Operational Data Quality Outweighs Algorithm Choice in Machine-Learning Solar Power Prediction

Abstract

Machine-learning models for photovoltaic (PV) power prediction are usually compared on a single cleaned dataset, and the resulting rankings are read as properties of the algorithms. This study tests that reading on two grid-connected PV plants in India observed over the same 34 days (15 May to 17 June 2020), using 15-min inverter-level AC output and plant weather sensors from a public dataset. Plant-level targets were built from fixed inverter sets without imputation, and a data-quality audit flagged timestamps at which any inverter reported zero output while irradiance exceeded 0.1 kW m⁻². Two tasks were evaluated on the final 10 days after training on the first 24: nowcasting power from same-time sensor readings, and forecasting 15 to 120 min ahead from information available at issue time. Multiple linear regression, support vector regression, random forest, XGBoost, LightGBM and a multilayer perceptron were tuned by day-blocked forward-chaining cross-validation and compared with persistence and smart persistence using normalized RMSE (nRMSE), skill scores and Harvey–Leybourne–Newbold-corrected Diebold–Mariano tests. In nowcasting, the six learners reached 4.7–6.2% nRMSE on Plant 1 (R² ≥ 0.989) but 29.2–37.8% on Plant 2, where 35.0% of daylight timestamps carried the zero-output flag against 1.7% on Plant 1. Excluding flagged test timestamps lowered Plant 2 nowcasting error to 22.6–27.8% without closing the gap. At 1 h ahead the learners improved on persistence by 30.0–32.1% (Plant 1) and 21.3–27.3% (Plant 2); only one pairwise difference from the best learner was significant at the 5% level, and at 15 min the gain over persistence fell to 3.3–10.8%. SHAP attribution and feature ablation placed most nowcasting information in irradiance and most 1-h forecasting information in the diurnal profile. Models transferred between plants were biased by +24% and −29% in nowcasting. In this record, learner choice moved nRMSE by a few points while inverter operating condition moved it by more than 20. Reporting data-quality audits and persistence-referenced skill alongside model rankings would make such comparisons more transferable across plants.