A water hammer control method and system for water supply network based on multifunctional module integration

Abstract

The present invention discloses a water hammer control method and system for a water supply network based on the fusion of multi-functional modules, which relates to the field of data recognition technology, including a pressure propagation map construction module: responsible for network topology modeling and causal graph generation, a time series graph neural network prediction and tracing module: based on a graph neural network model with causal constraints, water hammer risk prediction and propagation path tracing is realized; a reinforcement learning intervention decision module: generates active intervention strategies through reinforcement learning intelligent agents to form closed-loop control, and at the same time abstracts the water supply network into a graph structure and combines it with time series modeling, which can comprehensively reflect the propagation path of the pressure wave, overcome the blind spots of traditional local modeling, and improve the comprehensiveness and accuracy of water hammer event detection; uses the causal analysis results as the adjacency matrix input to effectively eliminate the interference of irrelevant edges in the topology on the prediction, and improves the learning efficiency and causal tracing ability of the model.

Description

Technical Field

[0001] The present invention relates to the field of data recognition technology, and in particular to a water hammer control method and system for a water supply network based on the fusion of multifunctional modules. Background Art

[0002] Water hammer is a pressure fluctuation phenomenon caused by sudden changes in fluid inertia. It is widely present in water supply pipeline systems. Its suddenness and instantaneous high-pressure characteristics pose a serious threat to the safety of the pipeline system. If not handled in time, it may lead to pipeline rupture, damage to pump station equipment and even large-scale water outages.

[0003] The water hammer warning methods commonly used in the prior art mainly include the following categories:

[0004] Early warning method based on pressure threshold: The system sets a fixed or dynamic threshold. For example, an alarm is triggered when the pressure exceeds 1.5 times the working pressure. However, the false alarm rate can reach more than 30% in complex pipe networks. For example, false alarms are frequently generated due to turbulent interference when the water pump is started and stopped.

[0005] Detection method based on local time series modeling: Long short-term memory network (LSTM), convolutional neural network (CNN) and other models are used to model single-point pressure time series data. Although it can identify abnormal fluctuations, the average prediction lags by 5-10 seconds. Moreover, because the pipeline network topology structure is not taken into account, the missed reporting rate in large-diameter pipelines above DN800 exceeds 20%.

[0006] Water hammer analysis method based on simulation and rule inference: Pipeline network response is simulated through software such as EPANET, and risks are judged in combination with engineering rules. However, in real-time scenarios, it takes 2-5 minutes from data collection to generating intervention instructions, which cannot meet the "golden 10 seconds" emergency response requirements for water hammer incidents.

[0007] In summary, existing methods generally have the following problems:

[0008] (1) The lack of modeling of the propagation path and spatial causal chain of pressure waves in the pipe network makes it impossible to effectively trace the source;

[0009] (2) Unable to provide targeted intervention strategies, most can only provide passive warnings and are unable to take proactive measures to prevent the disease;

[0010] (3) The influence of the pipe network topology on the water hammer propagation mode is ignored, and the model has poor interpretability and generalization ability.

[0011] Therefore, a water hammer control method and system for water supply network based on multi-functional module fusion is proposed to solve the above problems. Summary of the Invention

[0012] In view of this, the technical problem to be solved by the present invention is to propose a water hammer control method and system for a water supply network based on the integration of multi-functional modules to solve the problems mentioned in the background technology.

[0013] To achieve the above objectives, the present invention provides the following technical solutions: A water hammer control method for a water supply network based on the integration of multifunctional modules comprises:

[0014] S1, digital modeling of the pipeline network: Based on the geographic information system (GIS) and design drawings, the pipeline network topology map is constructed, sensors are deployed, and pipeline attributes are recorded;

[0015] S2, real-time data acquisition and preprocessing: The sensor collects pipeline pressure data at 0.5-2 Hz, fills missing values ​​through cubic spline interpolation, and uses the triple standard deviation principle (3σ) combined with EPANET software simulation to eliminate outliers;

[0016] S3, causal graph generation and model input: Perform causal testing (Granger) on the preprocessed data (lag order k = 2-5, time window set to T = 30-90 seconds) to generate a sparse causal adjacency matrix as the input of the graph neural network (TGNN);

[0017] S4, Water Hammer Risk Prediction and Tracing: The Graph Neural Network (TGNN) model uses a "graph convolutional neural network (GCN) + gated recurrent unit (GRU) + attention mechanism" architecture to output the probability of water hammer occurrence and the causal propagation chain, locating the trigger point and propagation path;

[0018] S5, Reinforcement Learning and Intelligent Intervention: A pipeline network hydraulic simulation environment is built through the EPANET platform. Within the EPANET platform's digital twin environment, a Deep Q Network agent generates intervention actions based on pressure distribution and predicted positions, with minimizing pressure fluctuations as the reward objective. Control strategies are executed in real time through the Supervisory Control and Data Acquisition (SCADA) system, forming a closed-loop control system.

