Skip to Main content Skip to Navigation
New interface

Prévision séquentielle par agrégation d'ensemble : application à des prévisions météorologiques assorties d'incertitudes

Abstract : In this thesis, we are interested in sequential prediction problems. As a previsionist, we seek to predict, day after day, a physical variable, for instance the pressure at a given location. Several elementary predictors, from different simulations are made available to resolve this prediction problem. Nowadays indeed, previsionnists always devise several physical and thermodynamical models useful in various contexts. Rather than selecting one of these simulations, we seek to weigh them with coefficients. Each weight may be linked to the past performance of the simulation, in a more or less intuitive fashion. To devise this weights, we rely on the formalization and the theoretical results given by the individual sequences , a branch of machine learning. This domain indeed offers algorithms, that is automatic strategies, drawing experience from the past. The automatization is a valuable asset since it implies little to no maintenance cost as soon as the algorithm is integrated in the programs. Furthermore, these algorithms are given with strong theoretical guarantees, valid in a wide range of situation. The analysis of these algorithms implies that, even in the worst case scenario, the quantitativ performances of prediction are only slightly deteriorated. Firstly, we explore a theoretical part of the problem : we study online prediction of bounded stationary ergodic processes. Taking examples from the regression trees, we develop an auto-regressiv strategy, only using the past observations. Then we show that these strategies are asymptotically optimal in a stochastic setting and we then draw links with existing methods. Secondly, we expose sequential aggregation methods of meteorolog ical simulation of mean sea level pressure and of wind speed 10 meter above ground. The aim is to obtain, with a ridge regression of the weights, better prediction performance than a reference prediction, namely the deterministic prediction. This goal is attained on the given dataset with performance gains at 18 % on the mean sea level pressure and of 9 % on the wind speed. In the last chapter, we present the tools used in a probabilistic prediction framework, before using two algorithms on the aforementioned datasets. First, we explain the relevancy of probabilistic prediction and expose this domain's state of the art and the second part presents popular probabilistic scores. The used algorithm are then thoroughly descibed. The most automatized results give a relative performance gain of 18 % for the pressure variable and of 13 % for the wind speed variable.
Document type :
Complete list of metadata
Contributor : Nathalie Gaudechoux Connect in order to contact the contributor
Submitted on : Monday, December 7, 2015 - 5:40:39 PM
Last modification on : Friday, January 21, 2022 - 3:16:50 AM
Long-term archiving on: : Saturday, April 29, 2017 - 9:24:58 AM


  • HAL Id : tel-01239436, version 1



Paul Baudin. Prévision séquentielle par agrégation d'ensemble : application à des prévisions météorologiques assorties d'incertitudes . Modélisation et simulation. Université Paris 11, 2015. Français. ⟨NNT : ⟩. ⟨tel-01239436⟩



Record views


Files downloads