Stratification of NWP Forecasts for Medium-Range Ensemble Streamflow Forecasting: Fid-Bau Portal

Stratification of NWP Forecasts for Medium-Range Ensemble Streamflow Forecasting

Khan, Mudasser Muneer / Shamseldin, Asaad Y. / Melville, Bruce W. / Shoaib, Muhammad

Improving river flow forecasts for longer lead times by incorporating numerical weather predictions (NWP) into streamflow forecasting systems has attracted hydrologists in recent years. The process turns considerably complex and resource hungry when ensembles of NWP forecasts instead of any single NWP output are used to feed the flow forecasting models in order to capture the uncertainties in hydrological forecasting. This paper presents, for the first time, a comparison of three statistical stratification techniques for simplifying the input precipitation ensemble forecasts driving a river flow forecasting system. A data-driven flow forecasting model developed for the Waikato River in New Zealand using genetic expression programming (GEP) was forced by the 10-days-ahead ensemble precipitation forecasts issued by three meteorological .centers in different parts of the world including the United Kingdom, Canada, and China. The three precipitation ensembles, comprising 51, 21, and 15 members respectively, were reduced to a smaller ensemble consisting of only 5 members by static, dynamic, and cluster stratification. The 10-days-ahead river flow forecasts resulting from the full and reduced precipitation ensembles were compared against the corresponding flows observed during the whole year 2012. The three stratification mechanisms vary in complexity, and two of them are novel in their application to streamflow forecasting. The forecasts were compared for four different attributes including accuracy, reliability, resolution skill, and the maximum flow limit, each targeting a different aspect of the forecast performance. The results indicate that, in general, the flow forecasts driven by all three smaller ensembles were comparable with their full counterpart for all three forecasting centers. The cluster-stratified approach outperformed all others for most of the tested forecast attributes. The performance of static-stratified ensembles was not too far from the cluster-stratified ensembles despite it being significantly simple compared to the latter. The dynamic-stratified ensembles were least in competing with the full ensembles except for the maximum flow limit.