Abstract ID: 172
Multi-model sub-seasonal forecasts of 2m-temperature over Europe using Wasserstein barycentre
Lead Author: Camille Marie-Jeanne Laurence Le Coz
Laboratoire de Météorologie Dynamique-IPSL, Ecole Polytechnique, Institut Polytechnique de Paris, ENS, PSL Research University, Sorbonne Université, CNRS, France
Keywords: multi-model, barycentre, Europe, temperature
Abstract: Multi-model methods have been shown to improve the skill of predictions at different time scales (from medium-range to climate). One of the goal of the sub-seasonal to seasonal (S2S) project and its database, gathering ensemble predictions from several operational centers, was to investigate the benefit and the construction of such multi-model methods for the S2S time scale. The most direct and often used method to combine ensemble forecasts from different models is to concatenate their members. The members of this new multi-model ensembles can be weighted so that every models have the same weight or based on the previous skill of the models. This method is actually equivalent to computing the (weighted-)barycentre between discrete probability distributions with respect to the L2 distance, where each single-model ensemble forecast is an input distributions.
In this work, we investigate whether a barycentre based on a different distance would improve the skill of multi-model ensemble S2S predictions. Thus, we consider a second distance between probability distributions, the Wasserstein distance. It is defined as the cost of the optimal transport between these two distributions, and has interesting properties in the distribution space such as the possibility to preserve the temporal consistency of the ensemble members.
In order to evaluate and compare the L2 and Wasserstein barycentres, we consider the combination of two models from the S2S database. More precisely we used the 2m-temperature forecasts from the European Centre Medium-Range Weather Forecasts (ECMWF) and the National Centers for Environmental Prediction (NCEP) models. The performance of the two single-model and two multi-model ensembles are evaluated over seven winters in Europe. The weights given to the models in the barycentres have an important impact on their skills. Thus, we use a cross-validation approach to estimate these weights. We show that despite NCEP having overall lower skill than ECMWF over Europe, the barycentre ensembles are generally able to perform as well or better than both of them. However, the best ensemble varies depending on the validation scores and on the location. This is an encouraging first step before implementing this method for the combination of more models.
Alexis Tantet (Laboratoire de Météorologie Dynamique-IPSL, Ecole Polytechnique, Institut Polytechnique de Paris, ENS, PSL Research University, Sorbonne Université, CNRS, France)
Rémi Flamary (Centre de Mathématiques Appliquées, Ecole Polytechnique, Palaiseau, France)
Riwal Plougonven (Laboratoire de Météorologie Dynamique-IPSL, Ecole Polytechnique, Institut Polytechnique de Paris, ENS, PSL Research University, Sorbonne Université, CNRS, France)