Quantitative Precipitation Forecast Experiment Based on Basic NWP Variables Using Deep Learning

The quantitative precipitation forecast (QPF) performance by numerical weather prediction (NWP) methods depends fundamentally on the adopted physical parameterization schemes (PS). However, due to the complexity of the physical mechanisms of precipitation processes, the uncertainties of PSs result in a lower QPF performance than their prediction of the basic meteorological variables such as air temperature, wind, geopotential height, and humidity. This study proposes a deep learning model named QPFNet, which uses basic meteorological variables in the ERA5 dataset by fitting a non-linear mapping relationship between the basic variables and precipitation. Basic variables forecasted by the highest-resolution model (HRES) of the European Centre for Medium-Range Weather Forecasts (ECMWF) were fed into QPFNet to forecast precipitation. Evaluation results show that QPFNet achieved better QPF performance than ECMWF HRES itself. The threat score for 3-h accumulated precipitation with depths of 0.1, 3, 10, and 20 mm increased by 19.7%, 15.2%, 43.2%, and 87.1%, respectively, indicating the proposed performance QPFNet improved with increasing levels of precipitation. The sensitivities of these meteorological variables for QPF in different pressure layers were analyzed based on the output of the QPFNet, and its performance limitations are also discussed. Using DL to extract features from basic meteorological variables can provide an important reference for QPF, and avoid some uncertainties of PSs.