Application of machine learning and time-series analysis for air pollution prediction

Abstract

Medical research studies show that low air quality can have a direct effect on the increased number of diseases, especially respiratory defects, but also on the increased mortality rate in people. Luckily, harmful particles and substances in the air can easily be detected and measured by using affordable sensors. The number of this type of sensors deployed in the city of Skopje, Macedonia continuously grows. The increased coverage of monitored regions, and the elevated public interest in solving this problem for obvious reasons, make the prediction of high levels of air pollution extremely beneficial. According to the available historical data, the problem of low air quality is proving to be more serious during the winter, that is during the heating season. If weather forecast is available, there is an opportunity to predict the air quality. This work reviews recent advances in air quality predictions using time-series analysis techniques, machine learning and deep learning. We proposes and evaluate two approaches for air quality prediction: combination of LSTM and convolutional neural networks and one-dimensional convolutional neural networks. The results show a promising accuracy of about 78% in predicting the level of air pollution.