Research on Power System Fault Diagnosis and Prediction Model Based on Deep Learning

Abstract

This paper presents a novel deep learning-based framework for power system fault diagnosis and prediction, addressing critical challenges in modern grids with high renewable energy integration. The proposed hybrid TCN-BiGRU-Attention model integrates temporal convolutional networks, bidirectional gated recurrent units, and attention mechanisms to achieve robust spatiotemporal feature extraction from multi-source heterogeneous data (SCADA, PMU, and relay signals). Key innovations include Gramian angular field transformation for unified electrical/non-electrical feature representation, Monte Carlo dropout-based uncertainty quantification, and a cooperative diagnosis-prediction optimization strategy. Evaluated on the IEEE 39-bus system and real-world grid data, the framework demonstrates superior performance: 98.7% classification accuracy under 20 dB noise, 42% reduction in false alarms compared to CNN-LSTM baselines, and sub-5 ms inference latency meeting real-time protection requirements. Case studies validate exceptional robustness in renewable-rich scenarios and graceful degradation under 30% data loss. The modular design enables seamless integration with energy management systems while providing interpretable fault indicators through SHAP value analysis. This work advances intelligent grid resilience by bridging data-driven learning with power system physics, offering practical solutions for evolving grid operational challenges.