Swin transformer-based transferable PV forecasting for new PV sites with insufficient PV generation data
Shijie Xu, Hui Ma, Chandima Ekanayake, Yi Cui
Abstract
To effectively operate solar farms, accurate photovoltaic (PV) generation forecasting is required. For a newly constructed solar farm (PV site), its generation data could be limited. Using the sky images obtained from ground-based whole-sky cameras, this paper proposes a transferable Double route Shifted window Cross-Attention Transformer (DSCAT) framework to provide PV forecasting of the newly constructed PV site. The framework is trained using the data of an established PV site and then provides ultra-short-term PV forecasting for the newly constructed PV site. In the proposed framework, a temporal difference parallel Shifted window (Swin) Transformer-based structure is designed to capture the cloud motion details and extract the static spatial features. Then, a cross-attention structure is utilized to analyze the temporal features and predict the future PV generation. A variety of transfer strategies are designed to transfer the trained model to provide the PV forecasting at the new PV site. The training and transfer experiments are conducted with real-world sky images and PV generation datasets. The result shows the proposed framework could be transferred between varied environments, and provide a reliable forecast which achieves a 49% enhancement over the persistence baseline and 13% improvement over the PV forecasting benchmarks on average. • A transferable ultra-short-term photovoltaic (PV) forecasting model was proposed. • A double route cross-attention structure utilized for sky image sequences. • Detailed transfer learning strategies are designed and evaluated. • Experiments for PV forecasting and transferring between two real-world datasets.