Machine Learning-Based Estimation of Photovoltaic Output Under Environmental Stressors in a Semi-Arid Climate

Accurate estimation of photovoltaic (PV) power output is crucial for improving the reliability and efficiency of solar energy systems, especially in regions exposed to high dust loads and temperature variability. This study is scoped as a concurrent power-estimation (soft-sensor) study, inferring instantaneous direct current (DC) power from synchronous environmental measurements rather than multi-step-ahead forecasting; extension to horizon-based forecasting is left for future work. A comparative modeling framework is presented to estimate the power output of a 280 W polycrystalline-silicon PV module in the semi-arid climate of Nakhchivan, using a sixmonth, high-resolution time series of synchronized environmental parameters, irradiance, ambient temperature, humidity, wind, and dust, collected via the Nakhchivan State Meteorological Service and a Benning PV-2 tester. Five supervised learning models were evaluated: multiple linear regression, Ridge, Lasso, Random Forest (RF), and Gradient Boosting, assessed using mean absolute error (MAE), root mean square error (RMSE), symmetric mean absolute percentage error (sMAPE), and Theil’s U1 (the inequality coefficient, bounded from zero to one; no random-walk baseline is claimed). Random Forest showed the lowest error among the models compared, with RMSE of 27.93 W and Theil’s U1 of 0.0896, though this evaluation used a random, non-blocked, train/test split; the magnitude of the RF advantage should be interpreted as an upper bound pending re-evaluation under a leakagefree, blocked time-series protocol. Plane-of-array irradiance, ambient temperature, and dust concentration emerged as the most significant predictors. Ensemble-based models show promising generalization within the sampled December to May window; performance under a leakage-free protocol and generalization to genuine forecasting horizons and to summer remain to be established. These findings can support solar power estimation, fault detection, and performance optimization in distributed energy systems, particularly in dust-prone or datasparse regions. Physics and Astronomy Classification Scheme (PACS) numbers: numbers: 88.40.−j [Solar energy], 88.40.fc [Modelling and analysis].