In recent years, the atmospheric pollution in most metropolitan cities has become a crisis and the necessity of air quality forecasting has increased. Among different air pollutants, PM2.5 is considered as the major air pollutant in urbanized regions, especially because of serious harmful health effects on human being. So, there is an urgent need to develop air quality forecast programs capable of providing accurate predictions over a long future horizon. Predicting PM2.5 concentrations for several steps ahead of time is of great interest, especially in decision-making related to control policies and emergency measures such as traffic limitations, school closures, or temporarily shutting down major polluting industrial units. In this paper, commonly used multi-step ahead prediction strategies, including Recursive -Rec-, Direct -Dir-, Direct-Recursive -DirRec-, Multi-Input Multi-Output -MIMO- and Direct-MIMO -DIRMO- along with Autoregressive integrated moving average with exogenous variables -ARIMAX- and Multi-Layer Perceptron -MLP- modelling techniques are examined. Also, the independent variables are considered as time series variables and are forecasted using ARIMA/MLP model in order to be used for prediction of the dependent variables in multi-steps ahead of time. The experimental study is performed using PM2.5 data in Mashhad, Iran. Daily PM2.5 forecasts for this city is provided for the next 10 days. Four different feature selection methods are also implemented and compared. The results indicate that recursive strategy with LASSO feature selection in ARIMAX model overcomes in most of time steps.

Keywords