Forecasting the air pollution concentration with neural networks

Abstract

Air pollution is a global problem, which has a major impact on human health. Every year concentrations of many air pollutants cause a large number of deaths. In Europe, particularly Poland, there is poor air quality. In this paper, we deal with forecasting the concentration of air pollutants. We propose four deep learning-based methods for forecasting, which include temporal convolutional network (TCN), Kolmogorov-Arnold Network (KAN), fully convolutional network (FCN), and gated recurrent unit (GRU). Each of the methods is used in three different configurations. We generate predictions for eight air pollutants, from eight cities during a heating season in Poland. Extensive twofold experimental evaluation, combining statistical hypotheses verification and error measures (MAE, MAPE, RMSE) for more than 740 forecast models confirmed high prediction accuracy. Moreover, experiments revealed the advantage of the proposed methods over several state-of-the-art algorithms from the literature.