Research On Regional Forecast Methods Of Urban Air Quality Based On Spatio-Temporal Network

Urban air pollution is a social hot issue related to people’s health.With the continuous advancement of refined air pollution management measures,it is now necessary to realize the timely discovery,tracing and prediction analysis of air pollution sources in the fine-grained grid area of the whole city.However,existing studies mainly conducted air quality forecast for single or multiple independent sparse locations,lacking fine-grained prediction analysis for urban areas and did not fully consider the spatial correlation of air quality between regions.Therefore,there are difficult to meet the needs of the refined management of urban environment and public life services.In view of the problems and limitations of existing studies,based on the dense air quality monitoring data collected by 456 micro monitoring stations deployed in Lanzhou City,China and other city data,this paper combined machine learning and deep learning technology to carried out a study on the regional forecast method of urban air quality.The specific research contents include:（1）Research on spatial map generation model of urban air quality based on XGBoost algorithm.Firstly,dense air quality monitoring data from 456 micro monitoring stations in Lanzhou City,related meteorological data and land use data for a total of 16months were collected,and grid processing was carried out on the urban area.Next,this paper constructed a spatial map generation model,named AirInferModel,based on XGBoost algorithm.This model used various city data to infer the air quality of blank grid areas,and then generated the spatial map data instance for the subsequent regional forecast research.The experimental results show that the AirInferModel can effectively infer the hourly concentrations of various air pollutants with a grid resolution of 500m×500 m.The R² of PM_2.5 hourly concentration inference was 0.80,that of CO was 0.79and that of SO₂ was 0.78.The visual analysis of spatial map clearly shows the spatial distribution and spatio-temporal variation of air quality,and provided a fine-grained visual image of urban air quality distribution.（2）Research on regional forecast model of urban air quality based on ConvLSTM.In this paper,a spatial-temporal network model,named AirPredNet,was designed based on ConvLSTM structure to meet the requirement of regional air quality forecast.This model takes the historical spatial map sequence as the input.Firstly,CNN is used to capture the spatial characteristics of air quality in the spatial map,and then LSTM is used to capture the temporal dependence in the spatial map sequence.Finally,the model directly outputs the predicted future spatial map sequence.The experimental results show that AirPredNet can effectively predict future air quality variations for 8 consecutive hours in any grid area in the city.With the increase of the prediction time step,the prediction performance gradually decreases.For the one-hour prediction,the R² of PM_2.5pollution prediction is 0.90,that of CO is 0.80,and that of SO₂ is 0.84.For the eight-hour prediction,the R²of PM_2.5 pollution prediction is 0.57,that of CO is 0.30,and that of SO₂is 0.26.（3）Research on regional forecast model of urban air quality based on CNN-GRU and feature fusion.Considering the potential scenarios of low computing power in regional air quality forecast and the capturing of air quality variations characteristics in adjacent moments,this paper proposed an improved spatio-temporal network model,named Fast-AirPredNet.This model used the feature fusion module designed by CNN to capture the air quality spatial characteristics and continuous time characteristics,which improved the multi-step prediction accuracy.and the lightweight spatio-temporal prediction structure combined with the GRU is used to reduced the model size and improved the operation efficiency.The experimental results show that compared with AirPredNet,the model training time cost of Fast-AirPredNet is reduced by 39.4%,the test time cost is reduced by 20.2%,the model parameters is reduced by 63.1%,the video memory occupation is reduced by 34.6%,and the RMSE of the eight-hour prediction is reduced by 7.1%at the highest.Meanwhile,Fast-AirPredNet is better than the popular spatio-temporal network models in the long-time multi-step prediction performance.The analysis of regional forecast results shows that the proposed method can accurately predict the future distribution and diffusion trend of urban air quality,and provided a fine-grained prediction image of urban air quality with a grid resolution of 500 m×500 m.