Design And Implementation Of Electricity Anomaly Detection Based On Federated Learning

The abnormal power consumption behavior of power users has seriously disrupted the normal order of power consumption and endangered the safe operation of the power grid.Nowadays,artificial intelligence has emerged in the field of electricity anomaly detection,gradually replacing traditional manual inspection methods with the rapid development of machine learning.However,due to the slow popularization of intelligent power equipment,some power companies have poor historical data due to various reasons such as backward equipment conditions,false alarms and omissions of terminal alarms,and judgement errors by inspectors,resulting in poor performance of the trained models.In order to combine the data of multiple power companies under the premise of protecting data security,and solve the problems of poor data quality and insufficient effective data of power companies,this project has designed and implemented a set of electricity abnormal detection system based on federated learning to assist power company manages the standardized electricity consumption of users and improves the efficiency of electricity inspection.Different from the traditional method of training models in centralized data,the joint electricity anomaly detection system designed and implemented in this project adopts a distributed model training method,using local data distributed in various power companies to train the electricity anomaly detection model.Aggregate model parameters uploaded by power company clients participating in federated learning to update and share the global model.At the same time,in order to solve the problem that the global model of federated learning is easily affected by the client model with poor data quality,this project proposes an automated anomaly detection method based on client model parameters,which can detect the poor quality of data participating in federated learning.Based on visual analysis technology,it overcomes the "black box" characteristics of federated learning,further analyzes the client’s behavior and makes corresponding decisions,and comprehensively improves the performance of the global model.At the end of this project,a large-scale historical power consumption load data of power users is used to simulate and contrast the system.The experimental results show that the average test AUC of the electricity anomaly detection system reaches 97%,which is far higher than the traditional electricity anomaly detection method,which proves the practicability and feasibility of the system.