Fine particulate matter is one of the major air pollutants in urban areas, which adversely affects people\\\\\\\'s health and is considered as a serious threat to the society. Effective prediction of this pollutant can provide information for sensitive people to avoid or reduce outdoor activities and on the other hand can help regulators for efficient decision-making related to precautionary measures. This paper develops a novel hybrid algorithm for multi-step-ahead prediction of PM2.5 in which two challenges of forecasting in real applications has been taken into account. First, the challenge of employing decomposition techniques in practice is addressed. Different real-time approaches have been explored and compared in decomposition-based and noise-removal-based frameworks. Also, a real-time approach which combines feature selection and noise-removal-based technique is implemented in the framework of the proposed algorithm. The second challenge is the lack of access to real values of meteorological parameters, as influential predictors, in practical cases and use of forecasted values instead. To address the uncertainty in model inputs, Monte Carlo simulation is employed in the framework of the proposed algorithm. Probabilistic forecasts of the proposed algorithm output a distribution of the predicted values for a given time-point which can be used to calculate prediction intervals and probability of exceeding PM2.5 warning threshold. These results provide valuable information for decision-makers and regulators to take precautions and put control measures in place. According to the results, if a public alert is issued on days with at least 50% probability of exceeding PM2.5 warning level, it is observed that in 76.88% of days the algorithm indicates a correct warning for 1-day-ahead prediction which decreases to 55.00%, 41.25% and 37.50% for 2-days-ahead to 4-days-ahead forecasts.

Keywords