Convolutional Neural Network post-processing to generate spatially correlated ensemble precipitation forecasts
The forecasts from a numerical weather prediction (NWP) model have systematic biases and cannot be used directly. Statistical calibration is needed in order to generate ensemble forecast that are accurate and reliable. Typically, it is carried out on an individual grid-cell basis, with subsequent application of ensemble reordering approaches to incorporate the spatial structures in the calibrated ensemble forecast. These ensemble reordering methods, notably the widely employed Schaake shuffle approach, are based on certain templates and have a few limitations. Using Convolutional Neural Network (CNN), we propose two models for post-processing the precipitation forecast and for generation of ensemble forecasts. These ensemble forecasts display spatial structure, thereby removing dependence on ensemble reordering. CNNs are used for forecast calibration and extracting the spatial information, Monte-Carlo (MC) dropouts are then used for producing ensemble forecasts. The traditional methods are implemented on individual grid-cells, whereas the models we propose are applied to the whole forecast field. The models are implemented on NWP raw forecasts for the Brisbane Drainage Basin situated in east of Australia. They are assessed on all the precipitation levels, including no, low and high rainfall events. Results demonstrate that ensemble forecasts are well calibrated at basin and grid-cell scales, for all precipitation ranges. The uncertainty is estimated reliably, leading to skillfully calibrated ensemble forecasts.