Jiang et al. (2026) Physics-constrained multimodal vision transformer for ultra-short-term solar radiation forecasting error correction

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-04-05
• Authors: Ziyao Jiang, Wangminzi Peng
• DOI: 10.1038/s41598-026-46558-y

Research Groups

• Meteorological Sciences Institute of Jiangxi Province, Nanchang, 330096, Jiangxi, China
• Key Laboratory of Climate Change Risk and Meteorological Disaster Prevention of Jiangxi Province, Nanchang, 330096, Jiangxi, China

Short Summary

This study develops a physics-constrained multimodal vision transformer framework to correct ultra-short-term solar radiation forecasting errors, achieving significant reductions in RMSE and systematic bias by integrating satellite data with fundamental physical principles.

Objective

• To develop and validate an integrated error correction framework using physics-constrained multimodal vision transformers to refine ultra-short-term satellite-based solar radiation forecasts, addressing systematic biases and accumulating errors, particularly during transitional weather conditions.

Study Configuration

• Spatial Scale: 47 meteorological stations (ground measurements); Himawari-8 geostationary satellite coverage (for satellite data input).
• Temporal Scale: Four years of Himawari-8 satellite observations and ground measurements.

Methodology and Data

• Models used: Physics-constrained multimodal vision transformer, physics-constrained neural networks, hierarchical cross-modal attention mechanisms.
• Data sources:
  • Himawari-8 geostationary satellite datasets (visible, infrared, and water vapor channels) from the Japan Meteorological Agency.
  • Ground-based observational data from 47 meteorological stations (from the China Meteorological Administration).
  • Custom code developed for the study, including model architecture, training pipeline, evaluation scripts, and ablation experiment configurations.

Main Results

• The framework achieved an 18.7% Root Mean Square Error (RMSE) reduction compared to best-performing baselines.
• Systematic bias was significantly reduced from 12.7 W/m² to 1.2 W/m².
• Ablation studies confirmed synergistic interactions between multimodal fusion and physics-aware learning, with the combined approach outperforming individual components.
• The model maintains computational efficiency suitable for operational deployment, processing 21 forecasts per second on consumer-grade hardware.
• Radiative energy consistency was respected across all predictions.

Contributions

• Development of a novel integrated error correction framework combining multimodal vision transformers with physics-constrained neural networks for ultra-short-term solar radiation forecasting.
• Introduction of hierarchical cross-modal attention mechanisms to effectively extract complementary spatiotemporal features from diverse satellite channels (visible, infrared, water vapor).
• Integration of fundamental physical principles (energy conservation, radiative transfer) as soft regularization terms during neural network training, enhancing model robustness and physical consistency.
• Demonstrated substantial improvements in forecasting accuracy and bias reduction over existing operational systems and baselines.
• Validation of the synergistic benefits derived from combining multimodal data fusion with physics-aware learning.
• Achieved high computational efficiency, making the model suitable for real-time operational deployment.

Funding

• Project 1: Discriminant Analysis of the Impacts of Typhoons Entering Jiangxi Province on Local Precipitation over the Past 10 Years, Youth Talent Development Program of Jiangxi Provincial Meteorological Bureau (Project No. JX2023Q08).
• Project 2: Research on Jiangxi Local Bias Correction and Operational Application of CMA-WSP2.0 Solar Radiation Forecast Products, Special Project of Key Laboratory of Climate Change Risk and Meteorological Disaster Prevention of Jiangxi Province, General Program of Jiangxi Provincial Meteorological Bureau (Project No. JX2025M11).

Citation

@article{Jiang2026Physicsconstrained,
  author = {Jiang, Ziyao and Peng, Wangminzi},
  title = {Physics-constrained multimodal vision transformer for ultra-short-term solar radiation forecasting error correction},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-46558-y},
  url = {https://doi.org/10.1038/s41598-026-46558-y}
}