Exploring Transfer Learning Techniques for Solar Irradiation Forecast across Geographically Diverse Locations in Brazil Using Reanalysis Data

This research explores the effectiveness of Transfer Learning (TL) for day-ahead solar forecasting across diverse locations in Brazil. Hourly data from 2017 to 2020 of air temperature, ground-level and top-of-the-atmosphere solar irradiation, and cloud coverage were collected from the MERRA-2 Reanalysis database for ten sites within Rio de Janeiro and ten others distributed across the country. A Random Forest Regressor (RFR) was trained using data from Galeão Airport
(GIG) to forecast solar irradiation a day ahead, with evaluation over one year. This pre-trained model was then applied via TL, without fine-tuning, to all other sites. Results show that the model’s forecast accuracy performs very well, but declines with increasing divergence between source and target, emphasizing the importance of domain-specific modeling in solar forecasting. The findings highlight the need for robust models that consider distance thresholds, climatic variability, and geographical features to ensure reliable TL performance across different regions.