[0019] Preferably, the sensor described in S1 is a pressure sensor, and is deployed at (bifurcation point) and upstream and downstream of the valve. The pressure sensor deployment spacing is ≤500 meters, and the recorded pipeline properties include material, length and diameter.

[0020] Preferably, the missing values ​​are filled by cubic spline interpolation as described in S2, and manual verification is performed when the correlation coefficient of adjacent nodes is less than 0.7, and the abnormal value deviation is greater than 20% and is determined to be an outlier.

[0021] Preferably, the lag order of the causal test (Granger) described in S3 is set to k=2-5, and the time window is set to T=30-90 seconds.

[0022] Preferably, the S5 specifically includes:

[0023] S5.1, Digital Twin Environment Construction: Build a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning;

[0024] S5.2, Reinforcement Learning Strategy Training: A Deep Q Network agent learns strategies for regulating valve opening and pump speed with the goal of minimizing pressure fluctuations, incorporating safety constraints to ensure compliance with engineering boundary conditions.

[0025] S5.3, Intervention Strategy Execution and Feedback: Control instructions are sent to edge devices, including valves and pumping stations, with a delay range of 10 to 500 ms through the data acquisition and supervision system (SCADA), to adjust the pipeline network status in real time and form a closed-loop control.

[0026] A water hammer control system for a water supply network based on the integration of multifunctional modules, comprising:

[0027] Pressure propagation map construction module: collects raw pressure data from sensors and constructs the raw data into a sparse relationship graph, collects pressure bus data, calculates pipeline interconnection information, and sends the pipeline interconnection results to the time series graph neural network prediction and traceability module for risk prediction and propagation path tracing;

[0028] Time-series graph neural network prediction and tracing module: Builds a graph neural network model based on causal constraints, calculates the results in the pressure propagation map construction module, realizes water hammer risk prediction and propagation path tracing, and passes the prediction results to the reinforcement learning intervention decision module.

[0029] Reinforcement learning intervention decision module: Analyzes the output results of the time series neural network prediction and traceability module, generates active intervention strategies through the reinforcement learning agent, and sends them to the pump station and valves through the control bus. After the strategy is executed, it re-reads the latest pressure data and makes learning adjustments based on the pressure data, forming a self-closed "prediction → control → re-learning" cycle.

[0030] Preferably, the pressure propagation map construction module includes:

[0031] Pipeline network digital modeling module: Based on the geographic information system (GIS) and design drawings, it builds a pipeline network topology map, deploys sensors, and records pipeline attributes;

[0032] Real-time data acquisition and preprocessing module: The sensor collects pipeline pressure data at 0.5-2Hz, fills missing values ​​through cubic spline interpolation, and uses the triple standard deviation principle (3σ) combined with EPANET software simulation to eliminate outliers;

[0033] Causal graph generation and model input module: performs causal testing (Granger) on preprocessed data (lag order k = 2~5, time window set to T = 30~90 seconds) to generate a sparse causal adjacency matrix as input to the graph neural network (TGNN).

[0034] Preferably, the temporal graph neural network prediction and traceability module adopts a causal-constrained graph neural network (TGNN) model, including:

[0035] Input layer: multidimensional time series, including pressure, first-order derivative, fluctuation intensity plus causal adjacency matrix;

[0036] Feature extraction layer: It consists of two layers of graph convolutional networks (GCNs) to capture the spatial features of the pipeline network topology;

[0037] Temporal modeling layer: includes a layer of gated recurrent units (GRUs) to extract the temporal dependencies of the pressure sequence;

[0038] Attention layer: A multi-head attention mechanism (4 heads) is introduced within the model to enhance the perception of key nodes and key paths, enable learning of pressure propagation paths, and enhance the weight distribution of key nodes and paths; the attention layer is configured to weight the features of key nodes and key paths within the model to enhance the weight distribution of key nodes and key paths.

[0039] Output layer: includes the probability of water hammer occurrence (0-1) plus the causal propagation chain vector (traceability path). It not only includes whether a water hammer event will occur, but also can trace back the corresponding causal propagation chain, that is, from the trigger point to the wave propagation path, to assist operation and maintenance personnel in accurate traceability and positioning.

[0040] Preferably, the reinforcement learning intervention decision module includes:

[0041] Digital twin environment building module: Builds a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning;

[0042] Reinforcement Learning Strategy Training Module: The Deep Q Network (DQN) agent learns strategies for regulating valve opening and pump speed, aiming to minimize pressure fluctuations. It also incorporates safety constraints to ensure compliance with engineering boundary conditions. Its input is connected to the output of the Graph Neural Network (GNN) prediction and traceability module, which generates intervention strategies based on water hammer risk probability and propagation path information.

