Adaptive selection of members for convective-permitting regional ensemble prediction over the western Maritime Continent

Abstract

A common issue faced by the downscaled regional ensemble prediction systems is the under-dispersiveness of the ensemble forecasts, often attributed to the lack of spread under the initial conditions from the global ensemble. In this study, a novel method that adopts an adaptive approach to selecting global ensemble members for regional downscaling has been developed. Instead of using a fixed set of pre-selected global ensemble members, the adaptive selection performs a sampling algorithm and selects the global ensemble members, which maximizes a fractions skill score (FSS)-based displacement between ensemble members. The method is applied to a convective-permitting ensemble prediction system over the western Maritime Continent, referred to as SINGV-EPS. SINGV-EPS has a grid spacing of 4.5 km and is a 12-member ensemble that is driven by the European Centre for Medium-Range Weather Forecasts (ECMWF) 51-member global ensemble. Month-long trials were conducted in June 2020 to assess the impact of adaptive selection on the ensemble forecast spread and rainfall verification scores. In both fixed pre-selection and adaptive selection experiments, SINGV-EPS was still under-dispersive. However, adaptive selection improved the ensemble spread and reduced the root-mean-square error (RMSE) of the ensemble mean in wind, temperature, and precipitation fields. Further verification of the rainfall forecasts showed that there was a reduction in the Brier score and a higher hit rate in the relative operating characteristic (ROC) curve for all rainfall thresholds when adaptive selection was applied. Additionally, the ensemble mean forecasts from adaptive selection experiments are more accurate beyond 24 h, with a higher FSS for all rainfall thresholds and neighborhood lengths. These results suggest that the adaptive selection is superior to the fixed pre-selection of global ensemble members for downscaled regional ensemble prediction.