Study Of Atmospheric Chemistry Data Assimilation Based On Ensemble Kalman Filter

Air pollution events are increasing in China with the rapid economic development in the recent 40 years.For air pollution control and alert,it is necessary to forecast air quality with numeric atmospheric chemistry models.Nowadays,the performance of atmospheric chemistry model is mainly limited by the uncertainties of initial chemistry conditions?CICs?,anthropogenic emission and chemical reaction parameters within the model.In this paper,the above three sources of uncertainties will be decreased with data assimilation of observations by ensemble kalman filter?En KF?,which is expected to improve the performance of numeric air quality forecast.1.We investigated the data assimilation of multiple observations?from near-surface,lidar and satellite?to improve the aerosol CICs.Three types of observations,aerosol optical depth from the Moderate Resolution Imaging Spectroradiometer,surface particulate matter with diameters less than 2.5(PM_2.5)and 10?m(PM₁₀),and aerosol extinction coefficient?AEXT?profiles from ground?based lidars,were separately and simultaneously assimilated using the Weather Research and Forecasting Model with the Chemistry/Data Assimilation Research Testbed?WRF?Chem/DART?chemical weather forecasting/data assimilation system.That is different from previous researches that only assimilate one or two types.Two cases in June and November 2018were selected over middle and eastern China.We also analyzed the impact of systematic bias between observation and simulation and its bias correction?BC?.?1?No matter whether BC was applied,compared to the experiment without data assimilation?DA?,DA of single?type observations is always closer to the type of observations assimilated.?2?Without BC,however,DA of aerosol optical depth or AEXT sometimes significantly degraded the error performance for PM_2.5.This problem is caused by the inconsistency of bias tendencies when modeling aerosol optical properties and surface aerosol mass.It is found that WRF?Chem tends to predict dryer air within the boundary layer over eastern China,which may have played a role in the underestimation of AEXT even when PM_2.5 was overestimated.?3?After applying a simple BC,the problem was alleviated.DA of multiple observations with BC gives the best overall error performance when validated against all types of observations.Given the performance in reproducing PM_2.5 and PM₁₀ is similar,DA of multiple observations could respectively decrease the RMSE in reproducing AOD and AEXT by 9%and 19%when compared with DA of single kind of observations.This suggests that BC is important in DA of multiple observations and that the simultaneous DA of aerosol observations with different vertical information can work synergistically to improve aerosol forecasts.2.We used data assimilation to adjust multiconstituent CICs and anthropogenic emissions together.Previous studies in this field usually assimilate only one constituent of observation and seldom considers secondary pollution like Ozone?O₃?.In this paper,we use WRF?Chem/DART system with multiconstituent data assimilation to constrain related emissions?including the emission of O₃ precursors?and investigate the possible forecast improvement of O₃,sulfur dioxide?SO₂?,nitrogen dioxide?NO₂?,carbon monoxide?CO?,PM_2.5 and PM₁₀ over eastern China.We assimilate surface in situ observations of SO₂,NO₂,O₃,CO,PM_2.5 and PM₁₀,and satellite aerosol optical depth to adjust the related anthropogenic emissions as well as the chemical initial conditions.We validate our forecast results out to 72 hr by comparison with the in situ observations.Results show that updated emissions become lower by 10-50%for most species over most regions.Updated CICs could improve the model forecast only for the first 24 hours.Updated CICs and emissions together could improve the model performance for the whole 72 hr forecast and reduce the root mean square error by about 22%,25%,60%,10%and 35%for O₃,SO₂,NO₂,CO and PM_2.5respectively.Particulate matter with a diameter between 2.5 and 10?m(PM_2.5?10)is slightly improved due to the limited anthropogenic contribution to it.In a sensitivity experiment with a different update interval,the CO improvement is found to be sensitive to the cycling time used to update the CO emissions.In another sensitivity experiment when NO₂ observations are not assimilated and nitrogen oxides?NOx?emission are adjusted by only O₃,NO₂ forecasts show similar root mean square error improvement but have lower spatial correlation,indicating the value and limitation of the O₃?NOx cross?variable relationship.3.We used data assimilation to adjust sulfate reaction rate,in which a new reaction parametrization replaced detailed mechanisms.The sulfate formation is usually underestimated in original WRF-Chem with MOZCART mechanism partly because MOZCART only considers SO₂ oxidation by hydroxyl radical?OH?in gas phase and hydrogen peroxide?H₂O₂?in cloud.To solve the problem,we added a new sulfate?SO₄?formation parameterization which includes two imagined reactions and then adjusted the unknown 6 parameters within the parameterization by assimilating in-situ SO₄,SO₂,NO₂,O₃,PM_2.5 and PM₁₀ observations with WRF-Chem/DART.The new parameterization with adjusted parameters was validated by comparing observations from different regions and time with simulations of the original model and the one with the new parameterization.Results show that the new parameterization with adjusted parameters consistently improved model's performance in reproducing the overserved SO₄ and SO₂ concentration.In our first validation case,the bias of modeling SO₄?SO₂?is-13.1?g/m³?17.0?g/m³?,which becomes 3.5?g/m³?6.3?g/m³?with the new parametrization added.Similar improvement could be seen in other cases.Detailed analysis suggests that the two imagined reactions dominate in different regions and displays the expected self-limiting and positive feedback features in previous studies.Our results also imply that chemical reaction parameterization could be potentially studied via first guessing the mathematical form and then adjusting the parameters by DA.In general,with the help of multiconstituent observations from multiple platforms,En KF DA is effective in producing more accurate spatial distribution in aerosol CICs,in decreasing the uncertainties within the multiconstituent CICs and anthropogenic emissions that includes O₃ precursors,and in calibrating reaction parameters within a new SO₄ formation parametrization.Such effects could all result in better numeric air pollution forecast.