Research On Anomaly Detection Of Edge Grid Data Based On Deep Neural Network

In recent years,with the increasing popularity of power grid equipment intelligence,the amount of data generated by the terminal equipment has shown explosive growth.However,in the face of large-scale power time series data,power users’ behavior of stealing electricity,equipment failure and power grid fluctuation will cause abnormal data in the power time series data.In addition,the resources of terminal devices cannot meet the needs of intelligent algorithms.Therefore,how to effectively detecting the abnormal data hidden in the massive power time series data,and using intelligent algorithms to process the power time series data on the edge equipment nearby has become a hot spot in the current academic research.Facing the above problems,this thesis proposes the power time series data anomaly detection method based on the combination of time series directed horizontal visibility graph algorithm and graph convolutional neural network model,and the edge-oriented anomaly detection graph convolutional neural network model segmentation algorithm,and finally realizes the edge power time series data Anomaly detection.The main research work of this thesis includes:1.Aiming at the abnormal data implicit in power time series data in the power system,this thesis proposes a power time series data anomaly detection method based on the combination of time series directed horizontal visibility graph algorithm and graph convolutional neural network model.Firstly,the method uses the time series directed horizontal visibility graph algorithm to process the power time series data into graph data.Secondly,the graph convolution neural network model is constructed.Finally,the constructed graph convolutional neural network model is used to detect the processed graph data.The experimental results show that the graph convolutional neural network model has better classification performance and anomaly detection ability than other models,which provides a solid support for the realization of edge power time series data anomaly detection.2.Aiming at the problems such as idle terminal equipment resources on the edge of the power system and serious load on the cloud computing center,this thesis proposes an edge-oriented anomaly detection graph convolutional neural network model segmentation algorithm,and further designs a graph convolutional neural network model segmentation framework based on edge server and terminal equipment collaborative inference.Edge anomaly detection is achieved through two stages of offline training and online collaborative inference.In this thesis,the proposed graph convolutional neural network model segmentation algorithm for edge anomaly detection is validated,and the experimental results show that the graph convolutional neural network model segmentation algorithm for edge anomaly detection has better anomaly detection effect in edge power time series data.In this thesis,the graph datasets and power time series datasets are used to verify the proposed method.The feasibility and effectiveness of these two methods are proved by solving the problems of abnormal data implied by power time series data in power system and idle resources of terminal devices at the edge side of power system as well as serious load on cloud computing center.