Analysis Of Spatiotemporal Dust Distribution And Simulation And Prediction Of PM₁₀ Concentration Based On Machine Learning In The Tibet Plateau

Dust event is one of the common disastrous weathers in the Tibet Plateau,which not only seriously reduces the quality of the ecological environment,but causes significant damage to the social economy and human health.Global climate change poses a direct threat to the ecological balance of the Tibet Plateau,and extreme weathers have greatly triggered the occurrence of dust events in the area,resulting in aggravated deterioration of the plateau ecological environment.Therefore,analyzing the spatiotemporal distribution of dust and air quality in the area is very important for effectively preventing environmental pollution problems and studying the particularity and sensitivity of plateau ecology.We use the FY-4A remote sensing dust fraction values in the Tibet Plateau from2018 to 2020,and combine with various elements that the spatiotemporal distribution characteristics of dust and the air quality of each city are analyzed to establish the correlation between dust and meteorological elements,vegetation coverage,etc.Secondly,based on diversified data,6 machine learning methods are used to simulate the spatial distribution of PM₁₀ mass concentration in the Tibet Plateau with a resolution of 0.01×0.01,and a database of 24 eigenvalues is established.Aiming at the dust weather affecting the northern Tibet Plateau in March 2021,we analyze the weather situation and changes of various elements,and the sources of polluted air masses are tracked through the trajectory of HYSPLIT.Finally,the PM₁₀ mass concentration in Xining during the dust weather is predicted based on the ECMWF model forecast products and PM₁₀ observation data from December 2020 to March 2021.The main conclusions are as follows:（1）The dust activity range in the Tibet Plateau has seasonal and geographical differences.In winter,the values of dust in the central and western regions are the highest,which are positively correlated with altitude;in spring,the dust activity center moves northward and the range expands;the fraction of dust is the highest;the wind speed is the highest in the high-altitude mountains in winter,which is consistent with the distribution of dust;the Qaidam Basin and the southern side of the Qilian Mountains show opposite changes in soil water and dust;the AQI of the urban agglomeration in the northeast is higher;The days with mild pollution and above are the most in 2018,and the least in 2020,showing a downward trend as a whole.（2）The Random Forest algorithm has the lowest deviation in simulating PM₁₀mass concentration.The ranking of the simulation effects of the six algorithms is:Random Forest>Gradient Boosting Regression>Adaboost>Multiple Linear Regression>K-Nearest Neighbor>Support Vector Machine,indicating that the tree-based algorithms have a lower simulation deviation for the Tibet Plateau with sparse sites and a small amount of data;the importances of each feature value in each month during the Random Forest algorithm simulation process are different:the elements at500h Pa,100m wind,10m wind,population distribution,PM₁₀ emission,low vegetation coverage,temperature and dust values are all important factors affecting PM₁₀ mass concentration.（3）The Qaidam Basin has the highest PM₁₀ mass concentration,followed by the eastern urban agglomeration and the southern part of the Tanggula Mountains;the lowest are in the south side of the Qilian Mountains,the south side of the Gangdise Mountains and the southeastern part of the Tibet Plateau.There are significant positive correlations between dust and PM₁₀ mass concentration in the Qaidam Basin,the southern side of Kunlun Mountains,and the Ali Plateau.The stronger the surface wind speed,the lower the soil water,and the higher the PM₁₀ mass concentration.The cities corresponding to the grid points with the best simulation effect are Haibei,Hainan,Guoluo and Lasa.（4）On March 16,2021,cold air intruded into the northern Tibet Plateau and carried a large amount of sand and dust.The urban agglomeration in the Hehuang Valley has the largest pressure gradient and the poor visibility.The trajectory of the air mass in the direction of the Hexi Corridor simulated by HYSPLIT is the main air mass path affecting the Xining area.The poor atmospheric diffusion conditions leaded to the retention of a large amount of dust carried by the air mass.Compared with Zhangye City,the occurrence time of dust particle pollution in Xining City lags behind.（5）The Multiple Linear Regression algorithm can accurately predict the daily mean value of PM₁₀ mass concentration in Xining city,with the highest index of agreement and correlation coefficient.Compared with other nonlinear models,the Multiple Linear Regression prediction effect is the best;and the Support Vector Machine algorithm has the lowest index of agreement and negative correlation coefficient.The Multiple Linear Regression algorithm predicted the PM₁₀ mass concentration pollution in such dust weather in a timely and effective manner,which has a high ability to predict the transition period and the dust occurrence day with the maximum value of the particle concentration in the dust weather.