Underground pipe network subarea metering and intelligent regulation and control system

Abstract

The invention discloses an underground pipe network partition metering and intelligent regulation and control system which comprises a data collecting and preprocessing module used for collecting and preprocessing multi-source time sequence data; the graph structure construction module is used for constructing a graph structure data set and forming a node set, an edge relation and a feature matrix; the graph neural network modeling module is used for constructing a graph neural network model containing a residual connection structure; the parameter optimization module is used for optimizing structure parameters of the graph neural network; the state prediction and instruction generation module is used for predicting the state of each node and generating a regulation and control instruction; the instruction execution module is used for issuing the regulation and control instruction and carrying out dynamic regulation and control; and the feedback updating module is used for collecting state feedback data, evaluating prediction deviation and updating the graph structure data set and the graph neural network model parameters. According to the method, the graph neural network and the dragonfly optimization algorithm are utilized, accurate prediction of the node state of the underground pipe network and intelligent partition regulation and control are achieved, and the operation efficiency and safety of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of underground pipeline network monitoring and intelligent control, and particularly to an underground pipeline network zoning metering and intelligent regulation system. Background Art

[0002] With the acceleration of the urbanization process and the continuous expansion of the scale of urban infrastructure, as an important part of ensuring the normal operation of the city, the safety, stability and intelligent management ability of the operation state of the underground pipeline network have received more and more attention. The underground pipeline network system usually covers various functional categories such as water supply, drainage, gas, heat, communication, etc. Its structure is complex, spatially widespread, and the operating environment changes violently dynamically, which brings huge challenges to daily monitoring, maintenance and regulation. Traditional underground pipeline network management mainly relies on static modeling and regular inspection methods, which have problems such as insufficient coverage of monitoring data, untimely state prediction, single and slow-response regulation instructions, and it is difficult to meet the needs of modern cities for efficient, safe and intelligent operation management.

[0003] In the prior art, the methods for monitoring and regulating the state of the underground pipeline network mainly adopt the method of collecting sensor data and formulating empirical rules. Some systems deploy devices such as flow meters, pressure gauges, and liquid level sensors at key nodes to collect operation parameters regularly and trigger simple control operations according to preset alarm thresholds. Although such methods achieve partial monitoring and early warning to a certain extent, there are many limitations: First, the distribution of data collection points is sparse, lacking a continuous description of the overall network state, which easily leads to the spread of local anomalies and slow overall response; Second, the regulation logic relying on static rule setting cannot effectively cope with dynamic environmental changes, especially in the case of sudden load changes or partial system failures, the regulation strategy is rigid, the response is lagging, and it is difficult to achieve refined and intelligent management; Third, the traditional method based on single-point monitoring or decentralized control lacks the understanding and utilization of the overall topological structure of the underground pipeline network, and cannot effectively mine the correlation and influence path between nodes, which limits the global optimality of regulation decisions.

[0004] To address the above problems, in recent years, some studies have begun to explore the introduction of artificial intelligence technologies, especially using machine learning methods to model and predict the state of underground pipe networks. For example, deep learning models such as long short-term memory networks (LSTM) and convolutional neural networks (CNN) are used to extract features and detect anomalies from sensor time series data. However, most of these methods still rely on independent node-level modeling and fail to fully integrate the graph structure characteristics of the pipe network itself, making it difficult to capture the spatial correlations and dynamic transfer rules between nodes. At the same time, traditional gradient descent-based methods are mostly used in the parameter optimization process, which is prone to falling into local optima and difficult to achieve efficient global search in complex non-convex spaces, resulting in limited generalization ability and insufficient robustness of the finally generated prediction and control models in practical applications.

[0005] For graph structure data modeling, some existing methods have introduced graph neural network models such as graph convolutional networks (GCN) and graph attention networks (GAT) into the field of infrastructure management, which can process the information of node and connection relationships to a certain extent. However, there are still several key problems when existing graph neural networks are applied to actual underground pipe network management: First, standard graph neural networks are prone to the over-smoothing phenomenon during deep training, that is, node representations tend to be the same, resulting in a decline in feature expression ability and difficulty in accurately distinguishing different node states; Second, the configuration of model hyperparameters usually adopts empirical settings or simple searches, lacking a systematic and intelligent optimization mechanism and being difficult to adaptively adjust in pipe network systems of different scales and different topological complexities; Third, existing models generally ignore the physical constraint characteristics in the operation of underground pipe networks, such as flow conservation and pressure balance, resulting in a deviation between the model prediction results and actual controllability and reducing the executability of regulation decisions.

[0006] In addition, in terms of the selection of optimization methods, although some studies have tried to introduce swarm intelligence methods such as particle swarm optimization (PSO) and genetic algorithms (GA) for hyperparameter search, these traditional algorithms have problems such as slow convergence speed, being prone to falling into local optima, and lacking the ability to adjust dynamic behaviors, making it difficult to fully explore the optimal configuration in high-dimensional complex search spaces. At the same time, the existing optimization processes generally do not make targeted improvements in combination with the unique structural characteristics of graph neural network models, resulting in a disconnection between the search effect and the network training effect and being difficult to achieve systematic overall performance improvement.

[0007] Therefore, how to provide an underground pipe network sectional metering and intelligent regulation system is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to provide an underground pipe network zoning metering and intelligent regulation system. The present invention makes full use of the deep modeling ability of the graph neural network for the topological structure characteristics of the pipe network, combines the dynamic global search ability of the dragonfly optimization algorithm, and through the introduction of a local behavior priority mechanism, a structure-aware neighborhood definition strategy, and a multi-dimensional fitness fusion technology, details a method for realizing node state prediction and zoning regulation instruction generation based on an optimized residual graph neural network model, which has the advantages of high prediction accuracy, fast regulation response, strong intelligent self-adaptability, and high operation reliability.

