CONFERENCEAndrejs Cvečkovskis – “Machine Learning Approach to Solar and Wind Forecasting”
This work explores a machine-learning-based approach to short-term wind and solar forecasting using the Adaptive Fourier Neural Operator (AFNO) architecture. A complete data-processing and model-evaluation pipeline was developed using ERA5 reanalysis data, supporting multiple temporal resolutions (6-hour and 1-hour timesteps). Three models were trained and evaluated through error-growth analysis, long-term rollout assessment, and annual statistical comparison. The results show substantial differences in model stability and error accumulation, particularly for meteorological variables with strict physical constraints (e.g., snow depth, cloud fraction), where Fourier-space methods struggle to enforce hard constraints. Several post-processing techniques are presented that improve predictive accuracy without retraining. The work also shows a hybrid AFNO–PINN concept aimed at embedding physical laws directly into the learning process to enhance stability and physical consistency. Future directions include developing a regional high-resolution AFNO model for Latvia and applying the forecasts to solar and wind power-generation estimation.