Ensemble Selection Strategies to Improve Precipitation Forecasts

An ensemble prediction system (EPS) provides probabilistic forecasts, particularly for precipitation. It also furnishes estimation for forecast errors based on standard deviation of the ensemble members called the ensemble spread (the circles in figure 1). An EPS with a small spread may not predict precipitation accurately (red circle does not include the observation in Fig.1) and increasing the ensemble size is one way to increase the spread (orange dashed circle with expanded members). However, a large ensemble size may not be practical in real-time precipitation forecasts. To improve accuracy and efficiency of ensemble prediction, we developed two methods for selecting sub-ensemble members from a full ensemble based on small surface precipitation error at an initial time (green dotted circle): Localized Ensemble Mosaic Assimilation (LEMA) and Principal Component Analysis (PCA) down-selection. LEMA identifies the best-performing members locally, based on precipitation patterns and intensity in a window region around each grid point. PCA evaluates the error of each member by decomposing the precipitation forecast error into member variation and spatial functions. Both methods were evaluated using 20 precipitation events simulated by the Weather and Research Forecasting model by comparison with 100 sets of randomly selected sub-ensembles (100RD) and the full ensemble (FEns). The deterministic verification in terms of root mean square error (RMSE) showed that both down-selection methods improved precipitation intensity relative to 100RD and FEns for 18 hours, especially in the PCA sub-ensemble due to a reduction in false alarms. The probabilistic results, including Brier score (BS), and Continues Ranked Probability Score (CRPS), indicated that the LEMA sub-ensemble is superior to other sub-ensembles for 18 hours and is competitive with FEns for the first 6 hours because of a relatively larger ensemble spread. In conclusion, both down-selection techniques improve the efficiency of precipitation forecasts, with LEMA performing well in probabilistic forecasts and PCA performing well in the intensity of precipitation.