Deep Neural Network for PM2.5 Pollution Forecasting Based on Manifold Learning

Accurately forecasting pollution concentration of PM2.5 can provide early warning for the government to alert the persons suffering from air pollution. Many existing approaches fail at providing favorable results duo to shallow architecture in forecasting model that can not learn suitable features. In addition, multiple meteorological factors increase the difficulty for understanding the influence of the PM2.5 concentration. In this paper, a deep neural network is proposed for accurately forecasting PM2.5 pollution concentration based on manifold learning. Firstly, meteorological factors are specified by the manifold learning method, reducing the dimension without any expert knowledge. Secondly, a deep belief network (DBN) is developed to learn the features of the input candidates obtained by the manifold learning and the one-day ahead PM2.5 concentration. Finally, the deep features are modeled by a regression neural network, and the local PM2.5 forecast is yielded. The addressed model is evaluated by the dataset in the period of 28/10/2013 to 31/3/2017 in Chongqing municipality of China. The study suggests that deep learning is a promising technique in PM2.5 concentration forecasting based on the manifold learning.