GREY BOX AND COMPONENT MODELS TO FORECAST OZONE EPISODES: A COMPARISON STUDY UWE SCHLINK and MARIALUISA VOLTA 1 Human Exposure Res. and Epidemiology, UFZ, POBox 2, 04301 Leipzig, Germany 1 D.E.A., Università degli Studi di Brescia, Via Branze, 38, 25123 Brescia, Italy Abstract. For the purpose of short-term forecasting of high ozone concentration episodes stochastic models have been suggested and developed in the literature. The present paper compares the quality of forecasts produced by a grey box and a component time-series model. The summer ozone patterns for three European urban areas (two continental and one mediterranean) are processed. By means of forecast performance indices according to EC and WHO guidelines, the following features of the models could be found: The grey box model is highly adaptive and produces forecasts with low error variance that increases with the time horizon of forecast. The component model is more ‘stiff’ that results in a higher forecast-error variance and poorer adaption in detail. The forecast horizon, however, could be enlarged with this model. The accuracy of predicting threshold exceedance is similar for both models. This can be understood from the assumption of a cyclical time development of ozone that was made for both models. Keywords: ozone, air pollution management, time-series model, cyclostationarity, Kalman filter, short-term forecasting 1. Introduction To protect the population from adverse health effects early information and warnings of high ozone concentration are to be given in good time. Methods for the short-term prediction of atmospheric pollution have been developed based on very different approaches. The most popular methods are based on neuronal nets or stochastic models. The purpose of this paper is to compare the performance of two stochastic techniques: grey box and component time-series models are used to forecast exceedances of ozone threshold values. The ozone time-series of two consecutive summers of three European urban areas (two continental and one mediterranean) are used for calculations. A comparison of results, in terms of forecast performance indices according to the European Community directive 92/72/EEC and following the WHO guidelines, is presented and discussed. 2. Models From the damped decay of its autocorrelation function it can be realised that the ozone time-series is autocorrelated, and therefore autoregressive models Environmental Monitoring and Assessment 65: 313–321, 2000. c 2000 Kluwer Academic Publishers. Printed in the Netherlands.