Given that fine particles are one of the main origins of respiratory disorders, it is considered that PM2.5 is among the important contributors to air pollution and is a serious global health concern nowadays. This paper considers a new analytical approach for the prediction of PM2.5 concentration in Hamadan, Iran, with hopes of finding some ways to reduce the negative impacts of air pollution. During the last two years, the PM2.5 hourly data was gathered; they were preprocessed, and the outlier values were imputed using K-Nearest Neighbors techniques. To increase the accuracy, the estimation was improved by applying four machine learning models, namely, random forest, decision tree, support vector machine, and linear regression. Originality is represented by merging machine learning models with the time series model ARIMA. Thus, each model hybrid takes the strengths from all, giving a higher value of prediction of PM2.5 concentration. In this study many metrics such as MSE, RMSE, MAE, precision, and recall are applied for finding out the best model performance. Probably the most relevant outcome of our results is that the combination of linear regression and ARIMA returned a significant performance boost: MSE improved by 58%, while RMSE improved by 35%. This dramatic improvement underlines the predictive potential of hybrid models for air quality forecasting and forms a milestone in the study of PM2.5 prediction for the region.
