; represents the node flowing to the node ; represents the node water consumption.

[0037] In other embodiments, since the factor can be used to replace, all are to ensure that at each node, the difference between the total inflow and the total outflow is equal to the water consumption of the node, but different mathematical processing methods are adopted. The subsequent calculation formulas are the same, and will not be elaborated here.

[0038] In one embodiment, the mathematical expression of the energy equation loss is as follows: ; where, represents the energy equation loss for correcting the pressure loss of the pipeline, represents the inlet pressure of the pipeline ; represents the pipeline outlet pressure; represents the pipeline pipe friction coefficient, represents the pipeline length, represents the pipeline flow rate, represents the pipeline diameter, represents the acceleration due to gravity.

[0039] represents the friction loss along the way. The friction loss along the way is the energy loss generated due to the friction of the inner wall of the pipeline when the fluid flows in the pipeline. This loss is related to the friction coefficient of the pipeline, the length of the pipeline, the flow rate of the fluid, the diameter of the pipeline, and the acceleration due to gravity.

[0040] The first weight coefficient and the second weight coefficient are custom parameters, but need to be dynamically adjusted through multiple trainings to make the scales of each loss in the same order of magnitude. The expression for dynamic adjustment is: ; ; represents the first parameter, represents the second parameter, making the first parameter the adjusted first weight coefficient and the second parameter the adjusted second weight coefficient for the dynamic adjustment of the first weight coefficient and the second weight coefficient; Exemplarily, set and to both be 0 initially. After the training is completed, check each loss. For example, = 0.01, , ; In subsequent trainings, , . It is possible to control and to both not exceed . The purpose of dynamically adjusting the weight coefficients is to prevent a certain loss term from dominating during the optimization process and causing other loss terms to be ignored. In this way, the model can more evenly consider various constraints such as data fitting, flow conservation, and pressure loss. Model training is prior art and will not be elaborated here.

[0041] S3: Generate the prediction result of the operation state of the pipe network according to the prediction model.

[0042] The prediction results include predicted values of node pressures, predicted values of pipeline flows, and / or inversion values of key parameters.

[0043] Obtain the intermediate results output during the training of the prediction model (the output of the subsequent physical constraint calculation module). The intermediate results refer to the phased results calculated based on the input data and physical constraint conditions. These results have considered the constraints of physical laws but have not yet undergone final optimization and verification. They are used as the input for subsequent modules to generate the final prediction results.

[0044] In one embodiment, an adaptive weighted fusion algorithm is used to further process the intermediate results (the function of the subsequent prediction output module) to generate the final predicted values of pressures and pipeline flows. Specifically, obtain the intermediate results of pressures and flows from the physical constraint module, and combine other relevant features (such as topological features, time series features, etc.). Through the adaptive weighted fusion algorithm, weights are assigned to different features to generate the final predicted values, and the weighted features are fused to generate the final predicted values. The adaptive weighted fusion algorithm is an algorithm that dynamically adjusts weights. Its purpose is to automatically assign weights according to the importance of features, thereby generating more accurate prediction results. The prior art will not be elaborated here.

[0045] By jointly modeling node features (pressures) and edge features (flows), a dynamic relationship between nodes and pipelines is constructed. Specifically, pressure prediction uses the pipeline energy equation as a constraint condition to correct the pressure loss between nodes; flow prediction is based on the continuity equation to ensure the conservation of the inflow and outflow flows of nodes. For example, for a certain node , the model verification is as follows: ; where is the pressure of node ; is the pressure of node ; Node is adjacent to node ; represents the pipeline friction coefficient; represents the pipeline length, reflecting the distance that the fluid flows in the pipeline; represents the flow rate; represents the pipeline diameter; represents the acceleration due to gravity. This equation ensures that the pressure of node is the pressure of its adjacent node minus the pressure loss in the pipeline. This calculation ensures that the prediction results can not only reflect the actual physical laws but also accurately predict the node states under complex dynamic working conditions.

[0046] Precisely locate leakage points in the pipe network through a flow anomaly detection algorithm. Identify flow anomaly points by comparing the flow distribution with the historical normal operating conditions. Combine the flow anomaly points with the pressure distribution map for analysis, and locate the leakage position through the change in pressure gradient. For example, if the actual pressure gradient of a certain section of the pipeline is significantly higher than the pressure gradient predicted by the model, mark this area as a possible leakage point. In addition, combine the change trend of the pressure distribution in multiple time periods, and further verify the position of the leakage point through time series analysis.