[0009] The underground pipe network zoning metering and intelligent regulation system according to an embodiment of the present invention includes: A data collection and preprocessing module, configured to collect multi-source time-series data of each zone in the underground pipe network system and perform preprocessing; A graph structure construction module, configured to construct a graph structure data set according to node attributes and pipe segment connection relationships, and form a node set, edge relationships, and a feature matrix; A graph neural network modeling module, configured to construct a graph neural network model including a residual connection structure and complete multi-layer graph convolution calculation; A parameter optimization module, configured to introduce the dragonfly optimization algorithm to perform search iteration on the structure parameters of the graph neural network; A state prediction and instruction generation module, configured to predict the state of each node based on the optimized graph neural network model and generate regulation instructions for the corresponding zones in combination with regulation rules; An instruction execution module, configured to send the regulation instructions to the zone execution devices in the underground pipe network system for dynamic adjustment control; A feedback update module, configured to collect the state feedback data after control execution, evaluate the prediction deviation, and update the graph structure data set and the graph neural network model parameters to complete closed-loop regulation.

[0010] Optionally, the modules are implemented by the following method: S1. Collect multi-source time-series data of each zone in the underground pipe network system, perform preprocessing on the multi-source time-series data, and generate a data structure including node attributes and pipe segment connection relationships; S2. Construct a graph structure data set of the underground pipe network based on the data structure, where the graph structure data set includes a node set, edge connection relationships, a node feature matrix, and an adjacency matrix; S3. Use the graph structure data set as an input, construct a graph neural network model including a residual connection structure, perform graph convolution operations on the nodes in the graph structure, set residual path connections between each graph convolution layer, and perform non-linear transformation on the intermediate layer output using a normalization function and an activation function; S4. Introduce the Dragonfly Optimization Algorithm to optimize the structural parameters of the graph neural network model. Evaluate the prediction error and training stability by constructing a fitness function, and execute the behavioral mechanism of the Dragonfly Optimization Algorithm to search and iterate the combination of structural parameters. S5. Apply the optimized graph neural network model to the graph-structured dataset, generate the predicted state data of each node, and generate zoning control instructions in combination with the control rules. S6. Send the control instructions to the corresponding execution devices, and dynamically adjust the relevant nodes in the underground pipe network system through the execution devices. S7. Collect the feedback data after the dynamic adjustment execution, including the difference data between the actual node state value and the predicted state value, update the feedback data to the graph-structured dataset, and retrain the graph neural network model to form a closed-loop optimization control mechanism.

[0011] Optionally, the multi-source time-series data specifically includes the flow rate, pressure, water quality, and conductivity of each partition node, which are used to comprehensively characterize the operation state of the underground pipe network and support map construction and state prediction analysis.

[0012] Optionally, the preprocessing of the multi-source time-series data specifically includes missing value filling, outlier removal, normalization, and time alignment processing, which are used to improve the quality of the multi-source time-series data and construct input data in a unified format to support graph structure construction and graph neural network model training.

[0013] Optionally, the node set represents measurement points, pump stations, or valves, the edge connection relationship represents the physical pipe segment connection between nodes, the node feature matrix contains the historical physical parameters of each node, and the adjacency matrix represents the topological connection relationship between each node.

[0014] Optionally, the S2 specifically includes: S21. Map the preprocessed multi-source time-series data to the node level and construct a node set , where represents the i-th measurement point, valve, or pump station in the underground pipe network, , with a total of n nodes; S22. Construct an edge set according to the actual connection relationship of the physical pipe segments in the underground pipe network, where represents that there is a physical pipe segment connection between node and node , and the edge set is used to define the topological structure of the graph; S23. Generate a node feature matrix based on the physical monitoring data of each node, where n represents the number of nodes and d represents the feature dimension of each node, and the feature vector , including node Combined representation of flow rate, pressure, water quality, and conductivity at multiple time points; S24. Construct an adjacency matrix A based on the connection relationships between nodes, where the element in the i-th row and j-th column represents that node and node have a physical connection, indicating no connection between the two; S25. Combine the node set V, edge set E, node feature matrix X, and adjacency matrix A to form a graph structure dataset G=(V, E, X, A) for input into a graph neural network model for state prediction and structure optimization.

[0015] Optionally, the specific steps of S3 are as follows: S31. Use the graph structure dataset G=(V, E, X, A) as the input, where V is the node set, E is the edge set, X is the node feature matrix, and A is the adjacency matrix; S32. Construct a graph neural network model with a residual connection structure. The graph neural network model consists of L layers of graph convolution structures, and residual paths are set between each graph convolution layer. Each graph convolution operation performs adjacency normalization weighted aggregation on the input features, and uses a normalization function and an activation function to perform a non-linear transformation on the output of the intermediate layer to form a node state representation matrix; S33. Set cross-layer residual connections in the graph neural network model. The cross-layer connections adopt cross-layer fusion depth series , indicating the number of historical layers that the current convolution layer backtracks and fuses forward, to generate a cross-layer residual output : ; Among them, is the fusion ratio coefficient, is the cross-layer residual span, is the node feature matrix of the -th layer for cross-layer residual connections, and L is the total number of layers of the graph neural network; S34. Introduce a partitioned residual fusion mechanism in the graph neural network model. Divide the node set into multiple partitions according to the actual physical structure of the underground pipe network, allocate an independent residual fusion coefficient for each partition, and perform regional fusion on the cross-layer residual output and the historical features of the previous layer within each partition to generate a partitioned residual output : ; Among them, is the residual fusion coefficient of the p-th partition, , m represents the number of physical partitions, Denote the feature matrix of the nodes in partition p after cross-layer residual fusion, which represents the feature matrix of the nodes in partition p after cross-layer residual fusion at the th layer; S35. Construct the conservation relationship and flow error in the operation of the underground pipe network into a physical deviation tensor , and generate a physical correction term through the physical deviation mapping function , and fuse the correction term for the partition residual output to generate the final node state representation matrix : ; where is the physical coupling weight coefficient, represents the output after cross-layer residual fusion of the th layer partition residual. The final node state representation matrix is used to characterize the operation state characteristics of each node in the underground pipe network at future moments, and serves as the basis for generating control instructions and making dynamic adjustment decisions.

