A hybrid framework for forecasting PM2.5 concentrations using multi-step deterministic and probabilistic strategy

PM2.5 concentration forecasting can provide powerful assistance to avoid health and economy loss. Recent research studies have proven that the PM2.5 concentration series is significantly non-stationary and does not obey standard distribution. In this study, a hybrid framework is proposed to solve the above difficulties. The proposed hybrid model consists of five algorithms: wavelet packet decomposition (WPD), gradient boost regression tree (GBRT), linear programming boosting (LPBoost), multi-layer perceptron (MLP), and Dirichlet process mixture model (DPMM). A hybrid preprocessing method WPD-GBRT is firstly applied to extract multi-resolution exogenous features, the LPBoost algorithm can assemble MLPs to achieve better forecasting performance, and the DPMM algorithm is finally built to model complex distribution of PM2.5 concentrations adaptively. Four pollutant data with different time intervals collected from Tangshan are utilized for data simulation. The results indicate the proposed model has excellent forecasting performance. In 1-h interval data, the index of agreement (IA) of the proposed model is 0.9490 and the coverage width-based criterion (CWC) with 99% confidence level is 0.3370.