Stochastic simulation of climatic data with the Direct Sampling method
Responsable du projet Philippe Renard
Julien Straubhaar
Collaborateur Fabio Oriani
Résumé A wealth of historic climatic data has been made available in the last decades, and this is especially true for rainfall data for which records of more than a century exist for a number of locations. The statistics based on such long records are quite robust and reliable. However, most methods proceed by first deriving parametric, low order statistics from these data, and in a second step use this model for prediction. The statistical model is a simplification of reality, including possibly restrictive assumptions. Therefore the statistical models traditionally employed can only represent models up to a certain degree of complexity, and are limited by their underlying hypotheses. We argue that this does not allow using the full richness of the data sets. Training-image based methods and especially multiple-point statistics simulation have been initially developed for subsurface applications (Boucher, 2009, De Vries, et al., 2009, Guardiano and Srivastava, 1993, Hu and Chugunova, 2008, Strebelle, 2002, Wu, et al., 2008). In this field, the major limitation of most numerical models resides in the lack of data on the subsurface. Multiple-point geostatistics introduced the concept of a training image that represents the typical structures that a geologist would expect. High-order statistics of the training image are then reproduced in the simulation domain. One of the major drawbacks of the multiple-point approach is the lack of availability of relevant training data. This problem does not exist with climatic applications where data are available in large amounts. In the recent years, the stochastic hydrogeology group of the University of Neuchâtel has developed a multiple-point simulation method (Direct Sampling, DS) that allows exploiting historical data (or training data) in their whole complexity using non-parametric high-order statistics (Mariethoz, et al., 2010). Contrarily to existing training-image based methods, the Direct Sampling can simulate either categorical or continuous variables, or a combination of both in a multivariate framework. We believe this ability is of tremendous potential for applications related to climatic time series. In particular, the method reproduces the statistical properties of the training data set up to a very high order, and regardless of the complexity of these statistics. If the example data inform several variables, all variables are jointly simulated and the multivariate statistics are reproduced as well, including possibly non-linear or even multi-modal correlations. The objective of this project is therefore to extend the current range of applications of the Direct Sampling method to climatic data sets and to evaluate its performances as compared to other well established or newly developed methods such as copulas based geostatistics. We expect that this work will lead to improve stochastic modeling of climatic data by increasing the realism of the simulations accounting for the physical complexity of the phenomena while simplifying the work of the modeler.
Mots-clés Spatial and temporal statistics, Climatic variables
Type de projet Recherche fondamentale
Domaine de recherche Hydrologie, limnologie, glaciologie
Source de financement FNS - Encouragement de projets (Div. I-III)
Etat Terminé
Début de projet 1-8-2011
Fin du projet 31-7-2015
Budget alloué 237'903.00
Autre information http://p3.snf.ch/projects-134614#
Contact Philippe Renard