Lightweight Multi-Modal Transformer for Photovoltaic Power Forecasting With Sky Imagery and Meteorological Data
IEEE Access · 2026 · 14 · 95289-95304
DOI: 10.1109/access.2026.3701814 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Peng Gao und 2 weiteren Autor:innen. Die Arbeit nennt 36 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
The increasing penetration of photovoltaic generation systems into modern power grids necessitates accurate short-term power forecasting to ensure reliable grid operations and optimal energy management. Traditional forecasting approaches relying solely on meteorological measurements often fail to capture rapid atmospheric changes that critically affect solar irradiance. This paper proposes a lightweight multi-modal transformer architecture that integrates sky imagery with meteorological time-series data through cross-attention fusion mechanisms for enhanced photovoltaic power forecasting. The framework employs parallel convolutional and transformer encoders to extract complementary spatial and temporal features from sky images and weather measurements, enabling the model to capture both gradual meteorological trends and abrupt cloud-induced irradiance fluctuations. Experimental validation on a real-world dataset, averaged over five independent random seeds, demonstrates that the proposed approach achieves a normalized mean absolute error of $0.0999\pm 0.0037$ and a weighted mean absolute percentage error of $11.81\pm 0.44\%$ , corresponding to a 16.4% improvement over the best unimodal baseline. The cross-attention mechanism enables weather-adaptive visual feature selection, with recurring high-response bands on the visual time axis indicating that the model concentrates on a limited subset of historical sky observations rather than weighting all visual tokens uniformly. The lightweight architecture with only $32\times 32$ pixel image resolution achieves superior performance compared to higher resolutions while maintaining computational efficiency suitable for operational deployment in photovoltaic forecasting systems.
Zitieren
APA
Gao, P., Li, Y., & Song, W. (2026). Lightweight Multi-Modal Transformer for Photovoltaic Power Forecasting With Sky Imagery and Meteorological Data. IEEE Access, 14, 95289-95304. https://doi.org/10.1109/access.2026.3701814
Vancouver
Gao P, Li Y, Song W. Lightweight Multi-Modal Transformer for Photovoltaic Power Forecasting With Sky Imagery and Meteorological Data. IEEE Access. 2026;14:95289-95304. doi:10.1109/access.2026.3701814
BibTeX
@article{Gao2026,
author = {Gao, Peng and Li, Ying and Song, Wenzheng},
title = {Lightweight Multi-Modal Transformer for Photovoltaic Power Forecasting With Sky Imagery and Meteorological Data},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {95289--95304},
doi = {10.1109/access.2026.3701814}
}
EUVIMED