Forecasting and Downscaling Solar Irradiance using Transformer Models

Accurate forecasting of solar irradiance is essential for integrating solar energy into the grid, particularly in light of the U.S. Federal Energy Regulatory Commission’s (FERC) Order 2222, which emphasizes the role of distributed energy resources such as rooftop solar in bulk power system operations. Existing forecasting methods often fail to capture the fine temporal variations in solar irradiance, limiting their effectiveness in informing local solar photovoltaic (PV) installations and optimizing renewable energy integration. This research presents a novel approach to solar irradiance downscaling and forecasting using deep learning Transformer models, which are designed to capture complex temporal dependencies and variability. The proposed method uses cubic spline interpolation to downscale solar irradiance data from a 15-min. resolution to a 5-min. local resolution. Subsequently, the Transformer model is applied to forecast the downscaled data. Different Transformer topologies were trained on historical data from Brookings, South Dakota (SD), and tested for a 24 hour forecast based on standard error metrics. The findings highlight the potential of Transformer models for improving solar irradiance forecasting while also shedding light on the challenges and benefits of using raw data for such tasks.