Ozone forecasting from meteorological variables: Part I. Predictive models by moving window and partial least squares regression

A multivariate modeling approach is presented for ozone forecasting from meteorological variables. For each prediction, a separate optimized multivariate regression model is constructed. The optimization involves the determination of the size of the training set by an internal validation procedure. A grid search is used in order to determine how many observations the training set should contain. A straightforward ordinary least squares procedure leads to systematic positive or negative deviations between measured and predicted ozone concentrations. Partial least squares regression (PLS) in combination with training-set selection and variable selection, gave an overall correlation coefficient of 0.83 between observed and measured ozone levels. Appropriate weighting of the observations in the training sets improved the result to give an overall correlation coefficient between measured and predicted ozone levels of 0.86. The dependence of the optimal size of the training set on the number and location of missing data in the data matrix was also investigated.