Deep-learning post-processing of short-term station precipitation based on NWP forecasts

Post-processing methods that rely on fusion-grided forecast products can reduce systematic biases from Numerical Weather Prediction (NWP) precipitation forecasts. However, these methods also limit the capability to forecast precipitation accurately at local stations. We constructed a Station-based Precipitation Post-processing Model (SPPM) that utilizes deep-learning algorithms, predominantly convolutional layers and ResNet modules. Based on 390 meteorological stations in North China and European Centre for Medium-Range Weather Forecasts Highest-resolution (ECMWF-HRES) forecast data, the SPPM utilizes multi-level atmospheric forecast variables and geographic variables in a small area centered on a station as predictors. The results show that the SPPM improved the threat score (TS) by 4.29%, 3.66%, 15.63%, 61.08%, and 295.83% for precipitation thresholds of 0.1, 3.0, 10.0, 20.0, and 50.0 mm/3 h, respectively. We then examined the sensitivity of predictors using the interpretable deep-learning technique Layer-wise Relevance Propagation (LRP). The results indicate that the NWP total precipitation (TP) from ECMWF is the most sensitive and important factor, followed by the low-level (850 hPa) field, single-level field, and geographic variables. Notably, TP becomes increasingly important with larger forecast grades, while the importance of variables at other levels remains relatively constant. The majority of stations exhibit consistent importance rankings as mentioned above. Finally, possible causes of variables' insensitivity at medium-level (500 hPa) and high-level (200 hPa) were discussed.