Short-term Prediction Of PM_2.5 Concentration Changes Based On Machine Learning Combined Model-A Case Study Of Shanghai

The rapid development of the social economy and the improvement of people’s living standards,the rapid increase in energy consumption and the number of vehicles,have led to increasingly prominent air pollution problems such as PM_2.5,and haze weather often occurs,which seriously affects people’s daily life,physical and mental health.It is of great significance to analyze changes in the concentration of air pollutants such as PM_2.5,as well as accurate prediction and early warning.Therefore,this study uses hourly air quality monitoring station data,Merra-2 reanalysis data and ECWMF meteorological data based on the analysis of changes in PM_2.5concentration in Shanghai and the correlation between PM_2.5 concentration data and meteorological data.Observe the data and innovatively use the STL decomposition method to build a Combined machine prediction model to predict the future PM_2.5 concentration.The main work and conclusions of this paper are as follows:（1）The hourly change of PM_2.5concentration in Shanghai showed a“w”shape,reaching the peak at around 9:00 am and the lowest value at around 17:00 pm in the daily concentration change.The daily average concentration varies between 15.75-75μg/m³,and the monthly average PM_2.5 concentration is 37.99μg/m³.In terms of seasonal changes,Shanghai’s winter PM_2.5 concentration is relatively high,reaching49.88μg/m³,while the summer average concentration value is lower,with an average value of 29.62μg/m³.On the whole,the air pollution in 2017 and 2018 was mainly excellent and good,and the air quality in Shanghai in 2018 was significantly better than in 2017.In 2018,the average daily PM_2.5 concentration in Shanghai with a decrease of5.8%than that in 2017,and the proportion of air quality being excellent has increased significantly.（2）There is a positive correlation between PM_2.5concentration and PM₁₀,SO₂ and NO₂ pollutants.Among them,it has the strongest correlation with PM₁₀,while it has a weaker negative correlation with CO and O₃.In the analysis of the correlation between PM_2.5 concentration and meteorological factors,the relationship between PM_2.5concentration and relative humidity,precipitation and wind speed are inverse,and the relationship between PM_2.5concentration and temperature,pressure and boundary layer height are positive.（3）In the controlled experiment,four single machine learning models including SVR,XGBoost,Random Forest and BP neural network were constructed to predict PM_2.5 concentration.The comparison and analysis of PM_2.5 concentration prediction by a single machine learning model shows that the XGBoost model performs better at all times.Respectively:2017 hourly forecast index value R²:0.95,RMSE:17.99,MAE:12.86;2018 hourly forecast index value R²:0.87,RMSE:14.50,MAE:11.34;2017-2018 daily forecast results are average,the index values are R²:0.45,RMSE:19.31,and MAE:13.81.（4）For the Combined model prediction,the Combined model constructed by the fusion of the XGBoost model and the LSTM network based on STL trend decomposition and wavelet decomposition has a better forecasting effect and higher accuracy,and it has a certain improvement with the XGBoost model with better forecasting effect in a single model.STL trend decomposition to build a Combined model has the most obvious effect in improving the prediction effect of PM_2.5concentration in Shanghai in 2018,with RMSE increasing by 40.4%and MAE increasing by 38.3%.Based on the wavelet decomposition Combined model prediction,the effect of improving the prediction effect of PM_2.5concentration in Shanghai in 2018was also the most obvious,with RMSE increased by 60%and MAE increased by 63.1%.On the whole,the two Combined models constructed in this paper have a better STL trend decomposition effect than the wavelet decomposition effect on the hourly scale,while the wavelet decomposition effect is better than the STL trend decomposition effect on the daily average scale.