A method and system for real-time water network pressure balancing control

Abstract

The present application relates to a kind of real-time water network pressure balance control method and system, decision-making method includes the following steps: obtaining each node real-time water pressure data and input pre-trained water pressure regulating model, obtain the pressure guide value of each node of water network, realize water network pressure balance control, wherein the water pressure regulating model includes: water network pressure evaluation network, for obtaining water network pressure score according to the each node real-time water pressure data;Water network pressure control network is used to obtain the pressure guide value of each node of water network according to the each node real-time water pressure data and the water network pressure score, the water network pressure control network includes one or more graph convolutional neural network layers.Compared with prior art, the present application designs neural network model by using more suitable graph convolutional neural network for water network structure, solves the problem that fully connected neural network cannot capture interdependent correlation message between nodes.

Description

Technical Field

[0001] This invention relates to the field of smart water management, and in particular to a method and system for real-time water network pressure balance control. Background Technology

[0002] In the era of intelligence, the development of smart water management is rapid. It is essential to fully explore the value and logical relationships of data to achieve intelligent control and precise management of water management systems, ensuring the safe operation of water facilities. Smart water management requires the establishment of an intelligent control system, building a water management model based on historical water network data.

[0003] Currently, water management models are mainly built through two methods: manual design by water management experts and construction based on traditional fully connected neural networks. The advantage of expert-designed models is that they require less historical data, and experts can extract model features based on their experience to quickly build the model. The disadvantage is that they require significant manpower for modeling and testing, and the model's quality largely depends on the expert's experience. The advantage of fully connected neural networks is that they can use data-driven feature extraction, facilitating modification and testing and reducing manual intervention. The disadvantage is that they require a large amount of data and the models are often overly simplistic.

[0004] Chinese invention publication CN113756394A discloses a performance testing method for a water supply network pressure management system based on a pressure reducing valve. The basic steps are as follows: Select the performance characteristics of the pressure management system to be tested; if testing the accuracy and sensitivity of pressure regulation, select the test condition, set the pressure after the pressure reducing valve, and compare the actual pressure with the set pressure using data displayed by an electromagnetic flowmeter to test the accuracy of pressure adjustment; test the sensitivity of pressure adjustment based on the speed at which the actual pressure reaches the set pressure; if testing the pressure stabilization effect of the pressure management system, keep the pressure after the pressure reducing valve constant, repeatedly adjust the pressure before the booster pump in the water tank, and test the pressure stabilization effect using pressure data from the electromagnetic flowmeter; if testing the leakage control effect of the pressure management system on the water supply network, simulate a leak point in the network using a faucet, adjust the pressure after the pressure reducing valve, and determine the relationship between pressure and leakage using flow data from the electromagnetic flowmeter to test the reduction in leakage before and after the use of the pressure management system. This invention designs different testing methods for different types of pressure management system performance tests. The method involves the design of parameters such as the control pressure after the pressure reducing valve. These parameters require human modeling and extensive testing, and the quality of the model largely depends on experience.

[0005] Chinese invention publication CN113970073A discloses a method for accurate location of leaks in water supply networks based on ResNet, comprising the following steps: constructing a network model for accurate location of leaks in water supply networks, wherein the network model includes a ResNet network, a classification module, and a regression module; training the network model to obtain an ideal network for accurate location of leaks in water supply networks; collecting pressure data from multiple nodes in the water supply network using pressure sensors and inputting the pressure data into the ideal network for accurate location of leaks in water supply networks; extracting features from the pressure data using the ResNet network, and inputting the results of the feature extraction into both the classification and regression modules; obtaining the predicted leakage probability value for each pipe through the classification module and the predicted leakage location for each pipe through the regression module; selecting the pipe with the highest predicted leakage probability value and using the predicted leakage location corresponding to that pipe as the network leak location result. This method uses a residual neural network to construct the location network model. Since fully connected neural networks cannot capture the inter-node dependencies, the feature extraction capability of the neural network model is poor, requiring a large amount of high-quality data for training.

