Power distribution network fault section positioning method and device based on graph reinforcement learning
By applying graph-based reinforcement learning methods in the distribution network, using GNN and GRL models to identify and locate faults, the problem of insufficient diagnosis accuracy in the face of different fault types or unknown network structures is solved, and stronger generalization capabilities and efficient fault location are achieved.
Patent Information
- Application Number
- CN202411782147.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-05
- Publication Date
- 2025-05-06
AI Technical Summary
The prior art is difficult to maintain high diagnostic accuracy when facing different fault types or unknown network structures.
A graph reinforcement learning method is adopted to obtain real-time multi-source fault data from the distribution network, build a graph model, convert the data into graph data, use graph neural network (GNN) to identify topological features and fault information, and combine graph reinforcement learning (GRL) models to interact with the agent and the environment to achieve fault location.
When facing new fault types or unknown network structures, it shows stronger generalization ability, maintains high diagnostic accuracy, and improves the system's adaptability and flexibility.
Smart Images
Figure CN119936552A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of distribution network fault location, and in particular to a distribution network fault section location method and device based on graph reinforcement learning. Background Art
[0002] In the field of power distribution systems, the concepts of fault diagnosis and location play a key role in ensuring reliable and uninterrupted power supply to consumers. Distribution network fault diagnosis involves identifying and pinpointing the exact location of a fault or anomaly within the network infrastructure, such as a short circuit or equipment failure. Timely and accurate diagnosis is essential to minimize downtime, prevent cascading failures, and quickly restore power supply. Without effective fault diagnosis, utility companies will face operational inefficiencies, increased outage duration, and impaired service reliability, ultimately affecting relationships between consumers and businesses. Therefore, developing reliable fault diagnosis and location methods is essential to maintaining the integrity and resilience of the distribution network.
[0003] At present, distribution network fault diagnosis methods can be roughly divided into two categories: methods based on matrix analysis and methods based on intelligent optimization algorithms. Methods based on matrix analysis mainly rely on mathematical modeling of network topology, such as using adjacency matrix, node feature matrix, etc. to describe network connection relationships and node attributes, and identify fault locations by solving linear or nonlinear equations. Methods based on intelligent optimization algorithms use advanced technologies such as machine learning, neural networks, and reinforcement learning to learn patterns from a large amount of historical data to achieve more accurate fault location and classification. Although both methods have achieved certain results in specific scenarios, they also have some significant limitations: methods based on matrix analysis usually construct a static model that is difficult to adapt to dynamic changes in the network. Methods based on intelligent optimization algorithms have weak generalization capabilities for new types of faults or different network structures that have never been seen before, and cannot maintain high accuracy in some cases.
[0004] Therefore, it has become a top priority to develop a fault location scheme for complex distribution networks, ensuring stronger generalization capabilities when facing different fault types or unknown network structures while maintaining high diagnostic accuracy. Summary of the invention
[0005] In view of this, it is necessary to provide a distribution network fault section location method and device based on graph reinforcement learning to solve the shortcomings of the existing technology that it cannot maintain high diagnostic accuracy when facing different fault types or unknown network structures.
[0006] A distribution network fault diagnosis and location method based on graph reinforcement learning, comprising:
[0007] Obtain real-time multi-source fault data of the distribution network;
[0008] Based on the graph model of the distribution network, the real-time multi-source fault data is converted into graph data for the GNN model;
[0009] The topological features and fault information in the graph data are identified by using a GNN model to obtain the fault type;
[0010] The topological layer information of the GNN model is input into the GRL model, and the fault location is performed through the interaction between the intelligent agent of the GRL model and the environment.
[0011] Preferably, the real-time multi-source fault data is converted into graph data for the GNN model based on the graph model of the distribution network, including:
[0012] Based on the graph model of the distribution network, a network adjacency matrix is constructed according to the connection relationship between each node in the distribution network, and a node feature matrix is constructed according to the specific information of the nodes in the distribution network;
[0013] Based on the graph model of the distribution network, the network adjacency matrix and the node feature matrix are combined to obtain the graph data for the GNN model.
[0014] Preferably, the topological features and fault information in the graph data are identified by a GNN model to obtain the fault type, including:
[0015] The graph data is processed by a GNN model to obtain an output vector; the first element of the output vector is used to represent the fault type; wherein the total length of the output vector is m+1, and m is the number of nodes in the distribution network.
