The research on remote fault diagnosis system for substations based on deep learning

Substation failures can trigger large-scale power outages and cause substantial economic losses. However, the scarcity of fault data forces existing substation failure detection methods to rely on single-modality information, thereby limiting their perceptual capabilities, while aggregating data from multiple substations for model training raises significant privacy concerns. To address these challenges, we propose a remote substation fault diagnosis framework. Specifically, we introduce S-informer, a model that integrates heterogeneous data modalities through a specialized encoder and a cross-modal attention mechanism for effective feature fusion. A multi-scale temporal attention module is incorporated to enhance the model’s ability to capture temporal dynamics. To tackle the class imbalance problem in fault data, we design a dual-path decoder architecture that performs normal-mode reconstruction and fault-mode classification. In addition, we develop a similarity-aware federated aggregation strategy that leverages inter-substation similarity metrics and incorporates differential privacy mechanisms, enabling privacy-preserving model training across distributed sites. Experimental results on a real-world dataset demonstrate that the proposed method achieves an accuracy improvement of 2.79% over the multimodal Informer, while reducing training time by 39.5% compared to FedAvg.