Research On The Identification Model And Method Of Electricity Theft In Distributed Photovoltaic Power Station

With the intensification of energy shortage and environmental problems,human beings have made great efforts to develop environmentally friendly renewable energy,especially solar energy,which has become the focus of attention and research in various countries.In recent years,in order to support and advocate distributed photovoltaic(PV)power generation projects,The National Energy Board announced the promotion of distributed PV development pilot in the whole county(city and district).Driven by interests,there are users who use certain technical means to make distributed PV on-line meters calculate more power generation,and then obtain high returns.This behavior is called PV electricity theft.This act of stealing electricity has a huge problem of power safety,and has caused huge economic losses to the country and society.Therefore,it is of great significance for the development of distributed PV power generation industry to carry out the corresponding electricity theft behavior identification research.Firstly,the basic principle of PV power generation and PV power generation system are briefly introduced,and the PV output characteristics under different light intensity and temperature conditions are analyzed.The typical distributed PV electricity theft methods and mathematical models are given,the characteristics of electricity theft are analyzed,and the algorithm principle of the distributed PV electricity theft behavior identification model used is introduced in this paper.It provides a basis for the subsequent research on distributed PV electricity theft behavior identification.Secondly,the accurate prediction value of distributed PV power generation is an important reference value for judging users’ electricity theft.Therefore,a PV power prediction model based on long short-term memory(LSTM)network is constructed.Compared with the experimental results of traditional BP neural network prediction model and RNN network prediction model,it shows that LSTM network prediction model can better deal with time series problems and reflect the dynamic characteristics of time correlation,with lower prediction error and higher accuracy.Thirdly,with the popularity of smart meters,the quality and quantity of power data of distributed PV are increasing,which provides conditions for the identification and detection of electricity theft based on data-driven.Therefore,the classification algorithm in traditional machine learning is studied.Based on the classification model of extreme gradient boosting(xgboost),a distributed PV electricity theft behavior classification method based on xgboost is proposed.The comparison of numerical simulation results shows that xgboost classifier has better recognition performance than KNN classifier and SVM classifier.Finally,the deep learning algorithm has better feature extraction ability and higher detection accuracy than the traditional machine learning algorithm.Therefore,convolutional neural network(CNN)is selected to identify the distributed PV electricity theft behavior.However,due to the imbalance of user sample type data collected by relevant power departments,it can’t meet the needs of electricity theft identification based on deep learning.Therefore,a distributed PV electricity theft sample data augmentation method based on Wasserstein generative adversarial network(WGAN)is proposed.According to the typical PV electricity theft model,CNN is constructed to identify the electricity theft behavior according to the data characteristics of electricity theft samples.Finally,through the example analysis and comparison of different data augmentation methods and classifiers,it shows that the electricity theft samples generated by WGAN can comply with the fluctuation law of real samples and the probability distribution characteristics of historical data,so as to effectively improve the recognition performance of various classifiers,among which the CNN classifier has a better recognition effect.