Listy Biometryczne - Biometrical Letters Vol. 35(1998), No. 1, 37-66
|
ON FORECASTING OZONE EPISODES IN THE PARIS AREA*
Liliane Bel1, Lise Bellanger1,2, Michel Bobbia1, Gabriela Ciuperca1,2, Didier Dacunha-Castelle1, Elisabeth Gilibert4, Patrick Jakubowicz1, Georges Oppenheim1 and Richard Tomassone2,3 1 Modelisation Stochastiąue et Statistiąue dOrsay, Bat.425, Universite Paris XI, 91405 Orsay Cedex 2 Mathematique et Informatiąue, Institut National Agronomiąue, 16 rue Claude Bernard, 75231 Paris Cedex 05 3CNRS-UMR 5558, Universite Claude Bernard Lyon I, 43 Bd du 11 novembre 1918,<>r69622 Yilleurbanne Cedex 4AIRPARIF, Surveillance de la Qualite de lAir en Ile-de-France, 10 rue Crillon, 75004 Paris |
This paper presents two different statistical strategies to elucidate the dependence of ozone on primary pollutants (nitrogen oxides) and on meteorology. The aim is to forecast, at Sam of a current day, the maximum ozone value occurring in the afternoon, using 6 years (1990-95) of pollutant and meteorological data for Paris. The first method is based on classical methods using simultaneously cluster analysis, analysis of variance, discriminant analysis and stepwise regression. We identify three distinct and homogeneous groups in Paris area. Within these groups, daily curves of ozone pollution form clusters of decreasing levels; these clusters arę well discriminated by previous ozone, primary pollutant and meteorological data. The second method is based on nonparametric methods using a kernel estimator of an autoregressive function with exogenous variables. It works by analogy on climatic and pollution conditions. The forecast is a weighted sum of maxima observed in the past. We compare the two methods on 1996 data, and propose some improvements to avoid forecast errors in particular cases.
air pollution, ozone concentration, prediction, linear model, kernel nonparametric forecasting
