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

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