Anomaly detection and disaster response model for real-time power data based on LSTM and attention mechanism

Abstract

Anomaly detection and disaster response based on real-time power grid data are essential for ensuring the stability and robustness of power grids. This paper proposes an integrated framework combining multi-scale 1D-CNN for local feature extraction, LSTM for temporal dependency modeling, and dual-branch temporal-channel attention for adaptive feature weighting, coupled with a D3QN-based disaster response module for low-latency sequential decision-making. Experimental validation on a real-time power monitoring dataset demonstrates an anomaly detection accuracy of 95.2 %, an F1 score of 93.8 %, a disaster recovery rate of 94.6 %, and an overall system response latency of 112.6 ms, outperforming CNN-BiLSTM-AdaBoost and four other baselines. Future work will focus on reducing computational overhead and improving model robustness under large-scale deployment conditions.