A hybrid deep learning and circuit-theoretic framework for module-level fault diagnosis and localization in pv arrays

The rapid global expansion of photovoltaic (PV) systems necessitates robust fault management for minimizing operational downtime and enhancing system reliability under fluctuating environmental conditions. Current diagnostic methodologies, however, often lack the necessary robustness and generalizability to data-driven scenarios or to differentiate sub-string faults under dynamic partial shading conditions (PSCs) and realistic irradiance gradients. To overcome these limitations, this paper proposes a hybrid diagnostic framework that is evaluated under non-ideal operating conditions, specifically accounting for realistic cable and busbar resistances. The proposed pipeline decouples discrete topological classification from continuous physics-based estimation. In the first stage, a hybrid CNN-BiLSTM-Attention architecture classifies the topology, resolving ambiguities in electrical profiles at critical healthy-to-PSC boundaries with 99.76% classification accuracy. Subsequently, the identified topology is fed into a deterministic circuit-analysis framework to perform module-level fault localization. The proposed non-iterative mathematical core directly maps terminal metrics ( \(\:{V}ᵢᶜᵃˡ\) , V 1PVi , I i and I Endi ) to physical fault location, achieving a discrimination accuracy of 100% for variable-resistance line-to-ground (LG) faults and 100% for zero-resistance line-to-line (ILL) faults within solid intra-string bypass configurations. Furthermore, due to the analytical nature of the localization equations, the proposed technique remains invariant to fluctuations in irradiance and temperature, provided these environmental parameters are uniform across the PV array. Simulation results on a 3 × 5 PV array confirm that the proposed approach sustains high diagnostic accuracy even under measurement noise and parameter uncertainty, demonstrating its promise as a robust diagnostic method for PV arrays.