[0016] Preferably, the topological layer information of the GNN model is input into the GRL model, and fault location is performed through the interaction between the intelligent agent of the GRL model and the environment, including:
[0017] The topological layer information of the GNN model includes the fault type and the potential fault point location, and the potential fault point location is obtained according to the last m elements of the output vector;
[0018] The intelligent agent of the GRL model determines the environmental feedback after executing the preset action according to the topological layer information, and performs iterative calculation;
[0019] The GRL model determines the optimal decision path according to the result of the iterative calculation;
[0020] The GRL model performs fault location according to the optimal decision path.
[0021] Preferably, it includes: using Q-learning reinforcement learning algorithm to update the strategy of the intelligent agent.
[0022] Preferably, the agent of the GRL model determines the environmental feedback after executing the preset action according to the topological layer information, including:
[0023] The agent determines the current state according to the topological layer information and selects a preset action;
[0024] The agent performs the selected action and determines the environmental feedback after performing the preset action. Preferably, it also includes: optimizing the parameters of the GNN model by back propagation algorithm and stochastic gradient descent method; and using the cross entropy loss function to improve the fault diagnosis performance of the GNN model; the cross entropy loss function satisfies:
[0025]
[0026] Where N is the number of samples or batch size, M is the number of fault categories, and y ij is the true label and L is the predicted probability of the model.
[0027] Preferably, the network adjacency matrix is an m×m matrix, where m is the total number of nodes in the distribution network, and the element a in the network adjacency matrix is ij represents the connection relationship between the i-th node and the j-th node; if there is a direct connection between the node i and the node j, then a ij =1; otherwise a ij =0;
[0028] The node feature matrix is an m×(d+n) matrix, where m is the total number of nodes, d is the number of branches directly connected to each node, and n is the feature information of the node; if x ij (j≤d)x ij， If the value of (j≤d) is 0, it means that the corresponding node j directly connected to the node i sends a fault signal; if the value is 1, it means that no fault signal is sent; the characteristic information n of the node includes the current, voltage, phase angle and fault label information of the node.
[0029] Preferably, it includes: generating a fault repair plan through an intelligent agent of the GRL model.
[0030] A distribution network fault diagnosis and positioning device based on graph reinforcement learning, comprising:
[0031] A data acquisition module, used to acquire real-time multi-source fault data of the distribution network;
[0032] A data processing module, used for converting the real-time multi-source fault data into graph data for a GNN model based on a graph model of a distribution network;
[0033] A fault type identification module is used to identify the topological features and fault information in the graph data through a GNN model to obtain the fault type;
[0034] The fault location module is used to input the topological layer information of the GNN model into the GRL model, and perform fault location through the interaction between the intelligent agent of the GRL model and the environment.
[0035] Compared with the prior art, the beneficial effect of the present application lies in that the technical solution provided by the present application combines the distribution network fault diagnosis and location of graph neural network (GNN) and graph reinforcement learning (GRL), overcoming the shortcomings of traditional methods in adaptability, computational efficiency, data dependence and generalization ability. The technical solution provided by the present application shows stronger generalization ability, especially when facing new fault types or unknown network structures, while maintaining a high diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a flow chart of a distribution network fault diagnosis and location method based on graph reinforcement learning provided in an embodiment of the present application.
[0037] Figure 2 This is an example diagram of the power distribution network structure provided in an embodiment of the present application.
[0038] Figure 3 It is a schematic diagram of the GRL-based fault diagnosis and location process provided in an embodiment of the present application.
[0039] Figure 4 It is a schematic diagram of the GRL model structure provided in an embodiment of the present application.
[0040] Figure 5 It is a structural schematic diagram of a distribution network fault diagnosis and positioning device 500 based on graph reinforcement learning provided in an embodiment of the present application. DETAILED DESCRIPTION
[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0042] See also Figure 1 , Figure 1 This is a flow chart of a distribution network fault diagnosis and location method based on graph reinforcement learning provided in an embodiment of the present application. The distribution network fault diagnosis and location method based on graph reinforcement learning includes the following steps:
[0043] S101: Acquire real-time multi-source fault data of the distribution network.
[0044] Electrical parameters can be obtained through various sensors installed in the distribution network, such as current transformers, voltage transformers, temperature sensors, etc. The timestamps of events such as switch status and protection device triggering, as well as environmental factors such as temperature and humidity, are collected through condition monitoring systems, including feeder terminal units (FTUs) and substation automation systems (SASs). In addition, archived historical fault data can also be obtained, which can be used for model training and verification.
