Research On Abnormal Electricity Consumption Detection Of CNN-LSTM Based On Attention Mechanism And Particle Swarm Optimization

The problem of non-technical losses in the power grid is a long-standing problem,among which power theft is particularly serious.Resident users have strong randomness in electricity consumption,and regression-based detection methods often result in poor detection performance due to limited prediction accuracy.In order to improve the performance of abnormal electricity consumption detection for residential users,this paper studies an abnormal electricity consumption detection model based on the hybrid network of convolutional neural network(CNN for the short in the following)and long short-term memory(LSTM for the short)network,combining particle swarm optimization(PSO for the short)and attention mechanism.The main research contents are listed as the following:Firstly,the behavior patterns of abnormal users’ electricity theft are studied.The abnormal electricity consumption behavior of residents has a certain regular pattern.According to the actual existence,the typical abnormal electricity stealing behavior patterns of the residents have summarized 6 kinds of electricity stealing patterns.Analyze these 6typical electricity stealing modes,and combine to get 4 compound electricity stealing modes.Combining the residential user power consumption data set and these 10 electricity theft modes,10 different abnormal user electricity consumption curves are simulated and generated,and the abnormal user electricity consumption data set is established.This data set provides data support in subsequent simulation examples to verify the model’s detection of different types of abnormal users.Secondly,a hybrid network model of CNN and LSTM network based on attention mechanism and PSO optimization is constructed.The detection model accurately predicts the electricity consumption of residential users,and judges the abnormal points based on the abnormal degree of the actual electricity consumption and the predicted electricity consumption.The detection model used in this article is mainly composed of PSO,attention mechanism,CNN and LSTM network.Among them,the LSTM network can learn the electricity consumption behavior of residential users,and has advantages in the prediction of time series data.This network is suitable for predicting the electricity consumption data of residential users,and uses it to construct an abnormal electricity consumption detection model for residential users.The CNN can filter the input features and fully mine the potential relationship of the data.Use this network to extract features from the input data and optimize the detection model.The attention mechanism can avoid the loss of sequence information,enhance the input features that play a key role in the model,and suppress the features that are not very relevant to the model results.The PSO can avoid improper manual tuning and optimize the structure and parameters of the model.The attention mechanism and PSO are used to optimize the abnormal power consumption detection model from different angles to improve the prediction accuracy and detection performance of the model.Thirdly,the simulation on the University of Massachusetts public data set is made to verify the detection performance and prediction performance of the model.According to the previously constructed hybrid network model of CNN and LSTM network based on the attention mechanism and PSO,the public data set is used to analyze the prediction accuracy and detection performance of the abnormal electricity use detection model.At the same time,the article uses a hybrid model of CNN and LSTM network,a long short term network model based on attention mechanism,a hybrid model of CNN and LSTM network based on attention mechanism,long short term network memory model based on attention mechanism and PSO,LSTM network model,gated recurrent unit(GRU for the short)model,support vector regression(SVR for the short)model,random forest(RF for the short)model and linear regression(LR for the short)model are nine models of the comparison model to verify the performance of the model used in the article.In this paper,a confusion matrix and a radar chart are used to represent the detection results of these 10 resident users’ abnormal electricity consumption detection models,which are used to intuitively show the detection performance of the detection model for the 10 electricity theft modes.In summary,the above method is feasible and effective in simulation practice.Compared with other models,the hybrid model of CNN and LSTM network optimized based on the attention mechanism and PSO has higher prediction accuracy,and both the positive rate and the false positive rate are improved.This model can provide certain help for reducing non-technical losses in the power grid and quickly troubleshooting abnormal users.It also has a certain reference value for subsequent abnormal electricity consumption detection.