Optimized data-driven strategy for fault region identification, type classification, and location prediction in a multi-source wide area power system under practical conditions

Fault identification in large-scale multi-machine power systems remains a critical challenge for maintaining grid stability and reliability. Reported methods that assume idealized solid fault conditions lack applicability to real-world scenarios where fault resistance significantly affects system behavior. This paper presents a comprehensive two-phase deep learning-based framework that addresses both theoretical solid fault and practical non-solid fault scenarios. Phase I develops enhanced Long Short-Term Memory (LSTM) architectures that significantly improve the findings reported in recent literature. Moreover, Phase II introduces the fault modeling in a large-scale multi-machine power system, considering a wide fault resistance range of 0.1–50 Ω. The model suggests bidirectional LSTM (BiLSTM) architectures with optimized hyperparameters. Extensive validation on two-area 11-bus benchmark system demonstrates that the proposed framework successfully detects, classifies and locates ideal and practical power system faults. The proposed data pre-processing methodology in addition to optimizing the LSTM design evidently heightens the fault identification accuracy in wide-area power systems.