[0045] In order to achieve real-time collection and acquisition, real-time data transmission can be achieved through the Internet of Things (IoT) technology and communication protocols (such as MODBUS, DL / T645, etc.).
[0046] Furthermore, after obtaining the real-time multi-source fault data of the distribution network, data preprocessing can be performed, including noise removal, missing value filling, standardization and other operations to ensure data quality.
[0047] S102: Based on the graph model of the distribution network, the real-time multi-source fault data is converted into graph data for the GNN model.
[0048] Based on the graph model of the distribution network, a network adjacency matrix is constructed according to the connection relationship between each node in the distribution network, and a node feature matrix is constructed according to the specific information of the nodes in the distribution network; based on the graph model of the distribution network, the network adjacency matrix and the node feature matrix are combined to obtain the graph data for the GNN model.
[0049] Specifically, they include:
[0050] Constructing a graph model: Figure 2 The distribution network structure example shown in the figure; the distribution network presents a complex topology, which contains interconnected components such as transformers, switches and lines. Based on graph theory knowledge, the real distribution network model can be simplified into a network composed of nodes and branches, where nodes correspond to vertices in the graph data and branches correspond to edges in the graph data. In order to apply GRL to distribution network fault diagnosis and location, the first task is to construct a graphical representation of the distribution network using data from different distribution status monitoring systems. This involves two main components: network adjacency matrix construction and node feature matrix construction. As shown according to Figure 2 The distribution network structure is used to explain the construction method of the adjacency matrix D and the node feature matrix X.
[0051] The network adjacency matrix is an m×m matrix, where m is the total number of nodes in the distribution network, and the element a in the network adjacency matrix is ij represents the connection relationship between the i-th node and the j-th node; if there is a direct connection between the node i and the node j, then a ij =1; otherwise a ij=0.
[0052] Then we have:
[0053]
[0054] The node feature matrix X is an m×(d+n) matrix, where m is the total number of nodes, d is the number of branches directly connected to each node, and n is the feature information of the node; if x ij (j≤d)x ij， If the value of (j≤d) is 0, it means that the corresponding node j directly connected to node i sends a fault signal; if the value is 1, it means that no fault signal is sent; the characteristic information n of the node includes the current, voltage, phase angle and fault label information of the node. Then:
[0055]
[0056] Finally, the preprocessed data is mapped to the corresponding nodes in the graph model to form complete graph data.
[0057] S103: Identify the topological features and fault information in the graph data through the GNN model to obtain the fault type.
[0058] The graph data is input into the GNN model, and the topological features and fault information are extracted through convolution operations. GNN can capture complex dependencies and interactions, thus providing more accurate feature representation.
[0059] Furthermore, the topological features and fault information in the graph data are identified by the GNN model to obtain the fault type, including:
[0060] The graph data is processed by the GNN model to obtain an output vector Y; the first element of the output vector Y is used to represent the fault type; wherein the total length of the output vector is m+1, where m is the number of nodes in the distribution network. The final output of the GNN model is a vector Y of length m+1:
[0061] Y=[2,0,0,1,1,0,…,0,0]
[0062] The subsequent m elements correspond to the states of the nodes in the network. The eigenvalue of each node is a binary variable ranging from 0 to 1, where a value of 1 indicates that the network segment connected to the node is faulty. When the eigenvalues of two consecutive nodes are both 1, the faulty segment can be directly located.
[0063] Furthermore, in the GNN training process, the model parameters are optimized through back propagation and stochastic gradient descent, with the goal of enabling the model to accurately distinguish different types of fault conditions. The cross entropy loss function is used to measure the difference between the model prediction and the true label to improve the fault diagnosis performance of the model. The cross entropy loss function is defined as:
[0064]
[0065] Where N is the number of samples or batch size, M is the number of fault categories, and y ij is the true label and L is the predicted probability of the model.
[0066] S104: Input the topological layer information of the GNN model into the GRL model, and perform fault location through the interaction between the intelligent agent of the GRL model and the environment.
[0067] Please refer to Figure 3-Figure 4 , Figure 3 The schematic diagram of the fault diagnosis and location process based on GRL is shown in the figure. Figure 4 The topological layer information of the GNN model includes the fault type and the potential fault point location, and the potential fault point location is obtained according to the last m elements of the output vector Y; specifically, when the eigenvalues of two consecutive nodes are both 1, the fault point location can be directly located.
