Wind and solar power are the fastest-growing generation technologies worldwide, but their integration into power grids requires understanding their inherent variability and predicting their availability across multiple spatial and temporal scales. While complex deep learning architectures have recently gained prominence in the literature, their computational cost can limit practical application. In this work, we propose a modelling framework based on Bayesian-optimised multilayer perceptrons for wind and solar power forecasting in Germany, formulated as a sequence-indifferent nonlinear regression problem with exogenous meteorological variables. Historical reanalysis data is used for training and evaluation, before operational applicability is assessed using numerical weather predictions from multiple models and comparing with both published studies and system operator forecasts. Point forecasts are complemented by adaptive prediction intervals, modelled with variance estimators based on histogram-based gradient boosting regression trees. The results show that the proposed MLPs achieve day-ahead accuracies (nRMSE of 3.7% for wind and 2.7% for solar) on par with or exceeding more complex state-of-the-art approaches. The resulting framework provides a computationally efficient and uncertainty-aware basis for decision-making in renewable energy integration. • Bayesian-optimised MLPs achieve state-of-the-art wind and solar power forecasts. • HistGBRT models forecast error variance from lead time and power magnitude. • Region-wide datasets combine ERA5, ICON, IFS, GFS, MaStR, and SMARD sources.
Milan Wanek (Sat,) studied this question.