Anomaly detection of power load data based on Transformer and generative adversarial networks

Power system planning, load forecasting, and energy utilization analysis are significantly impacted by power load anomalies, necessitating prompt detection and identification. Firstly, the abnormal classification, causes, and characteristics of power load data are analyzed. Secondly, it enhances the traditional Transformer encoder structure by replacing the mask multi-header attention layer with the multi-header attention layer and eliminating the feedforward network. These improvements aim to enhance the model's global attention to the load sequence. To capture trend characteristics more effectively and expedite convergence, an improved generative adversarial networks (GAN)-Transformer model is proposed based on the generator and discriminator game structure of traditional GAN. Additionally, a multi-stage mapping and training approach along with an integrated focus score scoring mechanism are introduced. These techniques facilitate phased load sequence reconstruction, enabling the model to better extract anomalous features of the load data. Finally, through an analysis of an arithmetic example, the results demonstrate that the GAN-Transformer model outperforms in terms of load data anomaly detection accuracy, recall, F1 value, and training time. These results validate the effectiveness and superiority of the proposed method. The findings of this research provide valuable insights for advancements in power load data anomaly classification and data repair based on deep learning.