An Empirical Study on Improving Operational PV Power Forecasting Accuracy for a Single Solar Power Plant

Abstract

This study verified whether the quality of input forecast data and the design of information availability by prediction horizon have a greater impact on prediction performance than the complexity of the model structure in a single power plant environment. To this end, an operational power generation forecast pipeline was constructed for a single 120 kW solar power plant by integrating actual inverter data, meteorological forecast data, astronomical information, and multiple numerical weather forecast data. Observed power output was converted into a clear-sky index, normalized based on the power generation expected under clear conditions, and set as the prediction target. Experimental results showed that the proposed pipeline demonstrated performance with an MAE of 5.87 kW in the 0–3 hour short-term forecasting interval and an MAE of 8.01 kW in the 3–24 hour leading forecasting interval. Furthermore, it showed improvements of RMSE skill of 0.11 and 0.47, respectively, compared to the baseline for each horizon. The results of this study demonstrate that in a single power plant environment where multiple numerical weather forecast data are available, the quality of input forecast data, the consistency between training and inference data, and the design of information availability per prediction horizon may play a more significant role in prediction accuracy than the complexity of the model structure.