Prediction of PM10 and SO2 exceedances to control air pollution in the Bay of Algeciras, Spain

In this paper, the authors apply different classification techniques in order to provide 24 h advance forecasts of the daily peaks of SO2 and PM10 concentrations in the Bay of Algeciras. K-nearest-neighbours, multilayer neural network with backpropagation and support vector machines (SVMs) are the classification methods used. The aim of this research is to obtain a suitable prediction model that would enable us to predict the peaks of pollutant concentrations in critical meteorological situations caused by the widespread existing industry and population in the area. A resampling strategy with twofold crossvalidation has been applied, using different quality indexes to evaluate the performance of the prediction models. SVM models achieved better true positive rate and accuracy (ACC) quality indexes. Results of ACC index value of 0.795 for PM10 and 0.755 for SO2 showed the ability of the model to predict peaks and non-peaks correctly.