Transfer learning in very-short-term solar forecasting: Bridging single site data to diverse geographical applications

Over the past decade, the rapid growth of solar energy penetration has posed significant challenges for grid balancing and scheduling, heightening the need for accurate and efficient short-term solar forecasting. While deep learning models have shown promise in improving forecasting accuracy, previous studies have often focused on data from specific sites, limiting their generalisability across different climatic and geographical conditions. This study addresses this limitation by employing a multimodal self-attention deep model, trained under the dry and clear climate conditions of Folsom, California, and integrating various transfer learning techniques. We examine the transferability of this model to a new dataset from Nottingham, UK, characterised by humid and rainy conditions. Specifically, we compare different transfer methods based on model architecture and validate performance with limited target site data (equivalent to two weeks of data). The model’s expertise can be effectively transferred, reducing the required data for successful model training by 80% (from four months to two weeks). Simulations under realistic scenarios demonstrate that the model, trained with just two weeks of data from the deployment site, achieved performance surpassing the baseline. This work demonstrates the feasibility of transferring deep learning models for solar forecasting across diverse climatic conditions, significantly reducing the data and time needed for model adaptation and deployment. This has the potential to enhance the reliability and efficiency of solar energy integration into power grids globally.