Interpretable Physics-Guided Region-Aware Attention Learning for Series and Shunt Resistance Estimation from Photovoltaic I-V Characteristics

Abstract

Photovoltaic (PV) parameter estimation is essential for understanding device performance, degradation mechanisms, and energy conversion behavior. However, accurate extraction of parameters such as series resistance (R s ) and shunt resistance (R sh ) from current-voltage (I-V) characteristics remains challenging due to the nonlinear, noise-sensitive, and operating-condition-dependent behavior of PV systems. Conventional optimization-based techniques are often computationally expensive and sensitive to initialization and measurement noise, while existing deep learning approaches treat the I-V curve as a globally uniform signal without incorporating its underlying physical structure. In this work, a physics-guided interpretable region-aware attention framework is proposed for estimating R s and R sh directly from noisy PV I-V characteristics under varying irradiance and temperature conditions. The I-V curve is decomposed into low-voltage, transition, and high-voltage regions corresponding to distinct photovoltaic operating behaviors, enabling region-specific feature learning and cross-region interaction modeling through attention mechanisms. Environmental conditioning is additionally incorporated to improve robustness across operating conditions. Experimental results demonstrate highly accurate estimation performance, achieving an MAE of 0.0097 and an R 2 score of 0.9966 for R s , and an MAE of 45.86 with an R 2 score of 0.7662 for R sh . The proposed framework consistently outperforms conventional deep learning and optimization-based baselines. Furthermore, interpretability analyses reveal physically consistent behavior: R s estimation is primarily influenced by the high-voltage region, whereas R sh estimation is dominated by the low-voltage region, demonstrating the effectiveness of embedding physical structure into deep learning frameworks for accurate and interpretable PV diagnostics.