Physics-aware Deep Neural Network for Bias Correction of 1–24-h
Precipitation Forecasts
Abstract
Accurate hourly precipitation forecasting is essential for local
socioeconomic activities. Here, a U-shaped neural network (U-Net) with a
customized loss function for precipitation is proposed to enhance
1–24-hour precipitation post-processing over Beijing, China. 36
precipitation-related predictors, informed by local forecasters’
expertise and grounded in physical constraints, are utilized as inputs.
Permutation importance (PI) and saliency maps are used to evaluate the
average and specific influence of selected predictors on U-Net
predictions. The findings indicate the inclusion of additional
predictors is crucial for bias correction, resulting in improvements of
~14.3\% in critical success index for
precipitation exceeding 10mm/h. PI analysis reveals that associated
atmospheric variables function as important supplementary indicators,
and saliency maps visualizations suggest that key region influencing a
given grid-point extends to ~30 kilometers for
1–24-hour precipitation forecasting. It demonstrates that integrating
knowledge-based predictors is promise in advancing precipitation
post-processing, and further understanding and interpretation of model
is crucial.