[0006] Chinese utility model publication CN215929235U discloses a regional pressure management system for a water supply network. This system uses multiple remote wireless pressure monitoring units to remotely monitor water supply pipeline pressure, obtaining timely and accurate pressure data. When a pipeline ruptures and leaks, the pressure drops, allowing for timely detection of leaks and ruptures. However, this patent uses GPRS for communication, which has a slow signal transmission speed and a narrow network spectrum, preventing complete coverage. Furthermore, it cannot automatically regulate the network pressure to ensure dynamic stability. This utility model can only determine the current state of the fault area and employs a two-stage decision-making process.

[0007] In summary, existing methods for real-time water network pressure balance control have the following drawbacks:

[0008] (1) Existing water model design mainly relies on expert experience. Model design requires a lot of manpower. At the same time, the performance of the model depends too much on the expert's own knowledge reserves, resulting in a large difference in the performance of different model designs. Moreover, the testing is complicated, which is not conducive to large-scale and rapid deployment in the era of smart water.

[0009] (2) Existing technologies use traditional fully connected neural network models, which cannot rely on the characteristics of the water network itself, resulting in poor feature extraction performance of the model. At the same time, a large amount of high-quality data is required during training.

[0010] (3) When training neural network models, the existing technology mainly sets the goal of classification tasks. The model obtained by this training can only reflect the problem of a certain node, but cannot give suggestions on the pressure information related to the node. In the end, the problem of how to regulate still depends on the experience of water experts, and cannot fully realize intelligent control. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and apparatus for real-time water network pressure balance control that can effectively improve detection accuracy without relying on expert model design.

[0012] The objective of this invention can be achieved through the following technical solutions:

[0013] According to one aspect of the present invention, a method for real-time water network pressure balance control is provided, comprising the following steps: acquiring real-time water pressure data of each node and inputting it into a pre-trained water pressure regulation model to obtain pressure guidance values ​​for each node of the water network and achieve water network pressure balance control, wherein the water pressure regulation model comprises: a water network pressure assessment network for obtaining a water network pressure score based on the real-time water pressure data of each node; and a water network pressure control network for obtaining pressure guidance values ​​for each node of the water network based on the real-time water pressure data of each node and the water network pressure score, wherein the water network pressure control network comprises one or more graph convolutional neural network layers.

[0014] As a preferred technical solution, the water network pressure control network includes one or more linear layers.

[0015] As a preferred technical solution, the number of graph convolutional neural subnetworks in the water network pressure control network is 1-5.

[0016] As a preferred technical solution, the water network pressure assessment network includes at least one graph convolutional neural network layer.

[0017] As a preferred technical solution, the water pressure regulation model further includes a supervised generative network, which is used to obtain training evaluation indicators for constructing a loss function based on the pressure guidance values ​​of each node of the water network during the model training process. The supervised generative network includes a sub-network with the same structure as the water network pressure assessment network and a mathematical calculation layer. The mathematical calculation layer is used to output the training evaluation indicators based on the sub-network.

[0018] As a preferred technical solution, the steps for obtaining the pre-trained water pressure regulation model include:

[0019] Obtain a training sample set, wherein each sample in the training sample set includes real-time water pressure data of each node and ideal water pressure adjustment value of each node;

[0020] The water pressure regulation model is trained based on the training sample set. After the loss function value reaches the preset convergence condition, the pre-trained water pressure regulation model is obtained.

[0021] The independent variables of the loss function used to obtain the loss function value include the training evaluation index, the pressure guidance value, and the ideal water pressure regulation value for each node.

[0022] As a preferred technical solution, the loss function is calculated as follows:

[0023]

[0024] Among them, b j To train the evaluation metrics, c i For the pressure guidance value, a′ i The ideal water pressure adjustment value for each node is given.

[0025] According to another aspect of the present invention, a system for real-time water network pressure balance control is provided, comprising:

[0026] The input module is used to acquire real-time water network status information and, based on the real-time water network status information, acquire real-time water pressure data for each node.

[0027] The control module is used to obtain the pressure guidance value of each node in the water network through the water pressure regulation model based on the real-time water pressure data of each node.

