Relative effects of statistical preprocessing and postprocessing on a regional hydrological ensemble prediction system
Abstract. The relative roles of statistical weather preprocessing and streamflow postprocessing in hydrological ensemble forecasting at short- to medium-range forecast lead times (day 1–7) are investigated. For this purpose, a regional hydrologic ensemble prediction system (RHEPS) is developed and implemented. The RHEPS is comprised by the following components: i) hydrometeorological observations (multisensor precipitation estimates, gridded surface temperature, and gauged streamflow); ii) weather ensemble forecasts (precipitation and near-surface temperature) from the National Centers for Environmental Prediction 11-member Global Ensemble Forecast System Reforecast version 2 (GEFSRv2); iii) NOAA’s Hydrology Laboratory-Research Distributed Hydrologic Model (HL-RDHM); iv) heteroscedastic censored logistic regression (HCLR) as the statistical preprocessor; v) two statistical postprocessors, an autoregressive model with a single exogenous variable (ARX(1,1)) and quantile regression (QR); and vi) a comprehensive verification strategy. To implement the RHEPS, 1 to 7 days weather forecasts from the GEFSRv2 are used to force HL-RDHM and generate raw ensemble streamflow forecasts. Forecasting experiments are conducted in four nested basins in the U.S. middle Atlantic region, ranging in size from 381 to 12,362 km2. Results show that the HCLR preprocessed ensemble precipitation forecasts have greater skill than the raw forecasts. These improvements are more noticeable in the warm season at the longer lead times (> 3 days). Both postprocessors, ARX(1,1) and QR, show gains in skill relative to the raw ensemble flood forecasts but QR outperforms ARX(1,1). Preprocessing alone has little effect on improving the skill of the ensemble flood forecasts. Indeed, postprocessing alone performs similar, in terms of the relative mean error, skill, and reliability, to the more involved scenario that includes both preprocessing and postprocessing. We conclude that statistical preprocessing may not always be a necessary component of the ensemble flood forecasting chain.