| The accurate anomaly detection of electricity usage is conductive to discover the suspected users.The detection provides reference for inspection work on stealing electricity of electric power company,which improves the success rate of on-site inspection and reduces operating costs.Meanwhile,large amount of manpower and material r-esources can be cut down,which brings huge economic benefits.However,there are still some problems such as the trained data labels are difficult to obtain and the number of normal samples and abnormal samples is not balanced in the current research work.This paper carries out the research work on anomaly detection based on the characteristics of the electric power load dataset.The main works of this paper are as follows:Firstly,the data preprocessing method is studied for the electric power load dataset.According to its characteristics,the proj ect starts with missing value complement and outliers detection.The preprocessing extracts the features of the load sequence,such as statistical indicators,trend indicators and frequency domain indicators.Then,the electric power load dataset is sorted with abnormality using the Isolation Forest algorithm to filter normal samples with high reliability.Secondly,the under-sampling technique and semi-supervised learning method are studied.A hierarchical nearest neighbor under-sampling algorithm based on clustering is proposed to solve the problem of imbalance between classes.The algorithm extracts the most representative samples with the idea of stratified sampling to balance the number of the trained samples.As the electric power load dataset contains a small amount of labeled data and a large amount of unlabeled data,a Co-Forest anomaly detection model based on semi-supervised learning is established.Experiment results are conducted on both electricity load dataset and public datasets.It is shown that the proposed algorithm outperforms on the problem of imbalance between classes.Finally,an improved Co-Forest anomaly detection model is studied.Based on clustering,the single-class anomaly transforms into multiple-class anomaly to improve the accuracy of anomaly detection model.Then,a Co-Forest anomaly detection algorithm based on sample detection is proposed to solve the problem of semi-supervised learning insecurity.In each iteration of the model,the labeled samples are detected by calculating the anomaly membership degree to correct the label of samples and improve the security of the semi-supervised anomaly detection model.The experiment is completed using the electric power load dataset and the public datasets.And the results show that the Co-Forest anomaly detection model based on sample detection improves both the detection accuracy and model security. |