[0047] In one embodiment, the key parameters are the roughness coefficient and / or the position of the leakage. The method for inverting the key parameters is: compare the predicted values generated by the prediction model with the monitoring data, and iteratively optimize the estimated values of the key parameters. Utilize the backpropagation ability of deep learning, combined with the physical constraint calculation of PINN, to invert the key parameters (such as the pipeline roughness coefficient).

[0048] To cope with sudden changes in the operating state of the pipe network (such as a sudden increase in water demand or equipment failure), the technical solution of this application has the ability to dynamically adjust. Specifically, through the update of real-time monitoring data, retrain some model parameters (such as the weights in the loss function), and quickly output the adjusted prediction results. This design ensures the adaptability of the model to sudden operating conditions.

[0049] S4: Verify the prediction results and optimize and adjust the model according to the verification.

[0050] Set up a multi-level verification mechanism, and the multi-level verification mechanism includes: flow conservation verification, pressure range verification, and historical data comparison.

[0051] Conduct flow conservation verification for each node of the prediction results, that is, check whether the total inflow flow and the total outflow flow of the node are equal. If the flow conservation is not satisfied, mark the relevant nodes and pipelines, and feedback this information to the physical constraint calculation module for optimization.

[0052] Conduct rationality verification on the pressure prediction values of all nodes to check whether they are within the preset physical range. For example, the pressure of the urban pipe network usually has upper and lower limits (such as 0.2 MPa to 0.6 MPa). If the predicted pressure of a certain node exceeds this range, mark this node as an abnormal point and further analyze its physical reasons (such as input data error or inaccurate model calculation).

[0053] Compare the prediction results with the historical monitoring data to evaluate their matching degree with the historical operating conditions. For example, under the same water supply conditions, if the difference between the node pressure and the historical value is significant, mark this node as possibly abnormal (such as equipment aging or pipeline blockage).

[0054] If abnormal situations are detected, the information of the abnormal points (including node numbers, predicted values, and their differences from physical laws or historical data) will be fed back to the physical constraint calculation module and the prediction output module. This feedback mechanism can achieve the adaptive optimization of the system and gradually improve the accuracy of the prediction results.

[0055] After the verification is completed, the finally verified prediction results are output to the user interface, including the node pressure distribution map, the pipeline flow distribution map, and the leakage point marking map. For the nodes marked as abnormal, the recommended correction values and possible reasons for the anomalies are also output.

[0056] The embodiment of the present application also discloses an intelligent prediction system for urban pipe network operation based on a physical information graph neural network, including: a data input module, a topological feature extraction module, a physical constraint calculation module, a prediction output module, and a result verification module. The data input module, the topological feature extraction module, the physical constraint calculation module, the prediction output module, and the result verification module are communicatively connected for data transmission.

[0057] The data input module can obtain the monitoring data and topological structure information of the pipe network and generate time series graph data; generating time series graph data is to provide standardized data input for subsequent prediction model calculations.

[0058] The topological feature extraction module can extract features from the time series graph data using a graph neural network.

[0059] The physical constraint calculation module can use a physical information neural network as the core to implement the constraints on node pressure and pipeline flow by introducing hydraulic equations, and the hydraulic equations include the continuity equation and the energy equation.

[0060] The prediction output module can generate predicted values of the operating state and inversion results of key parameters based on the high-dimensional feature vector output by the physical constraint calculation module; the prediction output module is one of the key parts of the present invention, and its task is to generate accurate predicted values of the operating state and inversion results of key parameters based on the high-dimensional feature vector output by the physical constraint calculation module. Its main functions include node pressure distribution prediction, pipeline flow prediction, and inversion of key parameters (such as pipeline roughness coefficient and leakage point location).

[0061] The result verification module can perform a rationality check on the results of the prediction output module through a multi-level verification mechanism and mark and feedback abnormal situations. The result verification module is an important link to ensure the reliability and accuracy of the prediction results, and its task is to perform a rationality check on the results of the prediction output module through a multi-level verification mechanism and mark and feedback abnormal situations.

[0062] An embodiment of the present application also discloses a terminal, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent prediction method for the operation of urban pipe networks based on the physical information graph neural network according to the present application is implemented.

[0063] The above terminal further includes a communication bus, a communication interface and other components well known to those skilled in the art. Their settings and functions are known in the art, so they will not be described in detail here.

