Research On Anomaly Detection Method Based On User Power Consumption

The research in the characteristics of user's data of daily power consumption helps the Power Grid Corporation to have a more thorough understanding of the user's electricity using behavior.According to the outlier features of the user's abnormal power data,if a system of anomaly detection for electricity using behavior based on machine learning can be established,it will be significant for the development of the power industry and the smart grid.Based on the theoretical research of data preprocessing and machine learning,this paper makes an analysis and studies on how to build an effective anomaly detection model to improve the performance of detection for user abnormal power data.The main contents of the work are as follows:First,it introduces the research meaning and application value of anomaly power detection,analyzes the research status and application situation of machine learning and anomaly detection methods at home and abroad,which provides the basis for the research of anomaly power detection.According to the high-dimensional and timing features of user daily power data,the related knowledge and basic theory of three types of data preprocessing methods are introduced,based on statistics,automatic encoder and principal component analysis.Two anomaly detection algorithms based on BP neural network and isolation forest are introduced,with principle interpretation and algorithm implementation.Then,a model of anomaly detection for users' power behavior based on BP neural network is proposed and established.This model combines statistics and PCA data preprocessing methods,which effectively extracts the features of the electrical data,reduces the data scale,and improves the detection precision of the anomaly detection model.Meanwhile,this model adopts an improved algorithm of double hidden layer to train electrical data,which further improves the training effect of the model.On the other hand,considering the data preprocessing method of deep auto encoder and Kernel PCA,a model of anomaly detection for users' power behavior based on isolation forest is proposed and established.Through these two data preprocessing methods,the effect of isolation forest model in the scene of anomaly power behavior detection is improved.Finally,a comparative validation experiment is designed for the two anomaly detection algorithm models above.In the view of data preprocessing and anomaly detection model,the experiment proves the effectiveness of the data preprocessing results,which improves the performance of the model detection.The experimental results of two anomaly detection models for power using behavior show that,the isolation forest anomaly detection model based on Kernel PCA method is more stable and excellent than the BP neural network model based on statistics and PCA method.