[0043] Intervention strategy execution and feedback module: includes an execution module and a feedback collection submodule; the execution module has its input connected to the output of the reinforcement learning strategy training module and is used to convert the intervention strategy into control instructions for valves or pump stations and issue them via the control bus. The feedback collection submodule has two output paths:

[0044] (1) The real-time pressure sensor readings after execution are fed into the time series graph neural network prediction and traceability module in the form of original multidimensional time series, and its sliding time window is updated as the input tensor for the next prediction cycle;

[0045] (2) Based on the real-time pressure sensor readings, a reward signal is calculated and output to the reinforcement learning strategy training module for adjusting the intervention strategy; through the cycle of "prediction-strategy generation-instruction issuance-data and reward feedback-re-prediction", a closed-loop adaptive water hammer control of the water supply network is formed.

[0046] Compared with the existing technology, the present invention provides a water hammer control method and system for a water supply network based on the integration of multifunctional modules, which has the following beneficial effects:

[0047] (1) Combining graph structure and time series modeling to improve prediction accuracy and positioning capabilities;

[0048] Abstracting the water supply network into a graph structure and combining it with time series modeling can comprehensively reflect the propagation path of pressure waves, overcome the blind spots of traditional local modeling, and improve the comprehensiveness and accuracy of water hammer event detection.

[0049] (2) Introducing causal graphs to guide temporal graph neural network (TGNN) modeling to improve model interpretability and stability;

[0050] The results of causal analysis are used as the adjacency matrix input to effectively eliminate the interference of irrelevant edges in the topology on prediction, thereby improving the learning efficiency and causal traceability ability of the model.

[0051] (3) Realize closed-loop control of prediction-traceability-intervention;

[0052] Through the reinforcement learning module, the system not only realizes passive detection, but also actively adjusts the operation status of the pipeline network, thereby reducing water hammer damage and forming an intelligent closed-loop control system.

[0053] (4) Engineering deployability and versatility;

[0054] The system has a clear structure and distinct modules. It can be deployed on the existing water supply monitoring platform and is adaptable to water supply networks of various sizes and complexities. It has broad prospects for promotion. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1This is a schematic diagram of the steps of a water hammer control method for a water supply network based on the integration of multifunctional modules of the present invention;

[0056] Figure 2 Schematic diagram of the overall process of the water hammer warning and intelligent intervention system of the present invention;

[0057] Figure 3 A schematic diagram of a process for constructing a cause-effect diagram of the present invention;

[0058] Figure 4 Schematic diagram of the structure of the temporal graph neural network (TGNN) of the present invention;

[0059] Figure 5 This is a schematic diagram of the reinforcement learning intervention training process of the present invention. DETAILED DESCRIPTION

[0060] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0061] The present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0062] Example 1, please refer to Figures 1 to 5 As shown:

[0063] To solve the problems mentioned in the technical solution, the embodiment of the present application provides a water hammer control system for a water supply network based on the integration of multifunctional modules, including:

[0064] 1. Pressure propagation map construction module, responsible for pipe network topology modeling and cause-effect diagram generation;

[0065] The pressure propagation map construction module includes:

[0066] (1) Pipeline network digital modeling module: Based on the geographic information system (GIS) and design drawings, the pipeline network topology map is constructed, sensors are deployed, and pipeline attributes are recorded;

[0067] (2) Real-time data acquisition and preprocessing module: The sensor collects pipeline pressure data at 0.5-2 Hz, fills missing values ​​through cubic spline interpolation, and eliminates outliers by using the triple standard deviation principle (3σ) combined with EPANET software simulation;

[0068] (3) Causal graph generation and model input module: Perform causal testing (Granger) on the preprocessed data (lag order k = 2~5, time window set to T = 30~90 seconds) to generate a sparse causal adjacency matrix as the input of the graph neural network (TGNN).

[0069] 2. The Time Series Graph Neural Network prediction and traceability module uses a causal-constrained graph neural network (TGNN) model to predict water hammer risks and trace the transmission path.

[0070] The time series graph neural network prediction and traceability module adopts the causal constraint graph neural network (TGNN) model, including:

[0071] (1) Input layer: multidimensional time series, including pressure, first-order derivative, fluctuation intensity and causal adjacency matrix;

[0072] (2) Feature extraction layer: It consists of two layers of graph convolutional networks (GCNs) to capture the spatial features of the pipeline network topology;

[0073] (3) Temporal modeling layer: including a layer of gated recurrent unit (GRU) to extract the temporal dependency of the pressure sequence;

[0074] (4) Attention layer: A multi-head attention mechanism (4 heads) is introduced into the model to enhance the perception of key nodes and key paths, realize the learning of pressure propagation paths, and enhance the weight distribution of key nodes or paths;

[0075] (5) Output layer: The probability of water hammer occurrence (0-1) plus the causal propagation chain vector (tracing path) not only includes whether a water hammer event will occur, but also can trace back the corresponding causal propagation chain, that is, from the trigger point to the wave propagation path, to assist operation and maintenance personnel in accurate tracing and positioning.

