Hybrid CNN–LSTM Architecture for High-Accuracy Fault Detection and Anomaly Identification in Renewable-Rich Distribution Networks

The increasing penetration of renewable energy sources and inverter-based distributed generators has significantly altered the transient characteristics of distribution networks, leading to weaker fault currents, increased harmonic distortion, and more complex disturbance signatures. Traditional protection and monitoring schemes fail to provide adequate sensitivity under these operating conditions. This paper presents a deep-learning-based hybrid CNN–LSTM architecture designed for high-accuracy fault detection, classification, and anomaly identification in modern distribution grids. The CNN component extracts discriminative spatial features from fault-induced spectrograms, while the LSTM layer captures temporal evolution in waveform patterns. The proposed architecture was trained on 4,800 simulated transient events and validated using high-frequency (20 kHz) waveform data generated from an IEEE 33-bus distribution system with 30% PV penetration. The hybrid CNN–LSTM classifier achieved a testing accuracy of 98.12%, outperforming standalone CNN, LSTM, SVM, and wavelet-based classifiers. The Autoencoder achieved an AUC of 0.98 for anomaly detection. The results demonstrate that the proposed hybrid architecture provides robust performance under noise, high-impedance fault conditions, and renewable-induced distortions, offering a strong foundation for next-generation adaptive protection and monitoring systems.