[0068] The intelligent agent of the GRL model determines the environmental feedback after executing a preset action according to the topological layer information, and performs iterative calculation; the GRL model determines the optimal decision path according to the result of the iterative calculation; the GRL model locates the fault according to the optimal decision path.
[0069] Among them, the agent of the GRL model uses Q-learning or other reinforcement learning algorithms to update the agent's strategy during the training process. The agent of the GRL model will select an action, such as opening or closing a switch. The action process mainly simulates the on and off of branches between nodes to modify the network topology, and determines the reward calculation by detecting the fault signals flowing through each node. Specifically, the reward value is calculated based on the environmental feedback after the action is executed (such as the change in the detected fault signal). The design of the reward function is crucial, which reflects the effect of the action: if the action helps to accurately locate the fault, the reward value is higher; if the action fails to improve the fault location or causes other problems, the reward value is lower.
[0070] The above uses Q-learning or other reinforcement learning algorithms to update the agent's strategy. The agent adjusts its behavior strategy based on the rewards received, with the goal of maximizing the cumulative rewards. The above process is repeated iterative training until the agent learns the optimal fault location strategy. In this process, after multiple iterations and optimizations, the agent can make the best decision based on the current network status and accurately locate the fault section. The fault location is accurately located with the least number of steps, that is, when the status of two consecutive nodes is 1, the fault section can be directly located; or the most likely fault location can be found based on the agent's optimal decision path.
[0071] Furthermore, after determining the section where the fault occurs, the section where the fault occurs can be marked in the graphical model of the distribution network to intuitively display the section where the fault occurs. The marking of the section where the fault occurs can include marking with line segments of different colors, where different colors represent different fault types. In addition, it can also be displayed in the form of annotation lines combined with annotation text.
[0072] Furthermore, the GRL model agent generates a fault repair plan. Based on the fault location results provided by the agent, a specific repair plan or suggestion is generated, such as dispatching maintenance personnel, adjusting load distribution, etc. Execute repair operations: Implement repair measures to ensure that the fault is handled in a timely manner and reduce the impact on power supply. Verify the effect: After the repair is completed, monitor the network status again to verify whether the fault has been completely resolved and evaluate the repair effect.
[0073] Therefore, by combining GNN with GRL, we can achieve:
[0074] Quick response: The optimal decision includes minimizing the time from fault detection to complete repair so that power supply service can be restored as soon as possible. Reduce maintenance costs: Accurate fault location can reduce unnecessary inspection and repair work, thereby reducing overall maintenance costs. Avoid cascading failures: The optimal decision should consider how to avoid new faults or the expansion of existing faults due to improper operation. Ensure safe operation: All decisions should be made under the premise of ensuring the safety of personnel and equipment, such as avoiding unnecessary operations in high-risk areas. In addition, the optimal decision is not only applicable to the current state, but should also be able to adapt to possible changes in the future, such as changes in network topology or the influence of other unknown factors. By continuously interacting with the environment, the intelligent agent accumulates experience and optimizes its decision-making strategy to cope with more complex and changing situations.
[0075] In summary, the embodiments provided in the present application can effectively locate faults and take appropriate measures to repair or mitigate the impact, while ensuring the stability and security of the system.
[0076] Based on the same inventive concept, the present application also provides a distribution network fault diagnosis and positioning device 500 based on graph reinforcement learning, such as Figure 5 The structural diagram of the distribution network fault diagnosis and positioning device 500 based on graph reinforcement learning shown in the figure includes:
[0077] The data acquisition module 501 is used to acquire real-time multi-source fault data of the distribution network;
[0078] A data processing module 502, configured to convert the real-time multi-source fault data into graph data for a GNN model based on a graph model of a distribution network;
[0079] A fault type identification module 503 is used to identify the topological features and fault information in the graph data through a GNN model to obtain a fault type;
[0080] The fault location module 504 is used to input the topological layer information of the GNN model into the GRL model, and perform fault location through the interaction between the intelligent agent of the GRL model and the environment.
[0081] The various modules and units of the above-mentioned distribution network fault diagnosis and positioning device 400 based on graph reinforcement learning specifically perform the above-mentioned steps S101-S104 and any optional embodiments thereof, which will not be repeated here.
