Forecasting O 3 levels in industrial area surroundings up to 24 h in advance, combining classification trees and MLP models

A two-step methodology was developed to forecast tropospheric ozone (O 3 ) concentration levels, k hours ahead (k = 1, 8, 12, 24), combining meteorological, air quality and industrial emissions data, across three air quality monitoring stations in Sines Portuguese region. Firstly, the best O 3 concentration predictors have been identified through Classification and Regression Trees techniques; then Multilayer Perceptron models were adopted to forecast O 3 levels for each monitoring site. The obtained generalization model performances are very good to classify in advance the expected class of O 3 concentration level. Performance results vary from 70% of success to forecast O 3 class above 70 μg/m 3 24 h in advance up to 99% to predict the next hour in advance. These successful results are favorable to be implemented in a real-time tool for health and environmental advisories, allowing the forecast of air pollutants concentrations up to 24 h ahead, improving the local air quality management systems. • A two-step framework based on CART techniques and MLP models to forecast O 3 levels. • Industrial emissions and meteorological factors stand as best pollutant predictors. • MLP models showed very good predictive success up to 24 h in advance. • Successful results promising to establish a public-health oriented Space-Time forecast tool.