Bias correction of temperature and wind forecasts from the NOAA Global Forecast System (GFS) and Global Ensemble Forecast System (GEFS) using machine learning

Abstract

Numerical Weather Prediction (NWP) models, such as the National Oceanic and Atmospheric Administration's (NOAA) Global Forecast System (GFS) and the Global Ensemble Forecast System (GEFS), are essential tools for modern weather forecasting. NWP models are the backbones of various applications in weather, climate, and water enterprises. However, due to model limitations, initialization errors, and discretizations of grids, large systematic biases still exist aside from advances in computing capabilities, spatial resolution, and physical parameterization of the models. This study presents a machine learning-based bias correction framework that is driven two models: Extreme Gradient Boosting (XGBoost) and U-Net. The target variables include 2-meter temperature (2m-T), 10-meter and 100-meter wind speed (10m-WS and 100m-WS). ERA5 reanalysis data are used as the reference for evaluating and correcting forecast biases. The models are trained on both seasonal (summer and winter) and all-season datasets to account for seasonal variability in forecast errors. The GFS-based experiments focus on the CONUS region (24.5°N–49.5°N, 125.0°W–66.75°W), while the GEFS-based experiments cover Germany, a climatically diverse region. Results show that U-Net significantly outperforms XGBoost in long-term forecasting, particularly beyond 120 hours, due to its capacity to learn complex spatial and temporal dependencies. In contrast, XGBoost exhibits superior performance in short-term forecasts (0–48 hours), especially when data are limited, offering efficient and interpretable bias correction. Seasonal training improves temperature correction across both regions and models—especially during summer—while all-season models enhance generalization for wind speed forecasts. Quantitative evaluation using root mean square error (RMSE) confirms that both models effectively reduce systematic forecast biases in GFS and GEFS outputs. This work has also indicated that without using sophisticated deep learning structures, a rather simple machine learning model may achieve decent performance when correcting weather forecast products.