Forecast-space decomposition for intra-hour solar irradiance prediction from sky images
David Hamlyn · Sunny Chaudhary · Jonathon Hare · Tasmiat Rahman
University of Southampton
Energy and AI · 2026
Read the paper ↗ · Project page · Citation · Code release
GateFormer forecasts global horizontal irradiance (GHI) from sky imagery and recent irradiance measurements. A shared, GHI-conditioned Vision Transformer feeds two complementary predictors: one forecasts the future trajectory, while the other predicts change at the final horizon. A learned gate blends their forecasts, followed by a residual correction.
Conceptual overview. See the paper for the full architecture and conditioning inputs.
| Evaluation | Reported result |
|---|---|
| NREL-TSI · +30 minutes | 125.01 W/m² RMSE |
| Comparison with SIAT · paired timestamps | 9.6% lower RMSE |
| Forecast skill over smart persistence · three sites | 22.97–28.86% |
Evaluated on NREL-TSI, Chilbolton, and SIRTA. Figures are from the published paper; comparisons follow its evaluation protocol.
Code coming soon. This repository currently serves as the paper’s project page. The implementation will be added here in a future update.
Star this repository to bookmark the project, or use Watch → Custom → Releases to follow future releases.
If this work is useful to your research, please cite:
@article{hamlyn2026gateformer,
title = {{GateFormer}: Forecast-space decomposition for intra-hour solar irradiance prediction from sky images},
author = {Hamlyn, David and Chaudhary, Sunny and Hare, Jonathon and Rahman, Tasmiat},
journal = {Energy and AI},
volume = {26},
pages = {100917},
year = {2026},
doi = {10.1016/j.egyai.2026.100917},
url = {https://doi.org/10.1016/j.egyai.2026.100917}
}Download the BibTeX entry, or use GitHub’s Cite this repository button.
Vision Transformers for Estimating Irradiance Using Data-Scarce Sky Images — our earlier work on solar irradiance estimation, published in Energy and AI (2025).