Analyse de séquences de variables aléatoires hydrologiques à l’aide de modèles de changement de régime exploitant des variables atmosphériques

Modelling hydrologic time series using regime switching models and measures of atmospheric circulation

¹ Hydro-Québec, Institut de recherche 1800 boul. Lionel-Boulet, Varennes, Québec, Canada, J3X 1S1
² EDF – DTG - Département Surveillance Eau et Ouvrages 21, Avenue de l’Europe - BP 41 - 38040 GRENOBLE CEDEX 09
³ Hydro-Québec, Prévisions et qualité des données hydroélectriques 75 ouest, boulevard René-Lévesque, Montréal, Québec, Canada, H2Z 1A4

Auteur de correspondance : perreault.luc@ireq.ca

Abstract

When making investment or management decisions in the hydroelectricity business, one must often consider the future state of water resources as a random variable. A frequent hypothesis is that of the stationarity of the process : the distribution function of past observations is representative of the distribution function of future observations. Observations of hydrologic regime changes during the twentieth century over various locations lead us to believe that we should consider such changes if we want to correctly assess hydrological risk, and that we should reconsider the stationarity hypothesis. The difficulty of integrating climate variability information into our models is compounded by the small samples that we must use to properly evaluate that variability. That problem created a lot of questioning in Québec during the last decade. When we study the annual inflow time series of a number of Québec’s watersheds, particularly in the northeastern part of the Québec-Labrador peninsula, we can identify alternating sequences of high and low inflows. Should we neglect possible shifts in the series when modelling and forecasting inflows ? Hydro-Québec has developed and implemented many models taking into account changepoints. Some models use hidden Markov chains to evaluate regime changes probabilities ; others depend on climate indices to issue forecasts. Finally, certain approaches combine the above models. It is shown that the use of climate indices allows better changepoint detection.