[0016] Optionally, the specific steps of S4 include: S41. Construct the search space of the dragonfly optimization algorithm, and define the set of structure parameters to be optimized as: ; where is the cross-layer residual span, is the set of residual fusion coefficients for each physical partition, , , m represents the number of physical partitions divided in the underground pipe network, is the physical coupling weight coefficient; S42. Based on the graph structure dataset G=(V,E,X,A), construct a structure-aware neighborhood set. Among them, V is the set of nodes, E is the set of edges, X is the node feature matrix, and A is the adjacency matrix. Set the topological hop count between nodes as: ; where represents the shortest path hop count between node and , which is used to measure the proximity under structure awareness. represents the set of hop counts in all paths between node and . represents the minimum number of hops required for node to in all paths. Define the node pair that satisfies as the structure neighbor relationship, where is the neighborhood radius threshold for determining the individual neighborhood; S43. According to the constructed structure-aware neighborhood set, calculate the behavior vectors of each dragonfly individual, including: separation behavior vector S, alignment behavior vector AI, aggregation behavior vector C, food attraction behavior vector F, and enemy evasion behavior vector E. The neighborhood ranges of each behavior vector are defined based on the topological hop count between nodes; S44. Construct a multi-dimensional composite fitness function for comprehensively evaluating the performance of the graph neural network structure parameter combination: ; where, MAE represents the node state prediction error, represents the mean deviation index of the physical deviation tensor Π, represents the fluctuation variance of the training loss, , , is the weighting coefficient, Y is the true value, is the predicted value; S45. Introduce a local behavior priority mechanism to adaptively update the weight of the j-th behavior of each dragonfly individual in each iteration. The update rule is: ; where, represents the fitness value of the i-th individual in the current iteration, is the weight adjustment sensitivity coefficient, represents the weight coefficient of the j-th behavior in the (t + 1)-th iteration, represents the weight coefficient of the j-th behavior in the t-th iteration, The fitness function value of the i-th dragonfly individual in the -th iteration; S46. Use the behavior vector and the updated behavior weight to perform a composite behavior update to update the position of the dragonfly individual: ; where, is the updated dragonfly individual position vector in the (t + 1)-th iteration, is the current dragonfly individual position vector in the t-th iteration, is the dynamic weight of the separation behavior, is the dynamic weight of the alignment behavior, is the dynamic weight of the aggregation behavior, is the dynamic weight of the food attraction behavior, is the dynamic weight of the enemy evasion behavior, w is the inertia weight, represents the historical trend inertia term of the position update; S47. After each round of iteration is completed, based on the evaluation results of the fitness function , the following parameters are updated respectively: For each dragonfly individual i, if the fitness function value in the current iteration is better than the historical record value, then the structural parameter combination is recorded as the current local optimal parameter; Among all dragonfly individuals, select the dragonfly individual with the smallest fitness function value, and record the structural parameter combination as the global optimal parameter; The updated current local optimal parameter and global optimal parameter are respectively used for the direction guidance of the aggregation behavior vector C and the food attraction behavior vector F in the next round of iteration; S48. If the fitness change rate of the dragonfly individuals is lower than the preset threshold during consecutive rounds of iteration, trigger the behavior weight reset mechanism to reset , enhance the behavior balance and improve the global search ability; S49. When the maximum number of iterations is reached or the fitness convergence error is less than the set threshold , terminate the iteration process and output the optimal structural parameter combination ; S410. Apply the optimal structural parameter combination to the graph neural network model structure, configure the cross-layer residual fusion, partition residual fusion and physical coupling path, and generate an optimized graph neural network model for predicting the final node state representation matrix.

[0017] Optionally, the S5 specifically includes: S51. Load the graph neural network model structure optimized by the dragonfly optimization algorithm and initialize the parameters of the optimized graph neural network model; S52. Input the node feature information and connection relationship in the graph structure dataset into the optimized graph neural network model, perform forward propagation, and obtain the predicted state vector of each node; S53. Classify the predicted state vectors of all nodes according to physical partitions to generate a partition-level node state set; S54. According to the operation objectives of different physical partitions, load the corresponding regulation rule templates and perform rule matching on the partition node state sets; S55. Based on the rule matching results, generate the regulation instructions for each physical partition, including control actions, adjustment parameters and target state ranges; S56. Uniformly package the generated partition regulation instructions into a regulation instruction set and send them to the corresponding partition execution devices for adjustment and control.

[0018] The beneficial effects of the present invention are: By introducing a graph neural network model based on the residual connection structure into the underground pipe network system, the present invention can fully exploit the deep - level feature associations among the nodes and their connection relationships in the underground pipe network, overcoming the problems of excessive node feature smoothing and insufficient prediction discrimination in traditional methods. By designing cross - layer residual fusion, partition - based residual fusion, and physically - rule - coupled residual paths, the richness and expressive ability of node state representation are enhanced, enabling the prediction model to maintain high prediction accuracy and stability in underground pipe network environments of different scales and complexities.

[0019] The present invention uses the dragonfly optimization algorithm to globally optimize the structure parameters of the graph neural network, and innovatively introduces a local - behavior - first mechanism, structure - aware neighborhood definition, and multi - dimensional fitness fusion strategy, effectively improving the search efficiency and result quality of the optimization process. By dynamically adjusting the behavior weight coefficient, the search direction can be optimized in real - time according to the fitness change, avoiding falling into local optima; the structure - aware neighborhood defined based on topological hop count makes the population search behavior more conform to the actual physical structure of the underground pipe network; fusing the three indicators of prediction error, physical consistency, and training stability as the fitness evaluation criteria ensures that the optimized model not only has advantages in accuracy but also performs well in physical feasibility and engineering application reliability.

