loading page

Empirical Inverse Transform Function for Ensemble Forecast Member Selection
  • +1
  • Weiming Hu,
  • Clemente-Harding Laura,
  • Young George,
  • Guido Cervone
Weiming Hu
Pennsylvania State University Main Campus

Corresponding Author:wuh20@psu.edu

Author Profile
Clemente-Harding Laura
U.S. Army Engineer Research and Development Center
Author Profile
Young George
Pennsylvania State University, Main Campus
Author Profile
Guido Cervone
Pennsylvania State University Main Campus
Author Profile

Abstract

With the improvement in numerical weather prediction models and high-performance computing technology, ensemble modeling and probabilistic forecasts have taken on some of the most challenging tasks, such as weather model uncertainty estimation and the global climate projection. High-resolution model simulations that were deemed impossible to complete within a reasonable amount of time in the old days are now running as an ensemble to better characterize the model uncertainty. However, with advances in computation which make large parallel computing widely accessible, important questions are being increasingly addressed on how to interpret each forecast ensemble member, instead of relying on a summarization of all ensemble members. The analysis of individual ensemble members allows for an in-depth analysis of specific possible future outcomes. Thus, it is desirable to have the ability to generate a large forecast ensemble in order to help researchers understand the forecast uncertainty. But it is also crucial to determine which ensemble members are the better ones and to identify metrics to assess the uncertainty captured by each ensemble member. This work proposes the Empirical Inverse Transform (EITrans) function to address these questions. EITrans is a technique for ensemble transformation and member selection based on knowledge from historical forecasts and the corresponding observations. This technique is applied to a particular ensemble forecast to select ensemble members that would offer a sharper and more reliable distribution without compromising the accuracy of the prediction.