In response to the growing need for accurate forecasting of electricity generation from PV installations, which is crucial both for enhancing self-consumption and for balancing the power grid, this study presents a comparative analysis of selected machine learning models. The research focuses on the XGBoost algorithm and LSTM neural networks, applied to predict PV energy production based on meteorological data and historical generation records from four medium-sized PV installations (30–50 kWp) located in Poland. Meteorological data were retrieved from open sources and combined with actual production measurements to build the training dataset. This paper discusses the challenges posed by these data at the given latitude, as well as issues related to processing data from newly launched installations. The performance of both approaches was evaluated in short- and medium-term forecasting, with particular attention to prediction accuracy, robustness to noisy data, and the ability to capture nonlinear relationships.
Kraska et al. (Wed,) studied this question.