[0020] Through node state prediction based on the optimized graph neural network model and combined with physical partition control rules, the present invention generates more accurate and executable control instructions, enabling intelligent dynamic adjustment of the nodes in each partition of the underground pipe network. The control process combines real - time prediction and partition characteristics, realizing distributed and refined control, and improving the autonomous adaptation ability and operation safety of the underground pipe network system under complex working conditions. By continuously collecting execution feedback data and updating the model, a closed - loop optimization control mechanism is formed, enabling the system to continuously self - adjust and improve performance during long - term operation, and having good engineering application value and promotion prospects.

[0021] In summary, the present invention is significantly superior to the prior art in terms of node state prediction accuracy, control instruction rationality, optimization search efficiency, system intelligent adaptive ability, and long - term operation reliability. It can effectively solve the problems existing in traditional underground pipe network monitoring and control systems, such as insufficient structure modeling, insufficient parameter optimization, slow control response, and poor system reliability, and has significant technological progress and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the flowchart of the underground pipe network partition metering and intelligent control system proposed by the present invention; Figure 2 This is the flowchart of optimizing the structural parameters of the graph neural network based on the dragonfly optimization algorithm for the underground pipe network zoning metering and intelligent regulation system proposed by the present invention. Detailed implementation manners

[0023] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only showing the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0024] Refer to Figure 1 and Figure 2 , the underground pipe network zoning metering and intelligent regulation system includes: A data acquisition and preprocessing module, which is used to acquire multi-source time-series data of each zone in the underground pipe network system and perform preprocessing; A graph structure construction module, which is used to construct a graph structure data set according to node attributes and pipe section connection relationships, and form a node set, edge relationship and feature matrix; A graph neural network modeling module, which is used to construct a graph neural network model including a residual connection structure and complete multi-layer graph convolution calculation; A parameter optimization module, which is used to introduce the dragonfly optimization algorithm to search and iterate the structural parameters of the graph neural network; A state prediction and instruction generation module, which is used to predict the state of each node based on the optimized graph neural network model and generate regulation instructions for the corresponding zone in combination with regulation rules; An instruction execution module, which is used to send the regulation instructions to the zone execution equipment in the underground pipe network system for dynamic adjustment and control; A feedback update module, which is used to collect the state feedback data after control execution, evaluate the prediction deviation and update the graph structure data set and the parameters of the graph neural network model to complete the closed-loop regulation.

[0025] In this embodiment, the modules are implemented through the following methods: S1. Collect multi-source time-series data of each zone in the underground pipe network system, preprocess the multi-source time-series data, and generate a data structure including node attributes and pipe section connection relationships; S2. Based on the data structure, construct a graph structure data set of the underground pipe network, and the graph structure data set includes a node set, edge connection relationship, node feature matrix and adjacency matrix; S3. Take the graph structure data set as the input, construct a graph neural network model including a residual connection structure, perform graph convolution operations on the nodes in the graph structure, set residual path connections between each graph convolution layer, and use a normalization function and an activation function to perform non-linear transformation on the output of the intermediate layer; S4. Introduce the dragonfly optimization algorithm to optimize the structural parameters of the graph neural network model. Evaluate the prediction error and training stability by constructing a fitness function, and execute the behavioral mechanism of the dragonfly optimization algorithm to search and iterate the combination of structural parameters. S5. Apply the optimized graph neural network model to the graph-structured dataset, generate the predicted state data of each node, and generate zoning control instructions in combination with the control rules. S6. Send the control instructions to the corresponding execution devices, and dynamically adjust the relevant nodes in the underground pipe network system through the execution devices. S7. Collect the feedback data after the dynamic adjustment execution, including the difference data between the actual state value and the predicted state value of the node, update the feedback data to the graph-structured dataset, and retrain the graph neural network model to form a closed-loop optimization control mechanism.

[0026] In this embodiment, the multi-source time-series data specifically includes the flow rate, pressure, water quality, and conductivity of each partition node, which are used to comprehensively characterize the operation state of the underground pipe network and support map construction and state prediction analysis.

[0027] In this embodiment, the preprocessing of the multi-source time-series data specifically includes missing value filling, outlier removal, normalization, and time alignment processing, which are used to improve the quality of the multi-source time-series data and construct input data in a unified format to support graph structure construction and graph neural network model training.

[0028] In this embodiment, the node set represents a measurement point, a pumping station, or a valve, the edge connection relationship represents the physical pipe segment connection between nodes, the node feature matrix contains the historical physical parameters of each node, and the adjacency matrix represents the topological connection relationship between each node.

[0029] In this embodiment, S2 specifically includes: S21. Map the preprocessed multi-source time-series data to the node level and construct a node set , where represents the i-th measurement point, valve, or pumping station in the underground pipe network, , with a total of n nodes; S22. Construct an edge set according to the actual connection relationship of the physical pipe segments in the underground pipe network, where represents that there is a physical pipe segment connection between node and node , and the edge set is used to define the topological structure of the graph; S23. Generate a node feature matrix based on the physical monitoring data of each node, where n represents the number of nodes and d represents the feature dimension of each node. The feature vector , including node Combined representation of flow rate, pressure, water quality, and conductivity at multiple time points; S24. Construct an adjacency matrix A according to the connection relationship between nodes, where the element in the i-th row and j-th column represents the node and the node have a physical connection, indicating no connection between the two; S25. Combine the node set V, edge set E, node feature matrix X, and adjacency matrix A to form a graph structure dataset G = (V, E, X, A) for input into a graph neural network model for state prediction and structure optimization.

