Short-term forecasting of PV power based on aggregated machine learning and sky imagery approaches

Nowadays, the balance of the electricity grid is threatened by the increasing penetration of solar sources, whose inherent variability can cause a significant fluctuation of energy in the grid. Nowcasting of the output power of photovoltaic (PV) plants has become a central task for intelligent power grid management. This study aims to increase the performance of intra-hour PV power forecasts. For this, a new weighted ensemble model is proposed, that dynamically aggregates machine learning and models based on sky images. Incorporating physical information from direct sky observation into the algorithms is intended to reduce persistence, an intrinsic limitation of statistical forecasting models. The study is carried out with high-quality radiometric data and high-resolution sky images recorded on the Solar Platform of the West University of Timișoara, Romania. The overall results show that the proposed ensemble achieves a significantly higher skill score over persistence than the individual models, and a moderate precision, suggesting that a weighted ensemble penalizes large errors at the cost of average precision. Improvements in nowcasted PV power are notable at longer forecast horizons (longer than 15 min) and for periods of high-variability in the state-of-the-sky, scenarios where traditional approaches fail.