An Experimental Deep Learning-based Postprocessing Scheme for Producing Probabilistic Quantitative Precipitation Forecasts

Recent advances in Machine Learning offer many opportunities to improve skill in the forecasts of heavy precipitation. In a prior effort, researchers at UT Arlington developed a hybrid ML-parametric postprocessing scheme, namely the Artificial Neural Network – Censored Shifted Gamma Distribution (ANN-CSGD), that produces probabilistic quantitative precipitation forecasts (PQPFs) from numerical weather prediction ensemble precipitation forecasts. Building on this success, we developed an enhanced version of this scheme by adopting a Graphical Neural Network (GNN) to better account for the spatial patterns of precipitation systems. We performed a set of hindcast experiments to compare the performance of the new version, i.e., GNN-CSGD, with the baseline version ANN-CSGD.

The experiments show that the GNN-CSGD outperforms ANN-CSGD in terms of overall forecast skill measured by the continuous ranked probability score (CRPS), but skill differences are seen mostly for light precipitation, and tend to diminish towards higher precipitation accumulation. Further comparisons over specific heavy rainfall events suggest that GNN-CSGD tends to produce PQPFs that are less dispersive in space, and feature more accurate spatial coverage.