Prediction of outdoor PM2.5 concentrations based on a three-stage hybrid neural network model

Abstract PM2.5 concentrations forecasting has become an effective method to deal with the severe air pollution ahead. This study proposes a three-stage hybrid neural network model to forecast outdoor PM2.5 concentrations. It subsumes outlier correction, decomposition, neural network, and metaheuristic optimization to guarantee high accuracy. The adopted methods serve different purposes and ultimately aim at improving the performance of model. Firstly, a robust outlier correction preprocessing method is employed to filter the abnormal information in the original PM2.5 concentrations series and reduce its complexity. Then, the processed series is adaptively decomposed into several subseries by empirical wavelet transform. The subseries are more stationary and their trends are more obvious than original series. To forecast each subseries and generate multi-step forecasting outputs, weighted regularized extreme learning machine whose parameters are optimally determined by neural network algorithm is used. Finally, inverse empirical wavelet transform is implemented to reconstruct the subseries and correct unexpected forecasting values beyond the frequency bands. The proposed model is tested by the actual data collected from four cities in China. Results demonstrate that: (1) the proposed model has accurate forecasting performance, whose root mean square errors of 5-step daily forecasting results in Beijing, Tianjin, Shijiazhuang, and Tangshan are 4.7555μg/m3, 3.3024μg/m3, 6.4326μg/m3, and 4.3191μg/m3, respectively; (2) all components of the hybrid model are indispensable to promote accuracy; (3) the proposed model significantly outperforms seven benchmark models and five existing models; (4) the forecasting results can be utilized to analyze the trend of pollution and perform better management measures.