[0076] 3. Reinforcement learning intervention decision module: The output results of the time series neural network prediction and traceability module are analyzed, and an active intervention strategy is generated through the reinforcement learning agent. The strategy is sent to the pump station and valve through the control bus. After the strategy is executed, the latest pressure data is re-read and learning adjustments are made based on the pressure data, forming a self-closed "prediction → control → re-learning" cycle.

[0077] Among them, the reinforcement learning intervention decision module includes:

[0078] (1) Digital twin environment construction module: Construct a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning;

[0079] (2) Reinforcement learning strategy training module: The deep Q network (DQN) agent aims to minimize pressure fluctuations and learns strategies for regulating valve opening and pump speed, combining safety constraints to ensure that operations meet engineering boundary conditions. Its input is connected to the output of the time-series graph neural network prediction and traceability module, which is used to generate intervention strategies based on water hammer risk probability and propagation path information.

[0080] (3) Intervention strategy execution and feedback module: It includes an execution module and a feedback acquisition submodule; the input end of the execution module is connected to the output end of the reinforcement learning strategy training module, and is used to convert the intervention strategy into a control instruction for the valve or pump station and issue it through the control bus. The feedback acquisition submodule has two output paths:

[0081] The real-time pressure sensor readings after execution are fed into the time series graph neural network prediction and traceability module in the form of original multi-dimensional time series, and its sliding time window is updated as the input tensor for the next prediction cycle;

[0082] The reward signal is calculated based on the pressure data and output to the reinforcement learning strategy training module for adjusting the intervention strategy. Through the cycle of "prediction-strategy generation-instruction issuance-data and reward feedback-re-prediction", a closed-loop adaptive water hammer control of the water supply network is formed.

[0083] The active intervention of the present invention adopts a discrete action deep Q network (DQN), whose input-output is defined as follows.

[0084] Input state vector: Valve real-time opening: θ1…θ m , m adjustable valves, each expressed as a percentage (%);

[0085] Pressure observation: instantaneous pressure P1…P of n sensors n (MPa);

[0086] Pressure change rate: corresponding first-order derivative dP1 / d t …dP n / d t (MPa·s -1 )

[0087] Therefore, the overall input variables are roughly = [θ1…θ m ,P1…P n , dP1 / d t …dP n / d t ] T

[0088] Output action vector: Give {x1, x2, x3...x n}N relative openings in total;

[0089] For each variable frequency pump, {y1, y2, y3...y n}There are n frequency levels in total.

[0090] Where θ represents the valve opening; P represents the instantaneous pressure; dP1 / d t Expressed as the corresponding first-order derivative, dP n / d t Expressed as the corresponding n-th order derivative, x represents the opening, and y represents the frequency.

[0091] The DQN network outputs the Q values ​​of all action combinations, selects the largest one and executes the corresponding valve-pump speed adjustment instruction.

[0092] Rewards (calculated internally during the training phase) are divided into two objectives: the first encourages reducing pressure deviations, while the second penalizes excessive operation. As long as the network keeps the overall pressure close to the design value and operates moderately, it receives positive incentives; otherwise, it is penalized. After repeated learning under numerous simulated conditions, the model learns to directly output control commands for the aforementioned valves and pumps within milliseconds. These commands are the only outputs of the reinforcement learning module of this invention.

[0093] Example 2, please refer to Figures 1 to 5 As shown:

[0094] This solution is based on a multifunctional module fusion water hammer control method for water supply network. The steps are as follows:

[0095] Step 1: Digital modeling of the pipeline network: Based on the Geographic Information System (GIS) and design drawings, a pipeline network topology map is constructed, sensors are deployed, and pipeline attributes are recorded;

[0096] The pipeline network is abstracted into a heterogeneous graph G = (V, E), and pressure sensors are deployed at bifurcations and upstream and downstream of valves. The pressure sensor deployment spacing is ≤ 500 meters, and the recorded pipeline attributes include material, length, and diameter.

[0097] Example: Node deployment: Node V = {v1,v2,...,v n Corresponding pressure sensors (deployed at bifurcations, upstream and downstream of valves, with a spacing of ≤500 meters);

[0098] Edge attributes: Edge E = {e ij} represents a pipeline, and the edge attributes include length L ij (m), pipe diameter D ij (mm), roughness (mm), historical leakage times C ij .

[0099] Where V represents a node list: the physical locations on the pipeline network where sensors can be installed or controls can be performed—such as bifurcations, upstream and downstream of valves, and pump station inlets—are numbered v1, v2…v n , this long string of numbers is called the node list V.

[0100] (1) E represents the pipe section list: as long as the number v i and v j There is a direct pipe connection between them, so we record it as an edge e. ij , and write the length, diameter, material and other information of that pipe on this side; all such e ij Together they make E.

[0101] (2) The subscripts i and j correspond to the starting node vᵢ and the ending node vⱼ, respectively; therefore, eᵢⱼ is “the specific section of the pipeline that flows from node i to node j”, which uniquely identifies the physical pipe segment between these two points and its properties.

