PM2.5 concentrations forecasting using a new multi-objective feature selection and ensemble framework

Multi-step PM2.5 concentrations forecasting can help reduce the negative impact of PM2.5 on public health. In this study, a hybrid method is proposed for multi-step PM2.5 concentrations forecasting. The proposed hybrid computing framework consists of three modules: hybrid data pretreatment, multi-objective feature selection and ensemble predicting. The hybrid data pretreatment can smooth the original series, generate more predictable sublayers. The multi-objective feature selection can produce the optimal input structure by rough selection and fine selection. The ensemble predicting can generate the forecasting results with the selected input structure. Four hourly pollutant data from four different cities in China are utilized to verify the effectiveness of the proposed model. The studying results indicated that (a) the computational framework of the proposed model is proved to be effective; (b) the selected algorithms are advanced when compared with the alternative algorithms; (c) the 1-step mean absolute errors of the proposed model and three existing models on data from Hohhot are 3.2804 mu g/m(3) , 7.0232 mu g/m(3) , 5.1644 mu g/m(3) and 3.4720 mu g/m(3) , respectively. The proposed hybrid computing model can generate accurate forecasting results with relatively small computational time, when compared with several existing models.