Research On Detection Method Of Non-Technical Loss Based On Electricity Consumption Curve And Deep Learning

The existence of non-technical loss in power grid not only has a significant negative impact on economic benefits of power companies,but also poses serious practical threats to power quality as well as operational safety of the power system.With the recent vigorous popularization of smart meters in power gird,it is possible to collect massive electric power information instantly.However,it also lays hidden dangers for energy fraud of measurement data.Some malicious users conduct various tampering behaviors in a contactless way on the information side,which makes traditional detection methods of non-technical loss gradually fall into limitation.This paper mainly studies detection and classification of non-technical loss in power grid,which is based on user’s electricity consumption curve and deep learning.The details are as follows:1.In view of the situation that malicious users take new types of interventions based on electricity metering data,a number of tampering strategies are formed to generate consumption information involving non-technical loss based on the original electricity consumption curves.Through analysis and research on the expected target and operation concept of implementing tampering behaviors,on the basis of a collection source with universality and representativity,false electricity consumption curves are computed based on various tampering strategies,providing basic content support for the follow-up research.2.In order to fully exploit features implicit in electricity consumption curves,a combination of feature sequences that can realize quantitative characterization of consumption behavior is proposed.After extracting typical electricity consumption curves that reflect overall consumption situation of power users during global statistical period,a quantitative description method for consumption behavior such as trends,scales and anomalies contained in metering samples is presented.Considering the time series structure of electricity consumption data,bidirectional long short-term memory network is utilized,which can retain long-distance input for a long time and can realize bidirectional transmission of information state,providing a powerful convenience for seeking correlations between feature sequences and specific tampering strategies.3.Considering the application of cross-entropy loss to transform the problem into multiclass classification of electricity samples,a non-technical loss detection system based on deep neural network architecture is constructed.By comparing predicted output results of network with their actual attribute labels,confusion matrix and evaluation metrics derived from it are utilized to describe classification performance of the model.Compared with other classifiers,the superior performance of the model in terms of prediction results and evaluation metrics is proved through example simulation.Generalization and applicability of the model are also verified,realizing effective detection and accurate classification of non-technical loss in power grid.