[0102] Step 2: Real-time data acquisition and preprocessing: The sensor collects pipeline pressure data at 0.5-2 Hz, fills missing values ​​through cubic spline interpolation, and uses the triple standard deviation principle (3σ) combined with EPANET software simulation to eliminate outliers;

[0103] Among them, missing values ​​were filled by cubic spline interpolation, manual verification was performed when the correlation coefficient of adjacent nodes was less than 0.7, and outliers were judged to be outliers when the deviation was greater than 20%.

[0104] Step 3: Causal graph generation and model input: Perform causal testing (Granger) on the preprocessed data (lag order k = 2-5, time window set to T = 30-90 seconds) to generate a sparse causal adjacency matrix as input to the graph neural network (TGNN);

[0105] Among them, the lag order of the causal test (Granger) is set to k = 2~5, and the time window is set to T = 30~~90 seconds.

[0106] Causality test (Granger) was performed on historical data (time window T = 30-90 seconds, sliding step Δt = 1 second), where the lag order k = 2-5, p value ≤ 0.05 and F ij Significant causal edges ≥ 2.5, calculate the causal influence value F between nodes ij , generating an N×N causal influence matrix C.

[0107] Sparse processing: remove edges with p-value > 0.05 and retain significant causal edges F ij ≥2.5, forming a sparse adjacency matrix A∈{0,1} n × n .

[0108] This causal graph serves as the sparse adjacency matrix of TGNN, which can improve the model interpretability and reduce redundant noise edges in the topology.

[0109] Step 4: Water hammer risk prediction and tracing: The graph neural network (TGNN) model uses a "graph convolutional neural network (GCN) + gated recurrent unit (GRU) + attention mechanism" architecture to output the probability of water hammer occurrence and the causal propagation chain, locating the trigger point and propagation path;

[0110] Among them, the system operation method adopts the TGNN model with causal constraints;

[0111] The input layer of the graph neural network (TGNN) model: multidimensional time series (pressure, first-order derivative, fluctuation intensity) + causal adjacency matrix;

[0112] Feature extraction layer: a 2-layer graph convolutional network (GCN) to capture the spatial features of the pipe network topology;

[0113] Temporal modeling layer: 1 layer of gated recurrent unit (GRU) to extract the temporal dependency of the pressure sequence;

[0114] Attention layer: Multi-head attention mechanism (4 heads) to enhance the weight distribution of key nodes and paths;

[0115] Output layer: water hammer occurrence probability (0-1) + causal propagation chain vector (tracing path).

[0116] This solution introduces an attention mechanism within the model to enhance the perception of key nodes and critical paths, and realize the learning of pressure propagation paths. At the same time, the output of the graph neural network (TGNN) model not only includes whether a water hammer event will occur, but also can trace the corresponding causal propagation chain (that is, from the trigger point to the wave propagation path), assisting operation and maintenance personnel in accurate tracing and positioning.

[0117] Step 5: Build a hydraulic simulation environment for the pipe network using the EPANET platform. Within the EPANET digital twin, the Deep Q Network agent generates intervention actions based on pressure distribution and predicted positions, with minimizing pressure fluctuations as the reward objective. The control strategy is executed in real time through the Supervisory Control and Data Acquisition (SCADA) system, forming a closed-loop control system. This includes the following steps:

[0118] (1) Digital twin environment construction: Construct a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning;

[0119] (2) Reinforcement learning strategy training: The Deep Q Network agent aims to minimize pressure fluctuations and learns strategies for regulating valve opening and pump speed, combining safety constraints to ensure that operations meet engineering boundary conditions.

[0120] (3) Intervention strategy execution and feedback: Control instructions are sent to edge devices, including valves and pump stations, with a delay range of 10 to 500 ms through the data acquisition and monitoring system (SCADA), to adjust the status of the pipeline network in real time, realize active control of water hammer risks and form a closed-loop control.

[0121] Example 3;

[0122] This solution differs from Example 2 in that:

[0123] This example selects three typical water supply network scenarios, covering different pipe diameters, materials, and water hammer types. Sensor deployment follows the design principles of this solution (at bifurcations / upstream and downstream of valves, with spacing ≤ 500m):

[0124] Table 1 is the experimental scenario and parameter statistics:

[0125]

[0126] Table 2 is a statistical table comparing the effects of accurate prediction and traceability:

[0127]

[0128] Table 3 is a statistical table of the effects of active intervention and closed-loop control;

[0129]

[0130] Detection accuracy is 15% to 20% higher than traditional LSTM, and traceability error is kept within a distance of ≤2 nodes (approximately 1000 meters), meeting the precision requirements of engineering positioning. By combining a causal adjacency matrix (graph structure), temporal features (GRU), and an attention mechanism, this method captures the spatial causal chain and temporal dependencies of pressure wave propagation, addressing the local modeling blind spots and lack of traceability inherent in traditional methods.

