An Urban Air Quality Prediction Model Based On Deep Learning

With the rapid development of industrialization and urbanization,air pollutants has increased a lot,leading to serious environmental problems.In recent years,the concept of smart city and urban management has been rising gradually,which makes traditional broadcast and post remediation methods unable to meet the real-world needs for air pollution control.Urban air quality prediction can help citizens make prevention measures and further avoid potential health hazards.However,due to the effects of diverse spatialtemporal correlations and multi-source factors,it is challenging to make accurate urban air quality prediction.With the rapid development of artificial intelligence,deep learning technology has played a significant role in spatial-temporal prediction tasks.Therefore,we apply deep learning technology with multi-source data to study urban air quality prediction.The main work and contributions are as follows:(1)To model spatial-temporal correlations of air quality monitoring stations,we propose a Dynamic Spatial Temporal Graph Convolution Network(DSTGCN).By constructing dynamic spatial-temporal graphs,diverse spatio-temporal correlations between these monitoring stations can be directly captured at the same time.Besides,instead of using traditional time series prediction model RNN,we designed a dynamic spatial-temporal graph convolution module with pure convolution named DSTGCM.Experiments on two realworld datasets prove that DSTGCN achieves superior performance beyond state-of-art models.Compared with benchmark models,DSTGCN improves MAE metric by about4.3% and 2.6% respectively on two real-world datasets including Beijing and Tianjin,which indicates its effectiveness for air quality prediction.(2)In view of the complex effects brought by multi-source data,we analyze different features’ effects on air pollutants and reconstruct their corresponding representations.Then we adopt both spatial attention mechanism and a temporal embedding matrix to optimize the construction process of spatial-temporal correlations between stations,so as to form a Geo-Context Dynamic Spatial Temporal Graph Convolution Network(GCDSTGCN)model.GC-DSTGCN combines multi-source information,resulting in a more comprehensive modeling and consideration of urban air quality prediction.Experiments proves that GC-DSTGCN further improves the prediction accuracy and outperforms other baselines in MAE and RMSE metrics.(3)At the application level,we choose air quality monitoring stations in public area as the application scenario.First we design the overall architecture of the urban air quality prediction system and descrice the whole calculation process.Then we propose a comprehensive solution by combining an intelligent decision-making module and a visualization platform.Futhermore,the research results of this paper will be applied to the national key project from the Ministry of Science and Technology called Key Technologies of Io T and Smart City,which can not only contribute to solve trend prediction problems of public infrastructures but also give support for risk classification and intelligent decision-making.