Deep-learning-based post-processing for probabilistic precipitation forecasting

Ensemble prediction systems (EPSs) serve as a popular technique to provide probabilistic precipitation in short- and medium-range forecasting. However, numerical models still suffer from imperfect configurations associated with data assimilation physical parameterization, which can lead systemic bias. Even state-of-the-art often fail high-quality forecasting, especially for extreme events. In this study, two deep-learning-based models—a shallow neural network (NN) deep NN convolutional layers (CNN)—were used alternative post-processing approaches further improve the forecasting of over China 1–7 days. A conventional method—the censored shifted gamma distribution-based ensemble model output statistics (CSG EMOS)—was baseline. Re-forecasts run using frozen EPS—Global Forecast System version 12—were collected raw ensembles spanning 2000 2019. The re-forecast were generated once per day consisted one control four perturbed members. We calendar year 2018 validation period 2019 testing period, remaining 18 years training. According results, terms continuous ranked probability score (CRPS) Brier score, CNN significantly outperforms model, well CSG EMOS approach ensemble, heavy or events (those exceeding 50 mm/day). remarkable degradation was seen when reducing size training samples years. spatial distribution CRPS shows that stations central better calibrated than those other regions. With time 1 day, found be superior (in CRPS) at 74.5% study stations. These results indicate NNs promising statistical