A New Model Using Multiple Feature Clustering and Neural Networks for Forecasting Hourly PM2.5 Concentrations, and Its Applications in China

Abstract Particulate matter with an aerodynamic diameter no greater than 2.5 μm (PM2.5) concentration forecasting is desirable for air pollution early warning. This study proposes an improved hybrid model, named multi-feature clustering decomposition (MCD)-echo state network (ESN) - particle swarm optimization (PSO), for multi-step PM2.5 concentration forecasting. The proposed model includes decomposition and optimized forecasting components. In the decomposition component, a MCD method consisting of rough sets attribute reduction (RSAR), k-means clustering (KC), and the empirical wavelet transform (EWT) is proposed for feature selection and data classification. Within the MCD, the RSAR algorithm is adopted to select significant air pollutant variables, which are then clustered by the KC algorithm. The clustered results of the PM2.5 concentration series are decomposed into several sublayers by the EWT algorithm. In the optimized forecasting component, an ESN-based predictor is built for each decomposed sublayer to complete the multi-step forecasting computation. The PSO algorithm is utilized to optimize the initial parameters of the ESN-based predictor. Real PM2.5 concentration data from four cities located in different zones in China are utilized to verify the effectiveness of the proposed model. The experimental results indicate that the proposed forecasting model is suitable for the multi-step high-precision forecasting of PM2.5 concentrations and has better performance than the benchmark models.