Spectrum Normalization-Based Adversarial Network and CNN for Electricity Theft Detection

Power theft is the main reason for the non-technical loss of the power grid. The massive user power consumption data provides a basis for the data-driven power theft detection. However, the imbalance of the data seriously affects the effect of power theft detection. Aiming at the data imbalance in power theft detection, an oversampling method and Convolutional Neural Network (CNN) power theft detection classification model are proposed. Through the confrontation training of SNGAN generator and discriminator, the generator is prompted to generate false samples similar to real power theft samples, judge the diversity, fidelity and accuracy of SNGAN generated samples by Inception Score and structural similarity index (SSIM), and use the generated samples with good validation effect to enhance the original power theft data. The CNN convolution classification model with the same network structure as SNGAN is trained on the enhanced data set, and conduct power theft detection on the test set. The results show that compared with the traditional sampling algorithm and GAN derived sampling algorithm, SNGAN oversampling method generates samples with more realistic value and has smaller false detection rate. Compared with the traditional classification model, the convolution classification model equipped with SNGAN sampling algorithm has higher detection rate and AUC value, which meet the requirements of power theft detection.