Research on User Electricity Theft Detection Method Based on Recurrent Neural Network

This article introduces an automated technology for identifying electrical theft, utilizing data mining to combat theft during the electricity consumption cycle. This approach involves the utilization of long short-term memory (LSTM) units, which are an extension of the recurrent neural network (RNN) framework. These LSTM units manage the retention of input-output data through mechanisms like input gates, output gates, and forget gates, thereby mitigating the issue of vanishing gradients during the training phase. To address the challenge of information loss in RNNs when dealing with extended data sequences, the network has been enhanced to a parallel architecture by partitioning the input feature vectors from lengthy time series. Simulation tests were performed using datasets provided by the State Grid Corporation of China. Findings indicate that this enhanced method, within the same computational complexity, has achieved a detection accuracy of 92.75% for electrical theft, reduced the model's cross-entropy loss to 0.252, and raised the AUC to 0.872, outperforming standard RNN models in overall algorithmic efficiency.