[0030] In this embodiment, the specific steps of S3 are as follows: S31. Use the graph structure dataset G = (V, E, X, A) as the input, where V is the node set, E is the edge set, X is the node feature matrix, and A is the adjacency matrix; S32. Construct a graph neural network model with a residual connection structure. The graph neural network model consists of L layers of graph convolution structures, and residual paths are set between each graph convolution layer. Each graph convolution operation performs adjacency normalization weighted aggregation on the input features, and uses a normalization function and an activation function to perform a non-linear transformation on the output of the intermediate layer to form a node state representation matrix; S33. Set cross-layer residual connections in the graph neural network model. The cross-layer connections adopt cross-layer fusion depth system , indicating the number of historical layers that the current convolution layer backtracks and fuses forward, to generate a cross-layer residual output : ; Among them, is the fusion ratio coefficient, is the cross-layer residual span, is the node feature matrix of the -th layer for cross-layer residual connection, and L is the total number of layers of the graph neural network; S34. Introduce a partitioned residual fusion mechanism in the graph neural network model. According to the actual physical structure of the underground pipe network, divide the node set into multiple partitions , assign an independent residual fusion coefficient to each partition, and perform regional fusion on the cross-layer residual output and the historical features of the previous layer within each partition to generate a partitioned residual output : ; Among them, is the residual fusion coefficient of the p-th partition, , m represents the number of physical partitions, denotes the feature matrix of the nodes in partition p after cross-layer residual fusion, denotes the feature matrix of the nodes in partition p after cross-layer residual fusion at the -th layer; S35. Construct the conservation relationship and flow error in the operation of the underground pipe network into a physical deviation tensor , and generate a physical correction term through the physical deviation mapping function , and fuse the correction term for the partition residual output to generate the final node state representation matrix : ; wherein, is the physical coupling weight coefficient, denotes the output after cross-layer residual fusion of the -th layer partition residual. The final node state representation matrix is used to characterize the operation state characteristics of each node in the underground pipe network at a future moment, and serves as the basis for generating control instructions and making dynamic adjustment decisions.

[0031] In this embodiment, S4 specifically includes: S41. Construct the search space of the dragonfly optimization algorithm, and define the set of structural parameters to be optimized as: ; wherein, is the cross-layer residual span, is the set of residual fusion coefficients of each physical partition, , , m represents the number of physical partitions divided in the underground pipe network, is the physical coupling weight coefficient; S42. Based on the graph structure dataset G=(V, E, X, A), construct a structure-aware neighborhood set. Among them, V is the set of nodes, E is the set of edges, X is the node feature matrix, and A is the adjacency matrix. Set the topological hop count between nodes as: ; wherein, denotes the shortest path hop count between node and , which is used to measure the proximity under structure awareness, denotes the set of hop counts in all paths between node and , denotes the minimum number of hops required for node to in all paths. Define the node pair that satisfies as the structure neighbor relationship, wherein, is the neighborhood radius threshold, which is used to determine the individual neighborhood; S43. According to the constructed structure-aware neighborhood set, calculate the behavior vectors of each dragonfly individual, including: separation behavior vector S, alignment behavior vector AI, aggregation behavior vector C, food attraction behavior vector F, and enemy evasion behavior vector E. The neighborhood range of each behavior vector is defined based on the topological hop count between nodes; S44. Construct a multi-dimensional composite fitness function for comprehensively evaluating the performance of the combination of graph neural network structure parameters: ; where, MAE represents the node state prediction error, represents the mean deviation index of the physical deviation tensor Π, represents the fluctuation variance of the training loss, , , is the weighting coefficient, Y is the true value, is the predicted value; S45. Introduce a local behavior priority mechanism to adaptively update the weight of the j-th behavior of each dragonfly individual in each iteration. The update rule is: ; where, represents the fitness value of the i-th individual in the current iteration, is the weight adjustment sensitivity coefficient, represents the weight coefficient of the j-th behavior in the (t + 1)-th iteration, represents the weight coefficient of the j-th behavior in the t-th iteration, The fitness function value of the i-th dragonfly individual in the -th iteration; S46. Use the behavior vector and the updated behavior weight to perform composite behavior update to update the position of the dragonfly individual: ; where, is the updated position vector of the dragonfly individual in the (t + 1)-th iteration, is the position vector of the current dragonfly individual in the t-th iteration, is the dynamic weight of the separation behavior, is the dynamic weight of the alignment behavior, is the dynamic weight of the aggregation behavior, is the dynamic weight of the food attraction behavior, is the dynamic weight of the enemy evasion behavior, w is the inertia weight, represents the inertia term of the historical trend of position update; S47. After each round of iteration is completed, based on the evaluation results of the fitness function , the following parameters are updated respectively: For each dragonfly individual i, if the fitness function value in the current iteration is better than the historical record value, then the structural parameter combination is recorded as the current local optimal parameter; Among all dragonfly individuals, select the dragonfly individual with the smallest fitness function value, and record the structural parameter combination as the global optimal parameter; The updated current local optimal parameter and global optimal parameter are respectively used for the direction guidance of the aggregation behavior vector C and the attracting food behavior vector F in the next round of iteration; S48. If the fitness change rate of the dragonfly individuals is lower than the preset threshold during consecutive rounds of iteration, trigger the behavior weight reset mechanism to reset , enhance the behavior balance and improve the global search ability; S49. When the maximum number of iterations is reached or the fitness convergence error is less than the set threshold , terminate the iteration process and output the optimal structural parameter combination ; S410. Apply the optimal structural parameter combination to the graph neural network model structure, configure the cross-layer residual fusion, partition residual fusion and physical coupling path, and generate an optimized graph neural network model for predicting the final node state representation matrix.

[0032] In this embodiment, the S5 specifically includes: S51. Load the graph neural network model structure optimized by the dragonfly optimization algorithm and initialize the parameters of the optimized graph neural network model; S52. Input the node feature information and connection relationship in the graph structure dataset into the optimized graph neural network model, perform forward propagation, and obtain the predicted state vector of each node; S53. Classify the predicted state vectors of all nodes according to physical partitions to generate a partition-level node state set; S54. According to the operation objectives of different physical partitions, load the corresponding regulation rule templates and perform rule matching on the node state sets of each partition; S55. Based on the rule matching results, generate the regulation instructions for each physical partition, including control actions, adjustment parameters and target state ranges; S56. Uniformly package the generated partition regulation instructions into a regulation instruction set and send them to the corresponding partition execution device to achieve adjustment control.

