Wang et al. (2026) Physics-guided neural network integrating uncertainty evolution for spatiotemporal wind speed prediction

Identification

• Journal: Communications Earth & Environment
• Year: 2026
• Date: 2026-07-28
• Authors: Junkai Wang, Yu Zhang, Yiming Ren, Qiong Li, Lianlei Lin, Hangyi Yu, Hanqing Zhao, Zongwei Zhang, Sheng Gao, Yupu Yang
• DOI: 10.1038/s43247-026-03844-x

Research Groups

• School of Electronics and Information Engineering, Harbin Institute of Technology, China
• University of Glasgow, UK
• School of Cyberspace Science, Harbin Institute of Technology, China

Short Summary

This study presents a physics-guided neural network called PD-WPN that improves the prediction of three-dimensional wind speed fields by learning both data patterns and physical consistency of wind evolution. The model combines information across time, space, and height, and uses physical constraints to guide the learning of wind speed, wind direction, and spatial structure.

Objective

• Develop a physics-guided neural network for accurate future three-dimensional wind field prediction from historical data.
• Integrate axial attention mechanisms with physical information units to improve spatiotemporal feature extraction and uncertainty evolution modeling.

Study Configuration

• Spatial Scale: Global (with focus on China)
• Temporal Scale: Short-term forecasting (up to 24 hours)

Methodology and Data

• Models used: Physics-Dynamics-guided Wind Prediction Network (PD-WPN), a physics-guided deep learning model that integrates axial attention mechanisms, physical information units, and uncertainty-weighted multi-constraint loss functions.
• Data sources: Historical wind field data from various locations in China.

Main Results

• PD-WPN outperforms classical spatiotemporal forecasting algorithms (CNN, ConvLSTM) and state-of-the-art models (E3DLSTM, PhyDNet, SimVP, TAU) in terms of RMSE, MAE, and ACC.
• The model achieves superior performance in predicting wind speed and direction at both 10m and 100m heights.

Contributions

• This study contributes to the development of a physics-guided neural network that can accurately predict three-dimensional wind fields.
• The integration of axial attention mechanisms with physical information units enables efficient multi-dimensional feature extraction and uncertainty evolution modeling.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 62176201).
• The authors also acknowledge support from the Harbin Institute of Technology's Technological Innovation Center of Littoral Test.

Citation

@article{Wang2026Physicsguided,
  author = {Wang, Junkai and Zhang, Yu and Ren, Yiming and Li, Qiong and Lin, Lianlei and Yu, Hangyi and Zhao, Hanqing and Zhang, Zongwei and Gao, Sheng and Yang, Yupu},
  title = {Physics-guided neural network integrating uncertainty evolution for spatiotemporal wind speed prediction},
  journal = {Communications Earth & Environment},
  year = {2026},
  doi = {10.1038/s43247-026-03844-x},
  url = {https://doi.org/10.1038/s43247-026-03844-x}
}