An explainable learning architecture for dynamic fault localization in large scale photovoltaic arrays

Photovoltaic (PV) systems play a pivotal role in sustainable energy generation but remain highly vulnerable to operational faults such as arc faults, shading, and inverter errors, which compromise reliability and performance. Conventional fault diagnosis models often depend on handcrafted features, emulator-driven multimodal data, or threshold-based logic, lacking the scalability and transparency required for real-time, safety-critical deployments. Recent learning-based models such as CNN-BiGRU and residual CNN-GRU have demonstrated promising accuracy, yet fail to offer interpretable insights and require complex data pipelines. To overcome these limitations, this study proposes ECBL-Net (Explainable Convolutional Bidirectional Learning Network), a novel deep temporal model designed for accurate and interpretable fault localization using only raw time-series sensor data. The architecture combines convolutional encoding, bidirectional abstraction, and a temporal attention mechanism to localize faults precisely while maintaining transparency through SHAP-based post-hoc explanation. Implemented using Python 3.10 and PyTorch 2.1.0, ECBL-Net is trained on publicly available PVDAQ Systems 2105 and 2107 datasets, incorporating six key sensor modalities. Experimental results validate the model’s superiority, achieving 99.84% accuracy, 99.65% F1-score, and over 1.2% improvement in classification accuracy compared to state-of-the-art methods like CNN-LSTM and MSFWOA-MSIFN. ECBL-Net not only boosts detection accuracy but also offers class-specific fault localization with temporal clarity, maintaining inference latency suitable for real-time PV monitoring despite its added interpretability components. Its modular, attention-guided design invites further research in interpretable energy analytics and edge-deployable PV fault diagnosis. This framework offers a compelling direction for real-world PV monitoring, promising to reshape intelligent fault-aware maintenance with transparent, high-performance analytics a crucial leap toward trustworthy AI in solar energy systems.