Research On Forecasting Method Of Air Pollutant Concentration In Northwest Provincial Capital Cities Based On Machine Learning

The concentration of air pollutants is closely related to the health of residents and is a key link in the modern refined air quality forecast business.Air pollution will directly or indirectly affect people’s production,life and health,especially for some respiratory diseases.Under the background of rapid economic development and residents’pursuit of higher happiness index,it has always been the goal of people to continuously improve the accuracy of urban air pollutant concentration forecast.In this paper,the provincial capital cities in Northwest China were taken as the target area,and air quality data and meteorological data were used to construct various forecast models.Through comparing with actual observations,a combined forecast model of multiple stepwise regression and random forest（MSR-RF）was finally constructed,which improved the forecast accuracy of air pollutant concentrations,and the average fit index（IA）reached 0.86.Firstly,the spatial and temporal distribution characteristics of air pollutants in five provincial capital cities in northwest China were analyzed;Then,the forecast equations of air pollutant concentration in 5 cities were constructed by multiple stepwise regression（MSR）,and the forecast results were compared with the moving average method;Next,the Spearman correlation coefficient method and the random forest importance assessment method were used to calculate the importance of the predictors,and the factors that have a greater impact on the predictors were selected as the input variables to build the machine learning forecast model.In this way,the prediction models of air pollutant concentration based on random forest（RF）and support vector machine（SVM）were constructed,and the influence of different factor screening schemes on the forecast results and the forecast effects of different forecast models were compared;Finally,the forecasting ability of the model was improved by constructing the MSR-RF forecasting model,and the model was combined with the numerical model for mesoscale weather forecasting（WRF）to carry out a case analysis.The main findings are as follows:（1）PM_2.5,PM₁₀ and O₃ are still important targets for air pollution control in northwest cities.For these five provincial capitals,the city with the worst air quality is Xi’an,and the cities with better air quality are Xining and Yinchuan;During the period from 2015 to 2020,the concentration of O₃ showed an upward trend year by year,while the concentration of other pollutants showed a downward trend;The concentration of O₃ reached the highest value in summer and lowest in winter,while other pollutants were on the contrary;Using the RF importance assessment method to evaluate the air quality of 5 cities,it is found that during the heating period,particulate matter(including PM_2.5 and PM₁₀)has the greatest impact on air quality,and is the main source of air pollution in cities in Northwest China;In the non-heating period,O₃ has the largest weight value,followed by particulate matter.Therefore,carrying out air pollution forecasting work can provide important references for cities to carry out targeted pollutant prevention and control in different periods.（2）The selection of predictors by RF importance assessment method and Spearman correlation coefficient method was compared and analyzed,and it was found that RF importance assessment method is more suitable for selecting predictors.In the RF forecast model,compared with the Spearman correlation coefficient method,the IA of the model is improved by 3.0%on average by using the RF importance assessment method to screen the predictors;In the SVM prediction model,the IA of the RF importance assessment method is improved by an average of 2.4%compared with the Spearman correlation coefficient method.（3）The order of forecasting ability of RF,SVM and MSR on air pollutant concentration from high to low is:MSR>RF>SVM.When comparing the RF model and the SVM model,for the forecast models with a total of 35 pollution indicators in5 cities,there are as many as 34 models whose prediction effect of RF is better than that of SVM.When comparing the RF model and the MSR model,there are 20models whose prediction effect of MSR is better than that of RF,there are 15 models whose prediction effect of RF is better than that of MSR,the prediction performance of the two models is not much different,and the performance of MSR is slightly better than that of RF.（4）Based on the above studies,a combined MSR-RF model was constructed to improve the forecast accuracy of air pollutant concentrations in Northwest China cities.Compared with the single forecasting model,the combined forecasting model has stronger stability and forecasting ability,and the forecasting effect is better.Construct the MSR-RF prediction model of air pollutant concentration in each city.First,MSR was used to predict pollutant concentrations,and then RF was used to correct the prediction residuals.The prediction results of the residual sequence were obtained by training the RF model.Finally,the prediction results of the two parts were added together to obtain the final prediction result.The mean IA of the predicted results of the MSR-RF combined model is as high as 0.86,Compared with the MSR model,the MAE and RMSE of the MSR-RF model are reduced by 9.5%and 8.4%,and the IA is increased by 1.7%.Combining the WRF model with the machine learning method for case analysis makes the constructed model more practical and provides a scientific basis for better air pollution prevention and disease prevention in Northwest China.