Robust Deep Learning Models for Fault Detection in Grid-tied Solar Inverters: A Comparative Study of LSTM, Bi-LSTM, and CNN-LSTM Architectures

Grid-tied solar inverters, which are essential for renewable energy integration, represent the most failure-prone components in photovoltaic systems, with failure rates reaching 28% of total system failures. Detecting faults with high efficiency with a low false alarm rate is a key factor for uninterrupted power delivery in grid-tied renewable energy sources. Rapid fault detection can not only save components but also ensure grid stability. However, in real-world operating conditions, such as noise, variable loads, and non-linear dynamics, conventional detection approaches are challenged. Unlike existing simulation-based studies, there is a critical gap in hardware-validated, real-time fault detection approaches under realistic noise conditions. This study evaluates and compares three advanced deep learning architectures, LSTM, Bi-LSTM, and a hybrid 1D CNN-LSTM for multi-class fault classification in grid-connected solar inverters. The models were trained and tested using time-series datasets generated through OPAL-RT simulation with Raspberry Pi 4 under diverse fault scenarios with added sensor noise to simulate realistic environments of grid-tied solar inverters. Experimental results show that the CNN-LSTM model outperformed LSTM and Bi-LSTM by achieving 99.01% test accuracy while maintaining robustness for noisy input with ultra-low false alarm rate of 0.27%, surpassing LSTM (83.85%) and Bi-LSTM (98.52%). The Bi-LSTM provided superior temporal context learning, whereas the CNN-LSTM excelled in feature extraction from complex fault signatures. The results show CNN-LSTM as a strong candidate for real-time, noise-resilient fault detection in renewable energy systems, clearing the path for intelligent inverter monitoring solutions.