Abstract
The paper discusses methods of data mining for prediction of air pollution. Two tasks in such a problem are important: generation and selection of the prognostic features, and the final prognostic system of the pollution for the next day. An advanced set of features, created on the basis of the atmospheric parameters, is proposed. This set is subject to analysis and selection of the most important features from the prediction point of view. Two methods of feature selection are compared. One applies a genetic algorithm (a global approach), and the other-a linear method of stepwise fit (a locally optimized approach). On the basis of such analysis, two sets of the most predictive features are selected. These sets take part in prediction of the atmospheric pollutants PM10, SO2, NO2 and O3. Two approaches to prediction are compared. In the first one, the features selected are directly applied to the random forest (RF), which forms an ensemble of decision trees. In the second case, intermediate predictors built on the basis of neural networks (the multilayer perceptron, the radial basis function and the support vector machine) are used. They create an ensemble integrated into the final prognosis. The paper shows that preselection of the most important features, cooperating with an ensemble of predictors, allows increasing the forecasting accuracy of atmospheric pollution in a significant way.