Spatiotemporal prediction of air quality based on LSTM neural network

• Published in: Alexandria engineering journal Vol. 60; no. 2; pp. 2021 - 2032
• Main Authors: Seng, Dewen, Zhang, Qiyan, Zhang, Xuefeng, Chen, Guangsen, Chen, Xiyuan
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.04.2021; Elsevier
• Subjects: Air quality prediction; Deep learning; LSTM; Supervised learning; Time series; Deep learning; Air quality prediction; LSTM; Supervised learning; Time series
• DOI: 10.1016/j.aej.2020.12.009
• ISSN: 1110-0168

Accurate monitoring of air quality is of great importance to our daily life. By predicting the air quality in advance, we can make timely warnings and defenses to minimize the threat to life. With a large number of environmental data, the air quality prediction based on deep learning technology is studied in depth. Based on long short-term memory (LSTM), a comprehensive prediction model with multi-output and multi-index of supervised learning (MMSL) was proposed. The particle concentration data (mainly PM2.5, means particles with aerodynamic diameter ≤ 2.5 mm) of the present monitoring station, as well as that of the nearest neighbor stations, the meteorological data, and the gaseous pollutant data in the air (mainly CO, NO2, O3, SO2) of the same period were integrated. All data were converted into the supervised learning format and normalized. The LSTM was used for training to obtain the predicted values of air quality pollution indicators (PM2.5, CO, NO2, O3, SO2). In the present study, the representative stations of the 35 monitoring stations in Beijing were selected, and input the air quality sequences of the representative stations with different data characteristics into the model to obtain the predicted concentration values of the air quality indicators of the representative stations, then calculated the average value as the overall air quality prediction result of Beijing. The air quality time series datasets collected from 35 air quality monitoring stations in Beijing from January 1, 2016, to December 31, 2017, were used to validate the performance of the model compared with other baseline models and the two most advanced models. Experimental results show that, overall, the performance of the present model is superior to other baseline models.