A novel hybrid model for multi-step daily AQI forecasting driven by air pollution big data

Abstract

Air pollution is considered as a serious issue for the society and economy. Air pollution forecasting can offer a timely and effective early warning. In the current mainstream studies, the information from high-resolution air pollution data is usually ignored. In this study, the useful information from high-resolution big data (1 h) is utilized sufficiently. In the proposed model, high-resolution data (1 h) of different air pollution indices (AQI, PM2.5, PM10, SO2, NO2, O3, and CO concentrations) are utilized to enhance the performance, including 7 × 24,000 data samples. Besides, the variational mode decomposition (VMD) and the least absolute shrinkage and selector operation (LASSO) are proposed to preprocess and reshape the input data. The stacked auto-encoder (SAE) is proposed to reduce dimension and extract features. The deep echo state network (DESN) is trained to generate daily forecasting results. The high-resolution input is converted to low-resolution output in this way. Air pollution data of four polluted Chinese cities are utilized to verify the effectiveness, stability, and universality of the proposed model. The proposed model can outperform low-resolution forecasting models and other benchmark models.