Zhao et al. (2025) The potential of AI global weather models for reference evapotranspiration forecasting: a comparison with numerical weather prediction models

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-10
• Authors: Shuting Zhao, Yuan Qiu, Zihuang Yan, Lifeng Wu, Rangjian Qiu, Yufeng Luo
• DOI: 10.1016/j.jhydrol.2025.134363

Research Groups

• State Key Laboratory of Water Resources Engineering and Management, School of Water Resources and Hydropower Engineering, Wuhan University, Wuhan 430072, China
• Agricultural Water Conservancy Department, Changjiang River Scientific Research Institute, Wuhan 430010, China
• School of Hydraulic and Ecological Engineering, Nanchang Institute of Technology, Nanchang 330099, China

Short Summary

This study evaluates the potential of AI global weather models (GraphCast) for daily reference evapotranspiration (ET0) forecasting in mainland China, comparing its performance against numerical weather prediction models (ECMWF, JMA) and demonstrating significant improvements through XGBoost post-processing. The GraphCast-XGBoost model is recommended for 1–7 day lead times.

Objective

• To construct and evaluate daily ET0 forecasting models for 1–10 d lead times across 94 stations in mainland China using the FAO-56 Penman-Monteith equation based on outputs from the AI weather model GraphCast and two numerical weather prediction datasets (ECMWF, JMA).
• To assess the optimization potential for these forecast datasets using XGBoost post-processing.

Study Configuration

• Spatial Scale: 94 stations across mainland China, covering seven climatic zones.
• Temporal Scale: Daily forecasts for 1–10 d lead times, using data from 2020–2021.

Methodology and Data

• Models used:
  • Reference Evapotranspiration (ET0) calculation: FAO-56 Penman-Monteith (PM) equation.
  • Global Weather Models: GraphCast (AI-driven), ECMWF (Numerical Weather Prediction), JMA (Numerical Weather Prediction).
  • Post-processing model: XGBoost.
• Data sources: Outputs from GraphCast, ECMWF, and JMA global weather models.

Main Results

• The raw GraphCast-based PM model outperformed ECMWF and JMA-based models across 1–9 d lead times, achieving the highest forecast accuracy at 53.2–77.7 % of stations.
• XGBoost post-processing significantly enhanced the performance of all three forecasting datasets at 1–10 d lead times, reducing mean RMSE to 0.72–0.99 mm d⁻¹.
• The most substantial improvement from XGBoost was observed for ECMWF at 1 d lead time (31.8 % reduction in RMSE).
• The GraphCast-based XGBoost model exhibited superior global performance index rankings (79.8–96.8 % of stations) over the GraphCast-based PM model (67.7–77.7 %) at 1–7 d lead times.
• The dominance of GraphCast-XGBoost declined to 28.7–53.2 % of stations at 7–10 d lead times, with JMA surpassing GraphCast performance at 10 d lead time.
• The GraphCast-XGBoost model is strongly recommended for ET0 forecasting at 1–7 d lead times.

Contributions

• First study to assess the potential of AI global weather models (GraphCast) for reference evapotranspiration (ET0) forecasting.
• Unveiled a synergistic optimization pathway for ET0 forecasting by combining AI/NWP model outputs with XGBoost post-processing.
• Provides irrigation optimization insights for mainland China and similar regions.

Funding

• Not specified in the provided text.

Citation

@article{Zhao2025potential,
  author = {Zhao, Shuting and Qiu, Yuan and Yan, Zihuang and Wu, Lifeng and Qiu, Rangjian and Luo, Yufeng},
  title = {The potential of AI global weather models for reference evapotranspiration forecasting: a comparison with numerical weather prediction models},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134363},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134363}
}