Research On Multi-model Forecast Method Of Air Quality In Qinhuangdao City

In recent years,the problem of air pollution in China is prominent,which poses a serious threat to people's production and life.With people's attention to the problem of haze and the demand for urban air quality prediction,the research on regional air quality prediction is becoming more and more urgent.Based on summarizing and analyzing the research trends and technical difficulties of air quality prediction at home and abroad,this paper takes Qinhuangdao as the research area to research air quality prediction based on the numerical model and neural network method and compares the prediction performance of the different numerical model and neural network method.On this basis,machine learning integration method is used to PM_2.5.With the concentration forecast as the research focus,the integrated forecast model of Qinhuangdao city is constructed,and the integrated experiments are carried out from July 3,2020,to July 20,2020,and the forecast effect of PM_2.5 are analyzed between different methods.The following achievements have been achieved:(1)Based on the WRF-Chem,CMAQ and CAMx,a multi-mode numerical prediction system for Qinhuangdao city is established.The system is used to simulate the spatiotemporal changes of six pollutants concentration in Qinhuangdao city from January 1,2020,to June 30,2020.The prediction performance of different models for each pollutant is analyzed comprehensively.The results show that the three models are suitable to forecast CO,PM_2.5 and O₃ concentration,the coefficient of determination range is 0.26-0.70.Among them,the CAMx has a large deviation on forecasting O₃,and the prediction effects of SO₂and NO₂ are better;the CMAQ has a well prediction result for the concentration of all pollutants,the prediction value is close to the observation value,and the model is stable;WRF-Chem has the best prediction effect for PM_2.5 and O₃,the worst prediction effect for SO₂,and the problem of underestimation of the prediction value exists.(2)Besides,based on BP,GRNN and GA-BP neural network,in this paper,the multi-mode statistical prediction models of Qinhuangdao are established,the PM_2.5Concentration prediction results of the models are compared and analyzed.The results show that PM_2.5 The average R value of BP,GRNN and GA-BP neural networks are respectively 0.80?0.86?0.81.Through the model accuracy and consistency test,we can get the prediction result of PM_2.5,the accuracy of GRNN prediction is higher,the BP neural network is more effective,and the GA-BP is more stable.(3)Finally,based on the three machine learning integration methods of random forest,gradient boosting and bagging,this paper proposes a strategy to combine the numerical and statistical methods to establish ensemble prediction models.By combining WRF-Chem,CMAQ and CAMx numerical models with BP statistical model,the ensemble prediction models are constructed,and used to simulate the PM_2.5concentration of Qinhuangdao city,from July 3,2020,to July 20,2020.The results show that the three ensemble models greatly improve the prediction accuracy,and the coefficient of determination is significantly improved to 0.84.Compared with the random forest algorithm and gradient boosting methods,the bagging algorithm has the highest degree of similarity between the prediction value and the observation value,and the most obvious improvement for the prediction result.As an effective method,it can be used to improve the accuracy of urban air quality prediction.The paper has 31 figures,17 tables,and 96 references.