Urban Air Pollution Monitoring System With Forecasting Models

Abstract

A system for monitoring and forecasting urban air pollution is presented in this paper. The system uses low-cost air-quality monitoring motes that are equipped with an array of gaseous and meteorological sensors. These motes wirelessly communicate to an intelligent sensing platform that consists of several modules. The modules are responsible for receiving and storing the data, preprocessing and converting the data into useful information, forecasting the pollutants based on historical information, and finally presenting the acquired information through different channels, such as mobile application, Web portal, and short message service. The focus of this paper is on the monitoring system and its forecasting module. Three machine learning (ML) algorithms are investigated to build accurate forecasting models for one-step and multi-step ahead of concentrations of ground-level ozone (O <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sub> ), nitrogen dioxide (NO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ), and sulfur dioxide (SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> ). These ML algorithms are support vector machines, M5P model trees, and artificial neural networks (ANN). Two types of modeling are pursued: 1) univariate and 2) multivariate. The performance evaluation measures used are prediction trend accuracy and root mean square error (RMSE). The results show that using different features in multivariate modeling with M5P algorithm yields the best forecasting performances. For example, using M5P, RMSE is at its lowest, reaching 31.4, when hydrogen sulfide (H <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> S) is used to predict SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> . Contrarily, the worst performance, i.e., RMSE of 62.4, for SO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sub> is when using ANN in univariate modeling. The outcome of this paper can be significantly useful for alarming applications in areas with high air pollution levels.