Study On Detection Method Of Electricity Theft Based On Smart Meters’ Data

Electricity supply is the basic guarantee of national economic construction and people’s normal life.For a long time,illegal electricity theft not only affects the planning of power supply,but also brings huge economic losses to the country.Traditional manual-based methods of electricity theft detection are inefficient and time-consuming.With the popularity of smart meters,massive amounts of power data are collected,providing a data base for more efficient AI detection algorithms.Based on smart meter data,the thesis analyzes the related statistics of electricity theft and builds intelligent electricity theft detection algorithms,and achieves the following outcomes:(1)Through the analysis of the various ways of electricity theft,the paper finds that both the ways of modifying meter circuits and using neutral line have missing data,and accordingly proposes a new characteristic statistic of electricity theft,the missing value information.Experiments based on real data show that the model using the missing value information has a 5% higher AUC and a 24% higher F1 score than the model using electricity usage information.(2)To address the problem that the data-based electricity theft detection algorithm is prone to misjudgment or omission due to dataset change,the paper proposes a electricity theft detection method based on neural architecture search.The algorithm responds to data changes by quickly and automatically constructing and updating the electricity theft detection model.Through comparison experiments with a variety of manually designed algorithms,it is found that the method not only can quickly and automatically construct the detection model,but also outperforms the optimal manual model(1.4%).(3)In response to the problem that the current feature extraction algorithm is poor and not highly optimized for extracting multi-category and high-dimensional electrical data,the paper proposes an electricity theft detection model based on Mask pre-training.The method uses Mask pre-training to build a neural network that has extracted feature vectors,and based on the extracted feature vectors,an integrated outlier point detection algorithm is built for electricity theft user detection.The effectiveness and superiority of Mask pre-trained neural networks for feature extraction of multi-category and high-dimensional data(higher than the next best 7.5%)are verified by comparison.The paper also compares the recall and precision of integrated outlier algorithm and five single outlier algorithms on electricity theft detection,and the results show that the integrated outlier algorithm has the highest recall and precision(higher than the next best 3.24%).