[0028] The execution module is used to obtain the water pressure adjustment value of each node according to the pressure guidance value of each node in the water network, and send a water pressure adjustment signal that matches the water pressure adjustment value of each node to realize the water network pressure balance control.

[0029] As a preferred technical solution, the real-time water network pressure balance control system is deployed on an edge computing device.

[0030] As a preferred technical solution, the execution module includes a control unit, which is used to change the opening degree of the pressure reducing valve core according to the water pressure adjustment signal.

[0031] Compared with the prior art, the present invention has the following advantages:

[0032] (1) By using neural network training models instead of expert models, the correlation in the data can be fully explored, and decisions can be made quickly to a certain extent, so that the control can meet the requirements of real-time control and adapt to the requirements of rapid and large-scale deployment in smart water management.

[0033] (2) By using a graph convolutional neural network that better fits the water network structure instead of a traditional fully connected neural network, all data sample nodes in the graph network are associated with any real data sample node in the graph through edges. This information can be used to capture the interdependencies between nodes. Compared with existing technologies, the detection accuracy is improved by an order of magnitude, and transfer learning experience can be reused.

[0034] (3) By using a generative adversarial model training mechanism, the traditional training method can only determine the fault range. It can directly generate a pressure guidance value based on the water network status to directly control the hardware devices in the water network to take action. At the same time, compared with the traditional training method, the training data is faster and requires less data. Attached Figure Description

[0035] Figure 1 This is an architecture diagram of the neural network system model in the embodiment;

[0036] Figure 2 This is a diagram of the system model training framework in the embodiment;

[0037] Figure 3 This is a framework diagram of the real-time water network pressure balance control system in the embodiment.

[0038] Among them, 100 is the input module, 200 is the control module, and 300 is the execution module. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0040] Example 1

[0041] This embodiment provides a method for real-time water network pressure balance control, including the following steps: acquiring real-time water pressure data of each node and inputting it into a pre-trained water pressure regulation model to obtain the water pressure regulation value of each node.

[0042] like Figure 1The above is a diagram of the architecture of the water pressure regulation model. The water pressure regulation model includes a pre-trained evaluation network, a generator network, and a discriminator network. The pre-trained evaluation network is based on a graph convolutional neural network and is used to pre-train an evaluation network of the water network state. The main function of the evaluation network is to receive the pressure of each node in the water network, calculate the water network pressure score, and use it to evaluate the current state of the water network. The generator network consists of one linear layer and three graph convolutional neural network layers. The main function of the generator network is to receive the pressure and score of each node in the water network, calculate the water network pressure guidance value, and use it to assist in the water network pressure balance. The generator network is the core network in this embodiment. The discriminator network consists of a sub-network with the same structure as the pre-trained evaluation network and one mathematical computation layer. The main function of the discriminator network is to evaluate the quality of the water network pressure guidance value calculated by the generator network, while providing prior experience to accelerate network training.

[0043] In this embodiment, a Graph Convolutional Neural Network (GCN) is used instead of a fully connected Neural Network (DNN) as the underlying neural network of the model in this computation. Traditional Graph Neural Networks (GNNs) obtain the relationship information between nodes through graph traversal, and then use the information obtained from the traversal to extract features using a DNN. This model is only suitable for discovering the degree of relationship between nodes. This invention introduces convolution operations into the GNN, focusing on the relationship between each node and its surrounding nodes, and then performing feature extraction. GCN and CNN models have a very close common relationship. Essentially, both are information operations in the aggregation domain, only the data objects they operate on are different. CNN essentially stacks the pixels around a pixel according to different weights; in this invention, GCN stacks the neighbors around a node according to different weights.

