Shading fault diagnosis in rooftop PV systems under real-world operating conditions using a physics-informed hybrid LSTM-KAN approach

Partial shading in rooftop photovoltaic (PV) systems is a major source of energy-yield loss, electrical mismatch, and accelerated module degradation. Different shading mechanisms produce distinct electrical signatures, yet reliably separating dynamic partial shading (such as fast cloud transients) from fractional obstructions and uniform low-irradiance operation remains difficult due to overlapping current-voltage ( 𝐼 - 𝑉 ) characteristics. This paper proposes a physics-informed hybrid LSTM-KAN diagnostic framework. Nine physics-informed operational indicators, spanning electrical, thermal, efficiency, and shading families, are derived from synchronized inverter telemetry and meteorological measurements and segmented into non-overlapping 10-minute observation windows. Windows are labeled by majority occupancy at a threshold of 0 . 50 which a sweep over window length and threshold confirms to be the optimal setting for minimizing labeling errors. A residual bidirectional LSTM with attention pooling encodes temporal evolution, while a Kolmogorov-Arnold Network (KAN) trunk with cubic B-spline activations resolves multi-class shading topologies across eight distinct operating states. Evaluation is conducted using a strict 5-fold cross-validation scheme at day and independent-event granularity, with class balancing applied strictly inside training partitions. Pooling out-of-fold predictions across all 28,769 test windows yields an overall balanced accuracy of 89.6%, a pooled macro F1-score of 0.741 (which is inherently precision-limited by the 2.7% fault prevalence), and a pooled weighted F1-score of 0.988. The model detects faulted windows at 90.9% against a 0.81% false-alarm rate on healthy operation, corresponding to an overall detection ROC-AUC of 0.9785. Granular class evaluation demonstrates high sensitivities across static integer-module shading (81.3% to 91.7%) and fractional obstructions (up to 89.2%), while maintaining robust dynamic classification for intermittent cloud transients (pooled F1-score of 0.775). The proposed framework consistently outperforms nine classical machine learning and deep learning baselines trained on identical splits. Physics-guided feature design combined with hybrid temporal-KAN modeling therefore provides an effective, highly stable route for multi-class shading diagnosis under operational real-world conditions.