LISBON AIR QUALITY FORECAST USING STATISTICAL METHODS Jorge Neto (1) , Pedro Torres (2) , Francisco Ferreira (2) and Filomena Boavida (3) (1) Instituto de Meteorologia / Departamento de Observação e Redes (2) Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa (3) Instituto do Ambiente Faculdade de Ciências e Tecnologia da Universidade Nova de Lisboa Departamento de Ciências e Engenharia do Ambiente Quinta da Torre, 2829-516 Caparica, Portugal Email: jorge.neto@meteo.pt, Phone : +351-21-2948374 ABSTRACT Ozone and particulate matter levels in Southern European countries are particularly high, exceeding the established limit values, and the information and alert thresholds (in the case of ozone). Therefore, it is relevant to develop a good prediction methodology for the concentrations of these pollutants. Statistical models based on multiple regression analysis and classification and regression trees analysis were developed successfully. The models were applied in forecasting the average daily concentrations for particulate matter and average maximum hourly ozone levels, for next day, for the group of existing air quality monitoring stations in the Metropolitan Area of North Lisbon in Portugal. Key Words: Statistical forecast, particles, ozone 1. INTRODUCTION An important commitment of Portugal in the area of air quality is the fulfillment of the Portuguese and European legislation. Since ozone and particulate matter levels in Southern European countries are particularly high, exceeding the established limit values and the information and alert thresholds (in the case of ozone), it is relevant to develop a good prediction methodology for the concentrations of these pollutants. The forecasting of air pollutant concentrations is very important for areas with air quality problems. Predictions can be developed through the integration of physico- chemical relationships from both meteorology and pollutants behaviour, or by using stochastic methods based on the analysis of data series. A combination of standard statistical methods was the selected process described in this paper. Statistical models based on multiple regression analysis (MR) and classification and regression trees analysis (CART) were developed successfully applied in forecasting the average daily concentrations for both particulate matter (PM 10 ) and average maximum hourly ozone (O 3 ) levels for next day, for the existing air quality monitoring stations in the Metropolitan Area of North Lisbon in Portugal.