Research On Fault Location And Load Transfer Of Distribution Network Based On Machine Learning

With the rapid expansion of urban scale and the continuous advancement of power substitution,all walks of life rely more and more on power supply,and users have higher requirements for the reliability of distribution network.However,in recent years,the number of distribution network nodes has increased greatly,the structure is more complex,the actual fault situation is complex and changeable,the wiring branch and equipment are very complex,the topology of the distribution network is uncertain,and the switch of the load transfer path has a combined explosion problem.When a fault occurs,it is necessary to quickly locate the fault of the distribution network,reduce the affected area of the fault,after isolating the fault,timely load transfer is carried out in the non-fault power outage area.However,the traditional fault location and load transfer algorithm of the distribution network is difficult to solve this problem.Therefore,efficient distribution network fault location and load transfer methods are urgently needed to reduce fault losses,reduce operation and maintenance costs,and improve power supply reliability and user satisfaction.The emergence of advanced machine learning technologies such as graph convolution neural network and reinforcement learning provides new methods and ideas for solving distribution network problems.Firstly,this thesis reviews the background and current situation of distribution network fault location and load transfer technology,compares and analyzes the characteristics of various algorithms,and introduces the theory of distribution network fault location and load transfer as the basis for subsequent algorithm model.Based on the graph convolution neural network theory,this thesis proposes a new idea to solve the fault location problem of distribution network.The distribution network is treated as non-Euclidean spatial graph data.After the distribution network is abstracted as graph data,the feature information on the nodes and edges of the distribution network is used as the input of the graph convolutional neural network.At the same time,this thesis compares different types of characteristic data of distribution network nodes and edges,there are more choices according to different distribution network automation levels.For the two forms of data of distribution network nodes and edges,the input layer of merging edge features is designed on the nodes;Then,the hidden layer constructed by Graph SAGE is used to transfer node features to achieve feature abstraction from local to global.The output layer adopts the design of edge aggregation,transforming the fault location problem of distribution network into the edge classification problem of the graph.Finally,the experimental environment of random fault sample generation based on Open DSS software is designed,which proves that the proposed algorithm has high accuracy and robustness,and has good stability and convergence ability in the training process.After determining and isolating the fault,the power supply of users should be restored through load transfer in time.In the past,the load transfer problem is regarded as an optimization problem or a random search problem,and it is difficult to solve the contradiction between the solution speed and the quality of the transfer strategy.This thesis regards the transfer process as a Markov decision process,and proposes a load transfer control method for distribution network based on deep reinforcement learning.The agent adopts Dueling DQN algorithm considering environmental impact,which makes the learning goal clearer.In order to improve the accuracy and convergence ability of the algorithm,an improved algorithm of pre-simulation-greedy action strategy is proposed for the action strategy of the algorithm,and the proportion of action and learning is adjusted,and an adaptive optimization algorithm is adopted.Finally,a dynamic simulation distribution network environment model based on Open DSS software is designed to provide an interactive training environment for the agent.The results show that the proposed method can cope with the topology changes of distribution network under different faults,and give the optimal transfer strategy considering speed and solution quality.