[0033] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the underground drainage system in the main urban area of a coastal city. The underground pipe network in this area covers an area of about 57 square kilometers, with 183 nodes and 272 pipe segment connections, and the structure is intricate. Due to historical construction faults, old equipment, and data island problems, multiple large-scale waterlogging events occurred during the continuous heavy rains in the summer of 2023 in this area, which had a serious impact on urban traffic and residents' lives. After summarizing the lessons, the municipal management department decided to launch a new round of intelligent transformation pilot project for the underground drainage system in March 2024. The project area selected 12 main drainage sub-areas in the main urban area, and the goal was to achieve accurate prediction and dynamic control response of the pipe network node status by introducing artificial intelligence and intelligent optimization algorithms.

[0034] In the pilot project, the underground pipe network zoning metering and intelligent control system and its method proposed by the present invention were adopted. First, based on the original design drawings of the pipe network and the results of supplementary sensor layout, information such as node attributes and pipe segment connection relationships was established to form a complete graph structure data set. At the same time, multi-dimensional time-series data such as water level, flow rate, and pressure were collected in real time at each node, and abnormal data was cleaned and standardized to ensure the quality of the input data. Through the relationship between nodes and edges, a node feature matrix and an adjacency matrix were constructed, laying a foundation for subsequent graph neural network training.

[0035] Subsequently, the residual connection graph neural network model designed by the present invention was used for node state modeling. By introducing cross-layer residual fusion, zonal residual weight adjustment, and physical conservation constraints, the feature expression ability of the model under complex pipe network structures was enhanced. At the same time, the dragonfly optimization algorithm was used to globally optimize the key structural parameters in the network. Innovatively, the local behavior priority mechanism was combined to dynamically adjust the individual behavior strategy, the information propagation range was controlled according to the structure-aware neighborhood, and the multi-dimensional fitness fusion function was used as the search evaluation criterion, comprehensively considering the prediction error, physical consistency, and training stability, and finally the optimal model configuration was obtained.

[0036] During the continuous testing period from May to September 2024, the system predicted the node water level and flow velocity status within the next 2 hours every day. According to the prediction results and the actual operation rules of each drainage sub-area, zonal control instructions such as flood storage scheduling, drainage pump startup, and gate opening and closing were automatically generated and sent to each execution device. Through comparison with the actual observed data, it was found that after applying the method of the present invention, the prediction error of the node status decreased significantly, and the response speed of zonal control increased substantially. Especially in the three heavy rains in July and August, the system predicted the possible overflow risks of key nodes 30 minutes, 45 minutes, and 25 minutes in advance respectively, and initiated relevant drainage measures in advance, effectively avoiding serious waterlogging events in the main urban roads and residential areas.

[0037] To verify the effectiveness of the system, the project team compared the key operation indicators of the pilot area during the flood seasons of 2023 (without applying the technology of the present invention) and 2024 (applying the technology of the present invention). The statistical data shows that the mean square error of node status prediction decreased by about 26.8%, the accuracy rate of overflow warning increased by 18.2%, the average emergency response time was shortened by about 21 minutes, and the overall operation efficiency of the drainage system increased by 19.5%. At the same time, through the feedback mechanism, the model was continuously updated and optimized, and the system maintained good stability and adaptability during long-term operation, significantly reducing the frequency of manual intervention and operation and maintenance costs.

[0038] Table 1 Comparison of the intelligent regulation effects of the underground pipe network before and after application in the pilot area ; In this embodiment, by comparing the actual operation data of the flood season in 2023 without applying the method of the present invention and that in 2024 after applying the method of the present invention, it can be clearly seen that the present invention has achieved remarkable results in improving the intelligent regulation effect of the underground pipe network.

[0039] First of all, in terms of node status prediction, before applying the present invention, the mean square error (MSE) of node status prediction was 0.0245, while in the flood season of 2024 after applying the present invention, this error decreased to 0.0179, and the overall reduction rate reached 26.8%. This result shows that the present invention, through the joint optimization of the residual structure graph neural network modeling and the dragonfly optimization algorithm, has significantly improved the accuracy of node status prediction, effectively avoiding the performance degradation problem caused by over-smoothing or local overfitting of the prediction model, thus providing a more reliable data basis for subsequent regulation decisions.

[0040] Secondly, in terms of the accuracy rate of node water level warning, it only reached 77.5% in the flood season of 2023, while it increased to 91.6% after applying the present invention, with an increase rate as high as 18.2%. The improvement of the warning accuracy rate directly reflects the enhancement of the present invention method in abnormal detection and risk identification capabilities, enabling the system to identify potential overflow risk nodes earlier and more accurately, and triggering emergency regulation measures in a timely manner, thus effectively preventing the occurrence of major safety accidents.

[0041] In terms of the emergency response time, when the traditional system faced emergencies, the average response time was 88 minutes, while the system of the present invention shortened the response time to 67 minutes, with an overall shortening of 21 minutes, and the response speed was significantly accelerated. The shortening of the response time reflects that the present invention has realized the transformation from "post-event response" to "pre-event prevention and control" through the prediction-based regulation instruction generation mechanism, greatly improving the initiative of regulation and the flexibility of the system.

[0042] The overall operating efficiency of the drainage system has also been significantly improved. By comprehensively analyzing the flow rate at system processing nodes, water level changes, and emergency response effects, the overall operating efficiency of the drainage system after applying the present invention has increased by 19.5%. This improvement indicates that the present invention not only has advantages in predicting the state of single-point nodes but also can play a role in system-level coordinated control, enabling each partition of the underground pipe network to work together and the flow distribution to be more reasonable, effectively alleviating local bottleneck problems.