[0131] Peak pressure was reduced by 38.9% to 40%, and peak arrival time was delayed by 40% to 59%, far exceeding traditional methods (which lack proactive intervention capabilities) and meeting the "golden 10 seconds" requirement for water hammer emergencies. By training a DQN agent in the EPANET digital twin environment and issuing millisecond-level control commands, a closed-loop "prediction-strategy generation-execution" system was achieved.

[0132] Table 4 shows the statistics of cross-scenario generalization capabilities:

[0133]

[0134] Accuracy fluctuates less than 5% across different scenarios, significantly outperforming traditional LSTM models (fluctuation >12%). The model is suitable for pipes ranging from DN200 to DN2000 and mixed-material networks. Causal graph-driven modeling eliminates redundant topological edges, allowing the model to focus on the actual pressure propagation path and reduce the impact of differences in pipe diameter and material on prediction.

[0135] The above table shows that this solution has the following advantages over the existing technology:

[0136] (1) Accurate prediction and traceability: By integrating the pipe network topology and time series data (sampling frequency 0.5-2 Hz), the water hammer detection accuracy is increased to over 90%, and the traceability error is ≤ 2 node spacing.

[0137] (2) Active intervention and closed-loop control: Generate a regulation strategy 1 to 3 seconds before the water hammer peak arrives, reducing the pressure fluctuation amplitude by 38.9% to 40% and delaying the peak propagation speed by 40% to 59%.

[0138] (3) Cross-scenario generalization capability: By driving modeling with causal graphs, the model’s prediction accuracy fluctuates by less than 5% in pipe networks with different pipe diameters (DN200-DN2000) and materials (steel pipes / plastic pipes), significantly outperforming traditional methods (fluctuation >12%).

[0139] Example 4;

[0140] In order to further verify the performance of this solution in complex pipe networks, a mixed material pipe network (DN200-DN2000, 200 nodes) was selected for experiment. Figure 3 As shown in the figure, the pressure time series data is first preprocessed, including missing value filling and outlier removal. Then, the data segments are divided through the time window sliding strategy (such as 30~90 second window, 1 second step), and the Granger causality test (lag 2~5 orders) is performed on each segment of data. The causal influence value between nodes is calculated and the causal influence matrix is ​​generated; redundant edges are removed through sparse processing (significant causal edges with p value ≤ 0.05 are retained), and finally a sparse causal adjacency matrix for the temporal graph neural network (TGNN) model is output.

[0141] like Figure 4 As shown in the figure, the input layer receives pressure time series data (including pressure value, first-order derivative, and fluctuation intensity) and the causal adjacency matrix. After the preprocessing layer calculates the dynamic features, the graph convolutional layer (GCN) extracts the spatial features of the pipeline network topology; then the GRU layer captures the temporal dependency, and the attention layer weights the key nodes and paths. Finally, the output layer outputs the probability of water hammer occurrence and the causal propagation chain to achieve risk prediction and traceability.

[0142] like Figure 5As shown in the figure, the rectangular box represents the EPANET digital twin pipeline network (environment), DQN network module and experience replay pool. The arrows mark the flow of data such as state, action, and reward, showing the reinforcement learning closed-loop training process: the environment outputs states such as pressure distribution to the DQN network module. The module generates actions such as adjusting valve / pump speed based on the state and acts on the environment. The environment feeds back rewards related to pressure fluctuations. The module stores the interaction experience (including the next state) in the replay pool and uses it to update the network parameters. The experience replay mechanism and DQN algorithm are used to achieve strategy optimization, and the intervention efficiency is improved with the goal of minimizing pressure fluctuations.

[0143] As shown in Table 5, in 50 simulated water hammer experiments on a mixed-material pipe network (DN200-DN2000, 200 nodes, sensor spacing ≤500 meters, sampling frequency 1Hz), our proposed solution achieved a detection accuracy of 93.2%, a 14.7% improvement over the traditional LSTM (78.5%) and an 11.1% improvement over the traditional graph convolutional network (GCN) (82.1%). The traceability error was ≤1.5 nodes, only 50% of the traditional GCN (≤3.0 nodes), demonstrating that it overcomes the blind spots of traditional local modeling and achieves comprehensive and accurate detection and location of water hammer events. Model training for our proposed solution took only 8 hours, a 20%-33% reduction compared to traditional LSTM (12 hours) and GCN (10 hours). The number of data dependencies was only 50,000, a 37.5% reduction compared to traditional methods, and the cross-scenario accuracy fluctuation was less than 5% (Table 3), demonstrating that causal constraints improve learning efficiency and model stability.

