Transfer Learning for Photovoltaic Power Forecasting Across Regions Using Large-Scale Datasets

Accurate photovoltaic (PV) power forecasting enables stable grid operation; however, acquiring sufficient and diverse data remains challenging. Although numerous deep learning models have been employed, most rely on proprietary datasets or data spanning only a few years, limiting the generalizability of forecasting applications. This study proposes a cross-continental transfer learning framework for PV power forecasting using a large-scale dataset with 4.5 million data points, including a source domain dataset with over 3.5 million data points. The prediction model, which is a transformer-based approach, compares the results of zero-shot learning, linear probing, fine-tuning, and training on the target dataset alone. The target datasets are located on different continents with different meteorological characteristics. Results demonstrate that the proposed transfer learning approach significantly improves prediction accuracy. It reduces the mean absolute percentage error by up to 38.8% compared with models trained solely on the target data. Further analysis reveals that freezing a few transformer blocks is beneficial when the target dataset is sufficiently large, whereas freezing most layers is effective for smaller datasets. This study analyzes the impact of source and target datasets on prediction performance, demonstrating that larger datasets in both domains enhance forecasting accuracy. These findings offer a robust approach for cross-continental knowledge transfer in renewable energy forecasting, particularly benefiting regions with limited historical data or newly installed systems. The source code is available on https://github.com/gist-ailab/transfer-learning-PV-forecasting.