A Machine Learning-Based Bias Correction Method for Global Forecast System Products

Accurate numerical weather forecasting is essential for forecasters to make predictions. However, operational forecast models often exhibit biases, including non-linear systematic errors. In this study, we introduce ML4BC, a convolutional neural network based on the renowned U-Net architecture, designed to correct biases in GFS forecasts over the contiguous United States (CONUS). ML4BC incorporates the convolutional block attention module (CBAM), which applies attention maps to the input feature map for adaptive feature refinement. We applied ML4BC to bias-correct 2-m temperature and other crucial variables such as Convective Available Potential Energy (CAPE), Convective Inhibition (CIN), etc. Our results demonstrate that the model, when trained on a single forecast lead time (e.g., f072), can be effectively applied to all other forecast lead times. The model’s performance was evaluated using metrics such as root-mean-square error (RMSE) and mean error, showing promising improvements in forecast accuracy.