Merging machine learning sky imaging methods with the two-state model in photovoltaic power nowcasting

Abstract Accurate photovoltaic (PV) power forecasting is vital for grid integration of PV plants. This study proposes a hybrid model for PV power nowcasting, combining a two-state empirical model with machine learning techniques. The two-state model forecasts PV power based on sunlight availability, indicated by the sunshine number (SSN). SSN is predicted using principal component analysis and decision tree ensembles, with the input derived from sky images. A simple neural network aggregates the available data with the SSN forecast to provide a PV power forecast. The model construction and operation are illustrated with data collected during the spring season. The data used to conduct this study was recorded at the Solar Platform at the West University of Timisoara, Romania. A training dataset from 2022 to 2023 and a test dataset from 2024 are used for illustrating forecasts at 15 and 30 minutes intervals. The results show skill score improvement between 10.7% and 18.5%.