[0144] Table 5 is a statistical table comparing detection accuracy and traceability:

[0145]

[0146] As shown in Table 6, in a simulated water hammer event, this solution can generate an intervention strategy 1 to 3 seconds before the pressure peak arrives, reducing the peak pressure by 41.7% and delaying the peak propagation speed by 38%. This is 1.7 to 2.5 times higher than manual intervention (reduction of 20 to 25% and delay of 10 to 15%). The error rate is only 0.2 times per month, far lower than the 3 to 5 times per month of manual scheduling, achieving a breakthrough from passive alarm to active control.

[0147] Table 6 is a statistical table comparing intervention response speed and effect:

[0148]

[0149] As can be seen from the above table, the specific advantages of this solution are as follows:

[0150] (1) Accurate and efficient: The detection accuracy is 93.2%, significantly higher than the traditional model (LSTM / GCN increased by about 11% to 15%); the traceability error is ≤1.5 nodes, and the positioning accuracy far exceeds the graph convolutional network (GCN) (≤3 nodes) and manual (>5 nodes).

[0151] (2) Rapid and proactive intervention: Generate strategies 1 to 3 seconds before the pressure peak arrives, with a response speed of milliseconds (manual response requires more than 10 minutes); reduce the pressure peak by 41.7% (manual response requires only 20% to 25%), delay the peak propagation by 38% (manual response requires 10% to 15%), and improve performance by 1.7-2.5 times.

[0152] (3) Low resource dependence: Training takes 8 hours (33% to 50% less than LSTM / GCN), and only 50,000 data points are required (lower than traditional models).

[0153] (4) High reliability: The misoperation rate is 0.2 times / month, which is less than 1 / 15 of manual scheduling (3-5 times / month), completely avoiding the passive defects of traditional alarm systems.

[0154] In summary, from Examples 1 to 4, it can be seen that compared with the prior art, this solution has the following significant advantages:

[0155] (1) Combining graph structure and time series modeling to improve prediction accuracy and positioning capabilities;

[0156] Abstracting the water supply network into a graph structure and combining it with time series modeling can comprehensively reflect the propagation path of pressure waves, overcome the blind spots of traditional local modeling, and improve the comprehensiveness and accuracy of water hammer event detection.

[0157] (2) Introducing causal graphs to guide temporal graph neural network (TGNN) modeling to improve model interpretability and stability;

[0158] The results of causal analysis are used as the adjacency matrix input to effectively eliminate the interference of irrelevant edges in the topology on prediction, thereby improving the learning efficiency and causal traceability ability of the model.

[0159] (3) Realize prediction-traceability-intervention closed-loop control;

[0160] Through the reinforcement learning module, the system not only realizes passive detection, but also actively adjusts the operation status of the pipeline network, thereby reducing water hammer damage and forming an intelligent closed-loop control system.

[0161] (4) Engineering deployability and versatility;

[0162] The system has a clear structure and distinct modules. It can be deployed on the existing water supply monitoring platform and is adaptable to water supply networks of various sizes and complexities. It has broad prospects for promotion.

[0163] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the process, method, article, or device comprising the element.

[0164] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A water hammer control method for a water supply network based on the integration of multifunctional modules, characterized in that: include: S1, Pipeline Network Digital Modeling: Based on the geographic information system and design drawings, a pipeline network topology map is constructed, sensors are deployed, and pipeline attributes are recorded. The sensors are pressure sensors, and are deployed at pipeline bifurcations and upstream and downstream of valves. The pressure sensors are deployed at a spacing of ≤ 500 meters. The recorded pipeline attributes include material, length, and diameter. S2, real-time data acquisition and preprocessing: The sensor collects pipeline pressure data at 0.5-2Hz, fills missing values ​​through cubic spline interpolation, and eliminates outliers by using the triple standard deviation principle combined with EPANET software simulation; S3, causal graph generation and model input: performs causal testing on preprocessed data to generate a sparse causal adjacency matrix as input to the graph neural network; S4, Water Hammer Risk Prediction and Tracing: The graph neural network model uses a "graph convolutional neural network + gated recurrent unit + attention mechanism" architecture to output the probability of water hammer occurrence and the causal propagation chain, locating the trigger point and propagation path; S5, Reinforcement Learning and Intelligent Intervention: A pipeline network hydraulic simulation environment is built using the EPANET platform. Within the EPANET platform's digital twin environment, the Deep Q Network (DQN) agent generates intervention actions based on pressure distribution and predicted position, with minimizing pressure fluctuations as the reward objective. Control strategies are executed in real time through the data acquisition and monitoring system, forming a closed-loop control system. The S5 specifically includes: S5.1, Digital Twin Environment Construction: Build a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning; S5.2, Reinforcement Learning Strategy Training: The Deep Q Network (DQN) agent aims to minimize pressure fluctuations and learns strategies for regulating valve opening and pump speed, incorporating safety constraints to ensure that operations meet engineering boundary conditions. S5.3, intervention strategy execution and feedback: Control instructions are sent to edge devices, including valves and pump stations, with a delay range of 10-500ms through the data acquisition and monitoring system to regulate the pipeline network status in real time and form a closed-loop control.

