1 Development of a new analysis and data assimilation of surface pollutants at ECCCRichard Ménard, Martin Deshaies-Jacques with contributions from Alain Robichaud Air Quality Research Division International Workshop on Air Quality Forecasting Research Toronto, January 11, 2016
2 Current operational AQ analyses An example of application of AQ monitoring 2D Optimum Interpolation and Offline : AirNow observations + additional Canadian sites GEM-MACH 10 km Adjustment of error stat based on chi-square 3-month diurnal varying bias correction based on past years O3, PM2.5 – operational since Feb. 2013, but running in experimental mode since 2002 (Ménard and Robichaud 2005: ECMWF Proceedings) (Robichaud and Ménard 2014, ACP) add NO, NO2, SO2, PM10 since April 2015 (Robichaud et al. 2015, Air Qual Atmos Health) see Yulia Zaitseva poster GEM-MACH 2.5 km – PanAM Games
3 Near real-time mapping of the Air Quality Health IndexCanadian Air Quality Health Index (Stieb et al. 2008, JA&WMA) Ten year old program that has evolved from an O3-only forecast in Eastern Canada to a Canada-wide O3, NO2, PM2.5 forecast program A map of AQHI is delivered operationally (each hour) AQHI = 10/10.4100[(exp( [NO2])-1) +(exp( [O3]) -1)+(exp( [PM2.5]) -1)]
4 For Health impact studies10 year AQ analyses using AirNow CHRONOS and GEM-MACH Robichaud et Ménard, 2014, Atmos. Chem. Phys., 14, Used in epidemiology studies Ambient PM2.5, O3, and NO2 Exposures and Associations with Mortality over 16 Years of Follow-Up in the Canadian Census Health and Environment Cohort (CanCHEC) Crouze et al. (2015), Environ. Health Perspect., 123, Exposure to ambient air pollution and the incidence of dementia Chen, H., et al. Alzheimer’s & Dementia, (submitted) + 3 more papers underway The Canadian Urban Environmental (CANUE) Health Research Consortium Jeff Brook (PI) (ECCC and U of T) with 15 Canadian Universities and Governments Develop an easy access geospatial data server (e.g. Google Earth) to support quantitative research on the effect urban environment on health. Data linked to postal codes will contain information on numerous metrics, NDVI, local climatic zones, building density, land use, noise level, air pollution, greenspace, walkability. Data from 1980’s up to now. 15 year of AQ analyses will be transferred on CANUE site in NetCDF format Looking for partners to have a GEM-MACH 10 km reforecast since 2001 with OA or assimilation ??
5 Next: Improved analysis and assimilation for monitoring and forecasting Main features 3D Optimum Interpolation with data assimilation cycling each 3 hours Background error covariance mix of homogeneous error correlations and error statistics from ensembles of model forecast β Bens + (1-β) Bhomogeneous in a way similar to EnVar Use of compact support error correlation functions Have the option of using sparse matrix computation: to increase speed-up and allow much larger number of observations without data selection Overall , little or any additional computational cost over the existing objective analysis
6 Assimilation cycles (6 hrs) using our operational environment software MAESTRO (here only 00Z 06Z 12Z 18Z obs are assimilated) Impact on forecast Reduced bias (yet same diurnal signature) Reduced error variance (but not persistent) Caution: Preliminary results only 00Z, 06Z 12Z and 18Z obs are used error statistics are rudimentary and not tuned frequency of cycling will be increased to 3 hr and using all observations
7 An important goal is to have an improved analysisUse more observations Use an advanced data assimilation scheme the gain actually comes from having more realistic error covariances between observations, and between observations and model use the model to generate an ensemble of forecasts and thus an error covariance = Ensemble Kalman filter that could mean ~ 60 GEM-MACH model integrations in meteorology the ensembles are generated by a lower resolution model - reducing the effective cost by an order of magnitude But is that a good strategy for AQ models ? Spatial resolution in AQ simulation do matter
8 Main point I want to discuss is:Is to show how to obtain some spatial and temporal variability and inhomogeneity in the error statistics at little additional cost ? Practically that means: To get an improved OI or 3D-Var ? There are two sources of data that provide information about error statistics (error with respect to the truth) Observation-minus-model residuals - Provide combined information of model-minus-truth and observation-minus-truth (not separately) - Sparse only available at the observation locations Ensembles - Provide only information of model-minus-mean (not model-minus-truth) - Covers all variables and all spatial and temporal scales of the model - Does not give information about observation errors
9 How can we obtain the spatial and temporal variability ofthe error statistics fairly cheaply Consider a sampling period of 60 days period for both innovations and model output Temporal variability Error statistics for each hour of the day Spatial variability Local estimates of error variances Non-isotropic error correlations Instead of having an ensemble of 60 forecast issued at the same time, we consider a daily sample of a single forecast over a period of 60 days Over a period of 60 days we mix different weather conditions (we do not expect to capture flow-dependent structures) But the terrain features (e.g. proximity to water surfaces, mountains) and emissions correlation structures should be captured
10 Use a similar procedure to EnKF on ensemble members: Use an ensemble of model forecast valid at a given time (over 2 months) Use a similar procedure to EnKF on ensemble members: localization + spatial smoothing of model output to degrade the resolution We get terrain-dependent (land/water, mountain) error correlations OAv2 OAv1 Impact of an observation O3 – 21 UTC No additional forecast needed Seem to address known problems with OA in costal areas and close to mountains Ménard, R. and M. Deshaies-Jacques, 2017: The use of a time-serie air quality forecast to develop an ensemble error covariance. (document in preparation, Atmosphere ?)
11 isotropic correlation Appalachian MountainsPlain terrain isotropic correlation Anisotropy south-east Appalachian Mountains Anisotropy south-east Coastal station
12 specifying background error statistics ?How representative is the time-series ensemble variance for specifying background error statistics ?
13
14 H-L method estimation of (B2 , Lc )Can we get a reliable estimate of the spatial distribution of error statistics ? Ménard, R., M. Deshaies-Jacques, and N. Gasset, 2016 (JA&WMA), IWAQFR Special Issue Application of H-L method to obtain the error variances at each sites Comparison of different methods to obtain correlation-length estimates; Hollingsworth-Lonnberg, Maximum likelihood, Chi-square Local Hollingworth-Lonnberg where H-L method estimation of (B2 , Lc ) about Toronto downtown using a SOAR time 16 UTC, January 2015
15 Improving computational efficiency of the analysis solverTo accomodate large volume of observations variational analysis – 3D Var, EnVar batch processing for compact support correlations (Houtekamer and Mitchell, 2001) sparse matrix computation of Choleski decomposition and backsubstitution by taking advantage of compact support correlations Example : inversion of a 10,000 x 10,000 elements matrix with 20% sparsity non-zero elements of HBHT Config Wall time Max. memory #CPU GNU 241 s 770 MB 1 PGI 772 s 764 MB INTEL1 236 s 780 MB INTEL2 121 s 790 MB INTEL3 32 s 795 MB 8 MUMPS 25 s typical sparsity 10%
16 Summary OAv1 produced multiyear 2D maps of air quality that has been useful for health impact studies OAv2 under development generates a 3D analysis in model variables, and has the added capability of assimilation. Useful for evaluation and to improve model forecast Time-series of a single model forecast is useful to obtain terrain-dependent correlation structures, although not flow-dependent correlation structures (unless they are persistent over a period of 2 months) By using a local version of the HL it is possible to get the spatial variability of error variances Large volume of observation data and computational speed-up is obtained by using sparse matrix computation with compact correlation models
17 Thanks for listening