Statistical post-processing and generation of spatially correlated precipitation forecasts with convolutional neural networks

The raw forecasts from a numerical weather prediction (NWP) model cannot be directly used because of systematic biases. Statistical calibration is performed to produce reliable and accurate ensemble forecasts. However, this is usually done on a grid-cell by grid-cell basis, followed by the use of empirical copula to embed a realistic spatial structure in the calibrated ensemble members. One drawback of these approaches is that it is difficult to select the empirical copula. In this paper, we propose Convolutional Neural Network (CNN) based models for post-processing precipitation forecast fields and generating ensemble forecasts. Unlike the traditional approaches which are applied to individual grid-cells, the model is applied to the whole precipitation field. Monte-Carlo (MC) dropouts are used to estimate uncertainty and generate ensemble forecasts. These ensemble forecasts preserve the inherent spatial structure, thereby eliminating the need for ensemble reordering. The model is applied to NWP forecasts of Brisbane drainage basin in eastern Australia. It is evaluated on all precipitation events, including no, low and high precipitation amounts. The results show that, for all levels of precipitation, the ensemble forecasts are skillful at both the grid-cell and basin scale, and the uncertainty is estimated reliably.