Towards an Accurate Assessment of Peru’S PV Potential: Comparing Physical, Empirical, and Ml-Based Models Across Diverse Climates
Abstract
This study compares three PV modelling approaches for predicting the direct-current (DC) power output (PDC) of heterojunction (HIT) systems across five Peruvian locations with diverse climates. Using multi-year, minuteresolution datasets, we evaluated the Araujo–Green (AG) empirical model, the Single Diode (SDM) physical model, and a linear regression (LR) machine learning model. Model assessment included plotting daily DC power output (PDC), which revealed systematic overestimation by deterministic models (AG, SDM) and a closer alignment of (LR) predictions with observed curves, with (LR) underestimating most of the time, except for Chachapoyas. Afterwards, model performance was assessed in two ways using RMSE, MAE, MBE, NRMSE, slope (β₁), and R², complemented by 95% confidence interval (95% CI) analysis of slopes on a monthly and annual basis to test significance against unity and across models. The empirical model (AG) consistently overestimated output (annual NRMSE = 5.79– 8.33%), while the physical model (SDM) reached “good” accuracy in most locations (e.g., NRMSE = 5.18% in Arequipa). The machine learning model (LR) achieved the best results for all locations in terms of NRMSE, reaching the “excellent” threshold (NRMSE < 5%) with values as low as 1.42% in Lima, though underestimating the most in Arequipa (MBE = – 59.84 W). Confidence interval analysis showed that (AG) and (SDM) were sometimes statistically indistinguishable at monthly scale, whereas annual results confirmed all models were statistically different, with (LR) achieving annual slopes closer to unity, except for Arequipa.
Keywords
PV modelling, multi-climate PV, physical models, empirical models, machine learning