[0082] In summary, the technical solution provided by this application introduces the GNN and GRL mechanisms, so that the model can dynamically adjust the strategy according to the current network status, enhance the adaptability and flexibility of the system, and is particularly suitable for handling situations where the network structure changes frequently. GNN can reduce the computational complexity while retaining the network topology information, improve the processing efficiency of large-scale complex networks, and ensure real-time response capabilities. Compared with traditional matrix analysis methods, GNN and GRL methods are more tolerant of a certain degree of data incompleteness or uncertainty, reducing the need for precise network parameters. Through continuous learning and optimization, the GRL model can show stronger generalization capabilities when facing new fault types or unknown network structures, while maintaining a high level of diagnostic accuracy. In addition, this solution not only achieves efficient fault type identification and fault location, but also can provide maintenance personnel with clear operational suggestions, enhancing the transparency and credibility of the decision-making process.
Claims
1. A distribution network fault diagnosis and location method based on graph reinforcement learning, characterized in that: include: Obtain real-time multi-source fault data of the distribution network; Based on the graph model of the distribution network, the real-time multi-source fault data is converted into graph data for the GNN model; The topological features and fault information in the graph data are identified through the GNN model to obtain the fault type; the topological layer information of the GNN model is input into the GRL model, and the fault is located through the interaction between the intelligent agent of the GRL model and the environment.
2. The method according to claim 1, characterized in that The real-time multi-source fault data is converted into graph data for the GNN model based on the graph model of the distribution network, including: Based on the graph model of the distribution network, a network adjacency matrix is constructed according to the connection relationship between each node in the distribution network, and a node feature matrix is constructed according to the specific information of the nodes in the distribution network; Based on the graph model of the distribution network, the network adjacency matrix and the node feature matrix are combined to obtain the graph data for the GNN model.
3. The method according to claim 2, characterized in that The topological features and fault information in the graph data are identified by the GNN model to obtain the fault type, including: The graph data is processed by a GNN model to obtain an output vector; the first element of the output vector is used to represent the fault type; wherein the total length of the output vector is m+1, and m is the number of nodes in the distribution network.
4. The method according to claim 3, characterized in that The topological layer information of the GNN model is input into the GRL model, and the fault location is performed through the interaction between the intelligent agent of the GRL model and the environment, including: The topological layer information of the GNN model includes the fault type and the potential fault point location, and the potential fault point location is obtained according to the last m elements of the output vector; The intelligent agent of the GRL model determines the environmental feedback after executing the preset action according to the topological layer information, and performs iterative calculation; The GRL model determines the optimal decision path according to the result of the iterative calculation; The GRL model performs fault location according to the optimal decision path.
5. The method according to claim 4, characterized in that include: Use the Q-learning reinforcement learning algorithm to update the agent's policy.
6. The method according to claim 4, characterized in that The intelligent agent of the GRL model determines the environmental feedback after executing the preset action according to the topological layer information, including: The agent determines the current state according to the topological layer information and selects a preset action; The agent performs a selected action and determines environmental feedback after performing the preset action.
7. The method according to claim 1, characterized in that Also includes: The parameters of the GNN model are optimized by back propagation algorithm and stochastic gradient descent method; and the cross entropy loss function is used to improve the fault diagnosis performance of the GNN model; the cross entropy loss function satisfies: Where N is the number of samples or batch size, M is the number of fault categories, and y ij is the true label and L is the predicted probability of the model.
8. The method according to claim 2, characterized in that: include: The network adjacency matrix is an m×m matrix, where m is the total number of nodes in the distribution network, and the element a in the network adjacency matrix is ij represents the connection relationship between the i-th node and the j-th node; if there is a direct connection between the node i and the node j, then a ij =1; otherwise a ij =0; The node feature matrix is an m×(d+n) matrix, where m is the total number of nodes, d is the number of branches directly connected to each node, and n is the feature information of the node; if x ij (j≤d)x ij , (j≤d) has a value of 0, which indicates that the corresponding node j directly connected to the node i sends a fault signal; if the value is 1, it indicates that no fault signal is sent; the characteristic information n of the node includes the current, voltage, phase angle and fault label information of the node.
9. The method according to any one of claims 1 to 8, characterized in that: include: Generate fault repair solutions through the intelligent agent of the GRL model.
10. A distribution network fault diagnosis and positioning device based on graph reinforcement learning, characterized in that: include: A data acquisition module, used to acquire real-time multi-source fault data of the distribution network; A data processing module, used for converting the real-time multi-source fault data into graph data for a GNN model based on a graph model of a distribution network; A fault type identification module is used to identify the topological features and fault information in the graph data through a GNN model to obtain the fault type; The fault location module is used to input the topological layer information of the GNN model into the GRL model, and perform fault location through the interaction between the intelligent agent of the GRL model and the environment.