Advanced data driven models based on machine learning for detection of faults and failures in solar based renewable energy systems.

The rapid expansion of renewable energy systems demands reliable fault detection and prediction to ensure operational efficiency and grid stability. This study presents a novel framework that integrates Extended Kalman Filter (EKF) state estimation with uncertainty-aware graph learning for photovoltaic (PV) array fault detection and localization. Raw sensor data are processed by the EKF to generate refined state estimates and uncertainty covariances for each PV module. These uncertainty measures dynamically modulate an attention-based graph construction module, enabling adaptive edge weighting that down-weights unreliable connections during noisy or transient conditions. The resulting dynamic graphs are analyzed by a temporal graph attention network to produce both node-level fault localization and global anomaly scores. The graph-construction, temporal-encoding, and prediction components were optimized jointly, while the EKF process and observation models and their noise covariances remained fixed after calibration. On the real-world dataset, it attains an AUC-ROC of 0.941 and F1-score of 0.918 for global detection, and a node-level F1-score of 0.865 with Exact Match Ratio of 0.738 for fault localization. The approach demonstrates strong robustness to sensor noise and transient faults by leveraging physical uncertainty to guide graph topology. This work offers a promising direction for reliable monitoring of large-scale PV systems and other sensor-rich energy infrastructures.