Impact of Ensemble Size on Forecasting Occurrence of Rainfall Using TIGGE Precipitation Forecasts

Abstract

Highly uncertain precipitation input greatly reduces the overall reliability of rainfall-runoff modeling for estimating stream flows, which in turn forms the basis for decisions leading towards sustainable management of water resources. Probabilistic weather forecasting using multiple numerical weather prediction (NWP) models known as ensemble prediction system (EPS), has gained popularity amongst meteorologists in recent years. However, keeping in mind the computational burden and complexities of hydrologic modeling, scientists are constrained to use a limited number of ensemble members to drive the operational river flow forecasting systems. In this paper, 12 weather ensembles of different size are evaluated for their performance in forecasting occurrence of precipitation over a wet month in a large New Zealand catchment by comparing them against point observations of the corresponding events at rain gauge stations. The authors investigate various configurations of 12-h accumulated precipitation forecasts including members from the grand global ensemble consisting of 179 member forecasts issued at eight data centers connected through the observing research and predictability experiment (THORPEX) Interactive Grand Global Ensemble (TIGGE) network. The ensembles are developed using four different techniques including a one-dimensional (1D) clustering approach to group the members into clusters of different sizes in order to study the impact of ensemble size on precipitation forecasting with lead time of 10 days. This approach groups the members into different clusters by minimizing the sum of square distances between the members and the cluster centre. The results indicate that there is generally little improvement in the forecast skill by increasing the ensemble size, and some smaller ensembles can reasonably replace the grand global ensemble without significantly compromising the overall quality of the probabilistic forecast for precipitation occurrence. The results further infer that the grand control ensemble (GCE) combining the control forecasts from all the centers is as good a candidate as the grand global ensemble despite being much smaller in size.