[0043] In addition, the system operation and maintenance costs have also decreased significantly. Before the present invention was applied, the average monthly number of emergency responses in the pilot area was 16 times, while after applying the present invention, it decreased to 9 times, with a decrease rate of 43.8%. At the same time, the average monthly system operation and maintenance costs decreased from 603,000 yuan to 495,000 yuan, a decrease of 17.9%. The reduction in the number of emergency responses indicates that the stability and self-adaptive ability of the system during normal operation have been greatly improved, and the reduction in operation and maintenance costs reflects the economic benefits brought by the intelligent prediction and dynamic regulation mechanism in practical applications.

[0044] In summary, the data in Table 1 fully demonstrate the superior performance of the present invention in the intelligent regulation application of actual underground pipe networks. By introducing the optimized graph neural network modeling and dragonfly optimization-driven mechanism, the present invention not only improves the accuracy of node state prediction and the timeliness of early warning but also speeds up the emergency response speed, improves the overall drainage efficiency of the system, reduces the operation and maintenance burden and economic costs, and effectively makes up for the deficiencies of traditional methods in aspects such as insufficient structural understanding, prediction lag, and single regulation. Overall, the present invention can greatly improve the intelligent level, resilience, and safety of the underground pipe network system, and has broad practical application prospects and important promotion value.

[0045] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent replacements or changes, and should be covered by the protection scope of the present invention.

Claims

1. Underground pipeline network partition metering and intelligent regulation system, characterized in that Including: A data acquisition and preprocessing module, which is used to acquire multi-source time-series data of each partition in the underground pipe network system and perform preprocessing; A graph structure construction module, which is used to construct a graph structure data set according to node attributes and pipe segment connection relationships, forming a node set, edge relationships, and a feature matrix; A graph neural network modeling module, which is used to construct a graph neural network model containing a residual connection structure and complete multi-layer graph convolution calculations; A parameter optimization module, which is used to introduce the dragonfly optimization algorithm to search and iterate the structural parameters of the graph neural network; A state prediction and instruction generation module, which is used to predict the state of each node based on the optimized graph neural network model and generate control instructions for the corresponding partition in combination with control rules; An instruction execution module, which is used to send the control instructions to the partition execution devices in the underground pipe network system for dynamic adjustment and control; A feedback update module, which is used to collect the state feedback data after control execution, evaluate the prediction deviation, and update the graph structure data set and the parameters of the graph neural network model to complete closed-loop control.

2. The underground pipe network zoning metering and intelligent regulation system according to claim 1, wherein The modules are implemented through the following methods: S1. Acquire multi-source time-series data of each partition in the underground pipe network system, preprocess the multi-source time-series data, and generate a data structure containing node attributes and pipe segment connection relationships; S2. Based on the data structure, construct a graph structure data set of the underground pipe network. The graph structure data set includes a node set, edge connection relationships, a node feature matrix, and an adjacency matrix; S3. Take the graph structure data set as input, construct a graph neural network model containing a residual connection structure, perform graph convolution operations on the nodes in the graph structure, set residual path connections between each graph convolution layer, and use a normalization function and an activation function to perform non-linear transformation on the output of the intermediate layer; S4. Introduce the dragonfly optimization algorithm to optimize the structural parameters of the graph neural network model, evaluate the prediction error and training stability by constructing a fitness function, and execute the behavioral mechanism of the dragonfly optimization algorithm to search and iterate the combination of structural parameters; S5. Apply the optimized graph neural network model to the graph structure data set, generate predicted state data of each node, and generate partition control instructions in combination with control rules; S6. Send the control instructions to the corresponding execution devices, and perform dynamic adjustment on the relevant nodes in the underground pipe network system through the execution devices; S7. Collect the feedback data after dynamic adjustment execution, including the difference data between the actual state value and the predicted state value of the node, update the feedback data to the graph structure data set, and retrain the graph neural network model to form a closed-loop optimization control mechanism.

3. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that, The multi-source time-series data specifically includes the flow rate, pressure, water quality, and conductivity of each partition node, which is used to comprehensively characterize the operation state of the underground pipe network and support map construction and state prediction analysis.

4. The underground pipe network zoning metering and intelligent regulation system according to claim 2, wherein The preprocessing of the multi-source time-series data specifically includes missing value filling, outlier removal, normalization, and time alignment processing, which is used to improve the quality of the multi-source time-series data and construct input data in a unified format to support graph structure construction and graph neural network model training.

5. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that The set of nodes represents measurement points, pumping stations or valves, the edge connection relationship represents the physical pipe segment connection between nodes, the node feature matrix contains the historical physical parameters of each node, and the adjacency matrix represents the topological connection relationship between each node.

6. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that The specific steps of S2 include: S21. Map the preprocessed multi-source time series data to the node level to construct a node set , where represents the i-th measurement point, valve or pumping station in the underground pipe network , with a total of n nodes S22. Construct an edge set according to the actual connection relationship of physical pipe segments in the underground pipe network , where represents a node and node are physically connected by a pipe segment, and the edge set is used to define the topological structure of the graph; S23. Generate a node feature matrix based on the physical monitoring data of each node , where n represents the number of nodes, d represents the feature dimension of each node, and the feature vector , including the combination representation of the flow rate, pressure, water quality, and conductivity of the node at multiple time points; S24. Construct an adjacency matrix A according to the connection relationship between nodes, where the element in the i-th row and j-th column represents the node and the node have a physical connection, indicating no connection between the two; S25: Combine the node set V, edge set E, node feature matrix X and adjacency matrix A to form a graph structure data set G=(V, E, X, A), which is used as the input of the graph neural network model for state prediction and structure optimization.

7. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that The specific steps of S3 include: S31: Take the graph structure data set G=(V, E, X, A) as the input, where V is the node set, E is the edge set, X is the node feature matrix, and A is the adjacency matrix; S32: Construct a graph neural network model with a residual connection structure. The graph neural network model consists of L layers of graph convolution structures. Residual paths are set between each graph convolution layer. Each graph convolution operation performs adjacency normalization weighted aggregation on the input features, and uses a normalization function and an activation function to perform a non-linear transformation on the output of the intermediate layer to form a node state representation matrix. S33. Set cross-layer residual connections in the graph neural network model. The cross-layer connection adopts cross-layer fusion depth series , indicating the number of historical layers that the current convolutional layer backtracks and fuses forward to generate a cross-layer residual output : ; Among them, is the fusion ratio coefficient, is the cross-layer residual span, is the node feature matrix of the th layer, which is used for cross-layer residual connection, and L is the total number of layers of the graph neural network; S34. Introduce a partition residual fusion mechanism into the graph neural network model, and divide the node set into multiple partitions according to the actual physical structure of the underground pipe network , and assign independent residual fusion coefficients to each partition , and perform regional fusion on the cross-layer residual output within each partition with the historical features of the previous layer to generate a partition residual output : ; Among them, is the residual fusion coefficient of the p-th partition, , m represents the number of physical partitions, represents the feature matrix of the nodes in partition p after cross-layer residual fusion, represents the nodes in partition p at the feature matrix after cross-layer residual fusion at layer S35. Construct the conservation relationship and flow error in the operation of the underground pipe network into a physical deviation tensor , and generate a physical correction term through the physical deviation mapping function , and fuse the correction term for the partition residual output to generate the final node state representation matrix : ; Among them, is the physical coupling weight coefficient, represents the output after the residual fusion of the -th layer partition. The final node state representation matrix is used to characterize the operation state characteristics of each node in the underground pipe network at future moments, and serves as the basis for generating control instructions and making dynamic adjustment decisions.

8. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that, The specific steps of S4 include: S41: Construct the search space of the dragonfly optimization algorithm, and define the set of structure parameters to be optimized as: ; Among them, is the cross-layer residual span, is the set of residual fusion coefficients for each physical partition, , , where m represents the number of physical partitions divided in the underground pipe network, is the physical coupling weight coefficient; S42: Based on the graph structure data set G=(V, E, X, A), construct a structure-aware neighborhood set, where V is the node set, E is the edge set, X is the node feature matrix, and A is the adjacency matrix. Set the topological hop count between nodes as: ; Among them, represents the number of hops of the shortest path between nodes and and is used to measure the proximity under structure perception. represents the set of hop numbers in all paths between nodes and ; represents the minimum number of hops required for node to in all paths. Nodes that satisfy are defined as having a structural neighbor relationship, where is the neighborhood radius threshold for determining an individual's neighborhood. S43: According to the constructed structure-aware neighborhood set, calculate the behavior vector of each dragonfly individual, including: separation behavior vector S, alignment behavior vector AI, aggregation behavior vector C, food attraction behavior vector F, and enemy escape behavior vector E. The neighborhood range of each behavior vector is defined based on the topological hop count between nodes; S44. Construct a multi-dimensional composite fitness function , which is used to comprehensively evaluate the performance of the combination of graph neural network structure parameters: ; Among them, MAE represents the node state prediction error, represents the mean deviation index of the physical deviation tensor Π, represents the fluctuation variance of the training loss, , , is the weighting coefficient, Y is the true value, is the predicted value; S45: Introduce a local behavior priority mechanism to adaptively update the weight of the j-th behavior of each dragonfly individual in each iteration. The update rule is: ; Among them, represents the fitness value of the i-th individual in the current iteration, is the weight adjustment sensitivity coefficient, represents the weight coefficient of the j-th behavior in the (t + 1)-th iteration, represents the weight coefficient of the j-th behavior in the t-th iteration, The fitness function value of the i-th dragonfly individual in the iteration; S46: Use the behavior vector and the updated behavior weight to perform composite behavior update to update the position of the dragonfly individual: ; Among them, is the updated position vector of the dragonfly individual in the (t + 1)-th iteration, is the position vector of the current dragonfly individual in the t-th iteration, is the dynamic weight of the separation behavior, is the dynamic weight of the alignment behavior, is the dynamic weight of the aggregation behavior, is the dynamic weight of the food attraction behavior, is the dynamic weight of the enemy evasion behavior, and w is the inertia weight, represents the inertia term of the historical trend of position update; S47. After each round of iteration, based on the evaluation results of the fitness function the following parameters are updated respectively: For each dragonfly individual i, if the fitness function value in the current iteration is better than the historical record value, then the structural parameter combination is recorded as the current local optimal parameter; Select the dragonfly individual with the smallest fitness function value among all dragonfly individuals, and record the combination of structure parameters as the global optimal parameters; The updated current local optimal parameters and global optimal parameters are respectively used to guide the directions of the aggregation behavior vector C and the food attraction behavior vector F in the next iteration; S48. If the change rate of the fitness of the dragonfly individuals is lower than the preset threshold during consecutive multiple rounds of iteration, the behavior weight reset mechanism is triggered to reset , enhance the behavior balance and improve the global search ability; S49. When the maximum number of iterations is reached or the fitness convergence error is less than the set threshold the iterative process is terminated and the optimal combination of structural parameters is output ; S410: Apply the optimal combination of structure parameters to the structure of the graph neural network model, configure the cross-layer residual fusion, partition residual fusion and physical coupling path, and generate an optimized graph neural network model for predicting the final node state representation matrix.

9. The underground pipe network zoning metering and intelligent regulation system according to claim 2, characterized in that The specific steps of S5 include: S51: Load the graph neural network model structure optimized by the dragonfly optimization algorithm and initialize the parameters of the optimized graph neural network model; S52: Input the node feature information and connection relationship in the graph structure data set into the optimized graph neural network model, perform forward propagation, and obtain the predicted state vector of each node; S53: Classify the predicted state vectors of all nodes according to physical partitions to generate a zone-level node state set. S54. Load the corresponding regulation rule template according to the operation objectives of different physical partitions, and perform rule matching on the set of node states in each partition; S55. Generate regulation instructions for each physical partition based on the rule matching results, including control actions, adjustment parameters, and target state ranges; S56. Uniformly package the generated partition regulation instructions into a regulation instruction set and send them to the corresponding partition execution device to achieve adjustment control.

Images