A deep learning framework for photovoltaic power forecasting in multiple interconnected countries

Accurate photovoltaic (PV) power forecasting is essential for enabling utilities, grid operators, and market participants to anticipate solar generation patterns and adjust operational strategies. Recently, Deep Learning (DL) has emerged as a promising approach for improving PV forecasts. However, training DL models typically require large datasets, which are often limited in PV power forecasting, restricting their full potential. Incorporating additional data from multiple heterogeneous sources, such as different countries, can help address this limitation but – at the same time − introduces challenges related to distribution shifts that can negatively impact forecasting accuracy. The rapid expansion of PV capacity further complicates forecasting, as generation patterns shift with new installations. Existing approaches often face challenges in effectively utilizing data from multiple sources. The main novelty of this paper lies in proposing a DL-based PV forecasting framework that integrates both foundational and country-specific training. This approach enables the framework to leverage data from multiple interconnected countries, improving generalization and enhancing forecasting accuracy. Additionally, we introduce an adaptation mechanism that dynamically recalibrates the forecasting model in response to changes in PV capacity, without the need for retraining, allowing our framework to seamlessly handle such variations. The proposed framework, evaluated on data from three interconnected countries (Greece, Bulgaria, and Romania), demonstrates significant improvements in accuracy by exploiting cross-country data and adapting to evolving conditions. Using data from these three interconnected countries is particularly important since each country’s domestic PV production can directly affect the clearing prices and the energy mix of the neighboring countries. This work highlights a critical yet underexplored research direction, demonstrating the potential for significantly more accurate DL-based PV forecasting by effectively utilizing big data, in line with recent advances in DL across various fields.