Quelques remarques sur l'utilisation des modèles mathématiques pour la prévision en hydrologie et météorologie

Division technique générale, Electricité de France, Grenoble

Résumé

Practical forecasting models applied by "Electricité de France" are described. The multiple regression computation is applied to the following two types of forecast : 1. Long-term forecasting of runoff volumes and the subsequent evaluation of low-water flows, based on monthly rainfall and temperature. 2. Short-term flood forecasting, based on antecedent daily or hourly rainfall, antecedent runoff and eventually a seasonal parameter. 3. Extreme floods' probability is evaluated by the "Gradex" method, in which extreme daily (or hourly) flow and extreme daily (or hourly) rainfall distributions are extrapolated together, being both assumed to decrease exponentially according to e-(x/a) (a is called "Gradex "). 4. Local daily rainfall is forecast, one and two days ahead, by multiple correlation of rainfall versus pressure at ground level, and 1000-700 mb thickness, inside a sample of analog days, selected by pattern recognition of the 700 millibars level over western Europe-North atlantic. Graphical correlation methods-including "co-axial relations" are considered to be subjective and ineffective. The main point in computing multiple regression is that it allows rapid selection of the most useful variables for the forecast and the principal part of the natural model to be isolated, even if it is non-linear. A satisfactory linear approximation can be achieved by means of simple transformations. Deterministic models are merely considered to be inventories of links pointing to possible correlations, rather than as representing actual forecasting models. There is nothing to be gained in relying exclusively on over-detailed physical schemes of debatable usefulness : a reliable choice of variables is much more critical than the exact analytical form of relations.