Analysis of statistical post-processing methods for multi-model ensemble runoff forecasts in flood seasons

Abstract

Statistical post-processing of ensemble forecasts could effectively improve their accuracy and reliability. In this study, three typical post-processing methods including equal weight (EW), model output statistics (MOS) and Bayesian model averaging (BMA) were applied to the raw multi-model runoff forecasts during the flood period (from 1 June to 30 September) of 2010-2013, and the processed results were compared and analyzed. It is shown that BMA is a promising post-processing method with highest accuracy, but this becomes completely different at lead times of 126-240 h in 2013. The main problems for the BMA performance instability were found to be outliers and low-correlation, which affect the linear regression model fitting in the bias correction procedure. Following this, a combination model of LR and EW was proposed to improve the bias correction procedure for lead times of 126-240 h. And the test results demonstrate that the combination model is effective and efficient, and it is able to lead to both accurate and reliable multi-model ensemble runoff forecasts for longer lead times.