Boosting Power Anomaly Detection: Overcoming the Limitations of Small Input Windows

The current methods suffer from a limitation where the input window is small and independent, causing failure when the mains power remains relatively stable within this window. Expanding the input window size dramatically would lead to an undesirable surge in the number of model parameters. This paper proposes a novel approach to address this limitation by introducing a dynamic input window that adapts to the changing power dynamics. By leveraging a sliding window technique, the proposed method effectively captures both short-term and long-term power fluctuations, enabling accurate anomaly detection. Furthermore, the dynamic window size is optimized through a reinforcement learning framework, minimizing the number of model parameters while maintaining high detection performance. This approach offers a significant improvement over existing methods, paving the way for more robust and efficient power anomaly detection systems.