[0064] In the present application, the aforementioned memory can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device or apparatus. For example, a computer-readable storage medium can be any suitable magnetic storage medium or magneto-optical storage medium. For instance, a resistive random access memory (RRAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), an enhanced dynamic random access memory (EDRAM), a high bandwidth memory (HBM), a hybrid memory cube (HMC), etc., or any other medium that can be used to store the required information and can be accessed by an application program, a module, or both. Any such computer storage medium can be part of the device or accessible or connectable to the device.

[0065] The above are all preferred embodiments of the present application. The protection scope of the present application is not limited hereby. Therefore, all equivalent changes made according to the structure, shape and principle of the present application shall be covered within the protection scope of the present application.

Claims

1. An intelligent prediction method for urban pipe network operation based on physical information graph neural network, characterized in that: Includes steps: Obtain monitoring data and topological structure information of the pipeline network and generate time series graph data; Construct a prediction model, which is a hybrid model combining graph neural network and physical information neural network; extract features from time series graph data based on graph neural network to model the dynamic characteristics of the pipeline network; embed physical constraints into the loss function based on physical information neural network; Generate the operation status prediction results of the pipeline network according to the prediction model; verify the prediction results and optimize the model according to the verification; Among them, the loss function is set as the cumulative sum of the data fitting loss, the product of the continuity equation loss for constraining the flow conservation of the pipeline node and the first weight coefficient, and the energy equation loss for correcting the pressure loss of the pipeline and the second weight coefficient; the continuity equation loss is determined according to the flow at the node of the pipeline network and the external water supply or water consumption; the energy equation loss is calculated according to the pipeline pressure and the physical parameters of the pipeline.

2. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The continuity equation loss is the sum of the absolute difference between the sum of the flow from each node to other nodes and the water consumption of the node.

3. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The energy equation loss is the sum of the absolute differences of the inlet pressure, outlet pressure and friction loss along the pipeline.

4. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The first weight coefficient and the second weight coefficient are custom parameters. During the training process, the first weight coefficient and the second weight coefficient are dynamically adjusted. The expression for dynamic adjustment is: ; ; Represents the first parameter, Represents the second parameter, represents the data fitting loss, represents the loss of continuity equations used to constrain the flow conservation of the network nodes, Express the energy equation loss used to correct for pressure losses in pipes; The first parameter is used as the adjusted first weight coefficient, and the second parameter is used as the adjusted second weight coefficient.

5. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The prediction results include node pressure prediction values, pipeline flow prediction values ​​and / or inversion values ​​of key parameters.

6. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 5 is characterized in that: The method for inversion of key parameters is to compare the differences between the predicted values ​​generated by the prediction model and the monitoring data, and iteratively optimize the estimated values ​​of key parameters.

7. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The method for generating time series graph data is as follows: construct a node feature matrix, where the feature of each node is a high-dimensional vector composed of pressure, flow and related historical values; construct an edge feature matrix, where the feature of the pipeline is a high-dimensional vector composed of diameter, length and roughness coefficient; construct an adjacency matrix to represent the topological structure of the pipeline network; integrate the node feature matrix, edge feature matrix and adjacency matrix to generate time series graph data and store it in tensor format.

8. The method for intelligent prediction of urban pipe network operation based on physical information graph neural network according to claim 1 is characterized in that: The topological structure information includes the connection relationship between nodes and pipes, the length of the pipes, the diameter of the pipes, and the roughness coefficient of the pipes.

9. The intelligent prediction system for urban pipe network operation based on physical information graph neural network is characterized by: include: Data input module, which obtains monitoring data and topological structure information of the pipeline network and generates time series graph data; The topological feature extraction module uses graph neural network to extract features from time series graph data; The physical constraint calculation module uses the physical information neural network as the core to implement the constraints on node pressure and pipeline flow by introducing hydraulic equations. The hydraulic equations include continuity equations and energy equations. The prediction output module generates the operation status prediction value and key parameter inversion results based on the high-dimensional feature vector output by the physical constraint calculation module; The result verification module verifies the rationality of the prediction output module’s results through a multi-level verification mechanism, and marks and provides feedback on abnormal situations; The data input module, the topological feature extraction module, the physical constraint calculation module, the prediction output module and the result verification module are connected in communication to perform data transmission.

10. A terminal, characterized in that include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the urban pipe network operation intelligent prediction method based on the physical information graph neural network according to any one of claims 1 to 8 is implemented.

Images