[0044] like Figure 2The above is a training framework diagram of the real-time water network pressure balance model based on the generative adversarial mechanism in this embodiment. It mainly consists of three parts: a data input part, a generator part, and a discriminator part. The input part is mainly for generating the data needed by the model; how to evaluate the state of the water network is a discrimination problem; how to calculate a more reasonable water pressure value for the main nodes of the water network is a generation problem. The generator part and the discriminator part together constitute the generative adversarial training mechanism proposed in this invention. The discriminator is used to accelerate the training of the generator model and supervise the performance of the generator model. In the training framework diagram of the real-time water network pressure balance model based on the generative adversarial mechanism, the original input signal is the real-time water pressure data of each node in the water network. The current sample data X1 is input into the evaluation model 1, and the evaluation model 1 will output a score based on the input sample data. The score will be sent to the generator as prior experience to help the generator G generate reasonable water pressure control data X2. The representation in the neural network is the weight value, which determines the weight of the real-time data. At the same time, the current sample data X1 is sent to the generator G. Data X1 and data X2 are data structures with actual physical meaning. In the task of balancing water supply network pressure, data X1 will constrain the generator's output to a reasonably meaningful space and serve as a penalty to limit the generator model from generating data with large fluctuations. Generator G, based on the current sample data X1 and the score, generates a pressure generation sample X3 for each node under the current ideal state of the water network, which is then input into discriminator D. Finally, discriminator D, based on its experience, judges whether the data generated by generator G conforms to a good pressure distribution of each node in the water network, thereby supervising generator G to train a better model. This is the main idea behind generative adversarial networks (GANs).

[0045] The loss function for training a system model based on generative adversarial mechanisms is shown in the following equation:

[0046]

[0047] Among them, b j To score the discriminator model, c i To generate sample data, a′ i The water pressure data is from the original sample data.

[0048] This embodiment replaces expert models with a data-driven approach, fully exploring the correlations within the data and enabling rapid decision-making to a certain extent. This addresses the real-time control requirements of smart water systems and adapts to the demands of rapid and large-scale deployment in smart water management. Furthermore, by using a graph convolutional neural network (GCNN) design that better suits the water network structure, this embodiment solves the problem that fully connected neural networks cannot capture the inter-node dependencies. In the GCNN, all data sample nodes are connected to any real data sample node in the graph via edges. Compared to existing technologies, the detection accuracy is improved by an order of magnitude, and transfer learning experience can be reused.

[0049] Example 2

[0050] like Figure 3 The above is an architecture diagram of a real-time water network pressure balance control system, which consists of three parts: an input module, a control module, and an execution module.

[0051] Input Module. The input module of the real-time water network pressure balance control system has two data streams, which correspond to the two inputs of the generator network in Example 1. The two data streams are the input after the real-time water pressure data is calculated by the pre-trained evaluation network (corresponding to input A in the figure) and the input of the directly input real-time water pressure data (corresponding to input A' in the figure).

[0052] Control Module. The input module of the real-time water network pressure balance control system receives two data streams from the input module. Using a trained graph neural network model, it calculates the expected pressure distribution data of each major node in the water network at the current moment. The core of this module lies in the design of the water pressure regulation model. In this embodiment, a graph neural network is used as the underlying neural network of the water pressure regulation model. Pre-evaluation through a pre-trained evaluation network improves the completeness of the data. Finally, a training framework based on generative adversarial mechanisms is used to improve the performance of the entire real-time water network pressure balance control system.

[0053] The execution module of the real-time water network pressure balance control system provides water pressure guidance values ​​for the main nodes of the water network to the water pressure regulation model, assisting the original water supply system in regulating pressure and effectively improving the unreasonable high pressure phenomenon in the water network. The original water supply system integrates sensor technology and edge computing technology. Sensors detect the state variables of the controlled object and feed them back to the edge computing device. The proposed real-time water network pressure balance control system is deployed on the edge computing device. The pressure regulation model deployed on the edge computing device calculates the water network outlet pressure regulation value based on real-time data, sends a change signal, and the control unit controls the opening degree of the pressure reducing valve core according to the change signal, regulating the pressure at the end of the water network, reducing the leakage rate in the water network, and controlling the leakage level in the water network.