2. A water hammer control method for a water supply network based on multifunctional module integration according to claim 1, characterized in that: As described in S2, missing values ​​were filled by cubic spline interpolation, and manual verification was performed when the correlation coefficient of adjacent nodes was less than 0.

7. The abnormal value deviation was greater than 20% and was determined to be an outlier.

3. The water hammer control method for a water supply network based on multifunctional module integration according to claim 1 is characterized in that: The lag order of the causal test described in S3 is set to k = 2~5, and the time window is set to T = 30~90 seconds.

4. A water hammer control system for a water supply network based on multifunctional module fusion, applicable to a water hammer control method for a water supply network based on multifunctional module fusion according to any one of claims 1 to 3, characterized in that: include: Pressure propagation map construction module: collects raw pressure data from sensors and constructs it into a sparse relationship graph. It also collects data based on the pressure bus, calculates pipeline interconnection information, and sends the pipeline interconnection results to the time series graph neural network prediction and traceability module for risk prediction and propagation path traceability. Time-series graph neural network prediction and traceability module: This module builds a graph neural network model based on causal constraints, calculates the results from the pressure propagation map construction module, implements water hammer risk prediction and propagation path traceability, and passes the prediction results to the reinforcement learning intervention decision module. Reinforcement learning intervention decision module: This module analyzes the output results of the time-series graph neural network prediction and traceability module, generates active intervention strategies through the reinforcement learning agent, and sends them to the pump station and valves through the control bus. After the strategy is executed, it re-reads the latest pressure data and makes learning adjustments based on the latest pressure data, forming a self-closed "prediction → control → re-learning" cycle.

5. A water hammer control system for a water supply network based on multifunctional module integration according to claim 4, characterized in that: The pressure propagation map construction module includes: Pipeline network digital modeling module: Based on the geographic information system and design drawings, it builds a pipeline network topology map, deploys sensors, and records pipeline attributes; Real-time data acquisition and preprocessing module: The sensor collects pipeline pressure data at 0.5-2Hz, fills missing values ​​through cubic spline interpolation, and eliminates outliers based on the triple standard principle combined with EPANET software simulation; Causal graph generation and model input module: performs causal testing on preprocessed data and generates a sparse causal adjacency matrix as input to the graph neural network.

6. The water hammer control system for a water supply network based on multifunctional module integration according to claim 4 is characterized in that: The temporal graph neural network prediction and traceability module adopts a causal-constrained graph neural network model, including: Input layer: multidimensional time series, including pressure, first-order derivative, fluctuation intensity plus causal adjacency matrix; Feature extraction layer: It consists of two layers of graph convolutional networks and is used to capture the spatial characteristics of the pipeline network topology; Temporal modeling layer: includes a layer of gated recurrent units to extract the temporal dependencies of the pressure sequence; Attention layer: A multi-head attention mechanism is introduced within the model to enhance the perception of key nodes and key paths, enable learning of pressure propagation paths, and enhance the weight distribution of key nodes and paths. The attention layer is configured to weight the features of key nodes and key paths within the model to enhance the weight distribution of key nodes and key paths. Output layer: includes the probability of water hammer occurrence plus the causal propagation chain vector. It not only includes whether a water hammer event will occur, but also can trace back the corresponding causal propagation chain, that is, from the trigger point to the wave propagation path, to assist operation and maintenance personnel in accurate tracing and positioning.

7. The water hammer control system for a water supply network based on multifunctional module integration according to claim 4 is characterized in that: The reinforcement learning intervention decision module includes: Digital twin environment building module: Builds a pipeline network hydraulic simulation environment on the EPANET platform to simulate real-time pressure distribution and intervention effects, providing training scenarios for reinforcement learning; Reinforcement Learning Strategy Training Module: The Deep Q Network (DQN) agent learns adjustment strategies for valve opening and pump speed with the goal of minimizing pressure fluctuations, incorporating safety constraints to ensure compliance with engineering boundary conditions. Its input is connected to the output of the Graph Neural Network (GNN) prediction and traceability module, which generates intervention strategies based on water hammer risk probability and propagation path information. Intervention strategy execution and feedback module: includes an execution module and a feedback collection submodule; the execution module has its input connected to the output of the reinforcement learning strategy training module and is used to convert the intervention strategy into control instructions for valves or pump stations and issue them via the control bus. The feedback collection submodule has two output paths: (1) The real-time pressure sensor readings after execution are fed into the time series graph neural network prediction and traceability module in the form of original multidimensional time series, and its sliding time window is updated as the input tensor for the next prediction cycle; (2) Based on the real-time pressure sensor reading, a reward signal is calculated and output to the reinforcement learning strategy training module for adjusting the intervention strategy; through the cycle of "prediction-strategy generation-instruction issuance-data and reward feedback-re-prediction", a closed-loop adaptive water hammer control of the water supply network is formed.

Images