A Deep Belief Network Based Model for Urban Haze Prediction

Abstract

In order to improve the accuracy of urban haze prediction, a novel deep belief network (DBN)-based model was proposed. Firstly, data pertaining to both air quality and the environment (e.g. meteorology) data was monitored and collected. The primary haze influencing elements were discovered by analyzing the correlations between each of the meteorological factors and haze. Secondly, a DBN combined with multilayer restricted Boltzmann machines and a single-layer back propagation network was applied. Thirdly, the meteorological data predictions were carried out by using a competitive adaptive-reweighed method. A stable model was established by big-data training and its accuracy was verified by experiments. Results demonstrate that the pollution haze occurs in accordance with regular laws, and is greatly affected by wind direction, atmospheric pressure, and seasons. The correlation coefficient (CC) between the actual haze value and the prediction of the proposed model is 0.8, and the mean absolute error (MAE) is 26 μg/m3. Compared with the traditional prediction algorithms, the CC is improved by 18 % on average, while the MAE is reduced by 15.7 μg/m3. The proposed method has a good prospect to predict haze and investigate the main causes of it. This study provides data support for urban haze prevention and governance.

Keywords

• big data
• deep belief network
• deep learning
• haze prediction
• model