[0054] This embodiment utilizes a generative adversarial model training mechanism, overcoming the limitations of traditional training methods that only determine the fault range. It transforms the two-stage decision-making problem into a one-stage one, directly generating a pressure guidance value based on the water network status to control the actions of hardware devices within the network. Furthermore, compared to traditional training methods, training data is faster and requires less data. In addition, this embodiment uses a data-driven approach to accelerate the model building cycle in smart water management, adapting to the requirements of rapid and large-scale deployment, making real-time decision-making possible. By using a graph convolutional neural network as the underlying neural network model in the real-time water pressure balance decision-making system, the accuracy of feature extraction is improved, thereby enhancing the computational performance of the water pressure balance decision-making system. Finally, a supporting training framework for the pressure balance model is proposed. Through a generative adversarial mechanism, the original two-stage decision-making mode is changed, directly calculating the data that smart water hardware devices can process, thus achieving real-time decision-making capabilities in smart water management at the technical level. Ultimately, by balancing water network pressure in real time, leakage problems in the water supply network are reduced, providing support for the development of smart water network technology.

[0055] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for real-time water network pressure balancing control, characterized in that, The method comprises the following steps: obtaining real-time water pressure data of each node and inputting the pre-trained water pressure adjustment model to obtain pressure guide values of each node in the water network, thereby realizing water network pressure balance control, The water pressure adjustment model comprises: a water network pressure evaluation network configured to obtain a water network pressure score based on the real-time water pressure data of each node; a water network pressure control network configured to obtain pressure guide values of each node in the water network based on the real-time water pressure data of each node and the water network pressure score, wherein the water network pressure control network comprises one or more graph convolutional neural network layers.

2. The method of real-time water network pressure balancing control of claim 1, wherein, The water network pressure control network comprises one or more linear layers.

3. The method of real-time water network pressure balancing control of claim 1, wherein, In the water network pressure control network, the number of graph convolutional neural network layers is 1-5.

4. The method of real-time water network pressure balancing control of claim 1, wherein, The water network pressure evaluation network comprises at least one graph convolutional neural network layer.

5. The method of real-time water network pressure balancing control of claim 1, wherein, The water pressure adjustment model further comprises a supervision generation network configured to obtain a training evaluation index for constructing a loss function based on the pressure guide values of each node in the water network during model training, wherein the supervision generation network comprises a subnetwork having the same structure as the water network pressure evaluation network and a mathematical calculation layer configured to output the training evaluation index based on the subnetwork.

6. The method of real-time water network pressure balancing control of claim 1, wherein, The pre-trained water pressure adjustment model is obtained by: obtaining a training sample set, wherein each sample in the training sample set comprises real-time water pressure data of each node and ideal water pressure adjustment values of each node; training the water pressure adjustment model based on the training sample set, and obtaining the pre-trained water pressure adjustment model when a loss function value reaches a preset convergence condition, wherein the independent variables of the loss function for obtaining the loss function value comprise a training evaluation index, a pressure guide value, and ideal water pressure adjustment values of each node.

7. The method of real-time water network pressure balancing control of claim 6, wherein, The calculation function of the loss function is as follows: wherein, is the training evaluation index, is the pressure guide value, is the ideal node water pressure adjustment value.

8. A system for real-time water network pressure balancing control, characterized by, The method and system for realizing real-time water network pressure balance control according to any one of claims 1-7 comprise: an input module (100) configured to obtain real-time water network state information and obtain real-time water pressure data of each node based on the real-time water network state information; a control module (200) configured to obtain pressure guide values of each node in the water network by using a water pressure adjustment model based on the real-time water pressure data of each node; an execution module (300) configured to obtain water pressure adjustment values of each node based on the pressure guide values of each node in the water network, and send water pressure adjustment signals matching the water pressure adjustment values of each node, thereby realizing water network pressure balance control.

9. A system for real-time water network pressure balancing control according to claim 8, characterized in that, The system for realizing real-time water network pressure balance control is deployed on an edge computing device.

10. The system for real-time water network pressure balancing control of claim 8, wherein, The execution module (300) comprises a control unit configured to change the opening degree of a pressure reducing valve core based on the water pressure adjustment signals.

Images