A power network anomaly alarm denoising method based on a hybrid LSTM-attention model

Power grid fault alarms play a crucial role in minimizing system damage and service disruptions. However, existing deep learning approaches frequently neglect the topological structure and physical characteristics of power system data, leading to suboptimal fault identification performance. To address this limitation, this paper proposes PhysLSTM-Attn, a novel physics-informed deep learning method for power network anomaly alarm denoising. The model incorporates Kirchhoff’s Laws directly within the feature embedding layer, ensuring that learned representations adhere to fundamental circuit conservation principles. A topology-aware bidirectional LSTM encoder captures both temporal dependencies and spatial relationships by integrating graph-Laplacian-based positional encodings into its gating mechanism. In addition, an electrical-distance-enhanced multi-head attention mechanism computes attention weights based on electrical coupling strength rather than semantic similarity, providing a more accurate reflection of device interactions. A multi-hop graph convolutional network further models cascading fault propagation across multiple electrical distances, while a confidence calibration module supplies reliability estimates to support decision-making. Comprehensive experiments on the East China Power Grid Alarm Dataset and the IEEE 118-Node Extended Dataset demonstrate accuracy improvements of 8.32 % and 7.43 % over the LSTM-Attention baseline, respectively.