Adversarial Transfer Learning of Power Data Based on One-Dimensional Pyramid Pooling

In the field of power anomaly detection, both high-frequency and low-frequency collected data contain important information about abnormal behaviors in the power system, but their feature representations differ. To improve anomaly detection performance on low-frequency data, this study adopts a one-dimensional convolutional neural network (1D CNN) as a feature extractor to separately extract features from high-frequency (64k/s) and low-frequency power three-phase voltage and current (96/d) data. Subsequently, we employ adversarial transfer learning to fuse the features of high-frequency and low-frequency data through adversarial training, followed by pyramid pooling to achieve feature scale unification. The goal is to enhance the anomaly detection performance on low-frequency data using the expressive power of the fused features. Through experimental evaluations and comparative analysis, we validate the effectiveness and superiority of this method in power anomaly detection tasks. This research provides an innovative feature fusion approach for power system anomaly detection, offering a practical solution to enhance the reliability and security of power systems.