Self-Learning Fault Diagnosis Method for Intelligent Substations Based on Large Models and Multi-Source Data Fusion

This paper proposes a self-learning fault diagnosis method for intelligent substations based on large models and multi-source data fusion. It ensures data quality by collecting multi-source alarm information and implementing timestamp alignment and redundant alarm elimination. Precise correction of clock deviations between substation SCADA and IED devices is applied, and a hybrid similarity method is used to reduce duplicate alarms. A dynamic weight fault feature extraction model is established, using entropy values to assess the effectiveness of features and adjusting weights based on historical fault frequency. By utilizing feature vectors in combination with power grid topology and real-time status data, it achieves rapid and accurate fault point localization. Adaptive optimization of fault correlations is realized through knowledge graph technology, dynamically adjusting the association strength between alarms and faults based on actual operational data. This effectively reduces false positives and missed detection rates, enhancing the system's self-learning capability and adaptability.