Photovoltaic (PV) equivalent-circuit models are widely used for performance evaluation and diagnostics, but their usefulness relies on both accurate calibration and interpretable understanding of how parameters shape current–voltage (I–V) behavior. For nonlinear and strongly coupled PV models, conventional global sensitivity analysis can be computationally demanding and offer limited insight into effect direction and operating-point dependence. This study presents an method-oriented framework that integrates nature-inspired optimization with surrogate-based explainable global sensitivity analysis under a specified operating condition. The Starfish Optimization Algorithm (SFOA) is first used for parameter identification by searching for the optimal parameter set that minimizes the discrepancy between measured and model-predicted I–V data for the Single-Diode Model (SDM) and Double-Diode Model (DDM). A Random Forest (RF) surrogate is trained to approximate the mapping from voltage and parameters to output current. Its accuracy is evaluated on an independent test set, achieving RMSE/R2 of 0.001331/0.999366 for SDM and 0.003090/0.999394 for DDM. Sensitivity is quantified primarily using Shapley Additive Explanations (SHAP), with mean decrease in impurity (MDI) and one-factor-at-a-time (OFAT) analysis used for cross-validation. Under identical settings, SFOA achieves the best accuracy among competing optimizers, with best root-mean-square error (RMSE) values of 0.0008818 for SDM and 0.0008811 for DDM. The integrated SHAP, MDI, and OFAT analyses yield consistent importance structures, and the overall ranking for DDM follows the same trend as that for SDM. Within the examined ±5% neighborhood around the calibrated optimum, the photocurrent is the dominant factor governing I–V behavior, whereas diode-branch parameters show secondary and condition-dependent effects, and resistive parameters mainly contribute fine-scale adjustments within the examined neighborhood. Overall, the proposed framework provides accurate calibration and interpretable global sensitivity insights that can support a practical workflow for model-based PV analysis under the considered condition.
Huang et al. (Thu,) studied this question.