Forecasting daily of surface ozone concentration in the grand Casablanca region using parametric and nonparametric statistical models

Abstract

Forecasting concentration levels is important for planning atmospheric protection strategies. In this paper, we focus on the daily average surface ozone (O3) concentration with a short-time resolution (one day ahead) in the Grand Casablanca Region of Morocco. The database includes previous day O3 concentrations measured at Jahid station and various meteorological explanatory variables for 3 years (2013 to 2015). Taking into account the multicollinearity problem in the data, adapted statistical models based on parametric (SPLS and Lasso) and nonparametric (CART, Bagging, and RF) models were built and compared using the coefficient of determination and the root mean square error. We conclude that the parametric models predict better than nonparametric ones. Finally, from the explanatory variables stored by the SPLS and Lasso parametric models, we deduce that a very simple linear regression with five variables remains the most appropriate for the available data at Jahid station (R2 = 0.86 and RMSE = 9.60). This resulting model, with few explanatory variables to prevent missing data, has good predictive quality and is easily implementable. It is the first to be built to predict ozone pollution in the Grand Casablanca region of Morocco.