Towards an Ensemble Forecast Air Quality System for New York State Michael Erickson 1 , Brian A. Colle 1 , Christian Hogrefe 2,3 , Prakash Doraiswamy 3 , Kenneth Demerjian 3 , Winston Hao 2 , Mark Beauharnois 3 , Jia-Yeong Ku 2 , and Gopal Sistla 2 1 School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY 2 New York State Department of Environmental Conservation, Albany, NY 3 Atmospheric Sciences Research Center, State University of New York at Albany, Albany, NY
19
Embed
Towards an Ensemble Forecast Air Quality System for New York State
Towards an Ensemble Forecast Air Quality System for New York State Michael Erickson 1 , Brian A. Colle 1 , Christian Hogrefe 2,3 , Prakash Doraiswamy 3 , Kenneth Demerjian 3 , Winston Hao 2 , Mark Beauharnois 3 , Jia-Yeong Ku 2 , and Gopal Sistla 2 - PowerPoint PPT Presentation
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Towards an Ensemble Forecast
Air Quality System for New York StateMichael Erickson1, Brian A. Colle1, Christian Hogrefe2,3, Prakash Doraiswamy3, Kenneth
Demerjian3, Winston Hao2, Mark Beauharnois3, Jia-Yeong Ku2, and Gopal Sistla2
1 School of Marine and Atmospheric Sciences, Stony Brook University, Stony Brook, NY2 New York State Department of Environmental Conservation, Albany, NY
3 Atmospheric Sciences Research Center, State University of New York at Albany, Albany, NY
Motivations and Goals
- Project Goal: Develop an air quality ensemble forecast system to aid operational
forecasters for New York State.
- Motivation: Could errors in the atmospheric models impact air quality forecast
simulations? Can these errors be corrected via post-processing?
- Goal of this talk: Evaluate the air quality models (AQM) and Stony Brook (SBU)
ensemble with a focus on similar biases and errors within each ensemble.
- Future: Use a post-processing technique
called Bayesian Model Averaging (BMA) to
improve the deterministic and probabilistic
forecasts within the ensembles.
Air Quality Model (AQM) Ensemble: Air Quality Model (AQM) Ensemble:
Ensemble of Air Quality Model Ensemble of Air Quality Model
ForecastsForecasts
Ensemble Air Quality Model Flowchart
Atmospheric Model Ensemble: SBU, Atmospheric Model Ensemble: SBU,
NCEP NAM, ASRC, NYSDECNCEP NAM, ASRC, NYSDEC
AQM Operational Ensemble Members
*Currently two SBU members are run in the operational AQM ensemble. Retrospective simulations
used all SBU members except those with the Ferrier microphysics.
**ASRC model was not run in the retrospective simulations.
Member Name
Met. Emis. Inv. AQM Grid Res
Initial-ize
Start Date
NCEP_12z WRF-NMM
EPA CMAQv4.6
12-km 12z Summer 2004; Winter 2004-2005; everyday since June 2005
NCEP_00z WRF-NMM
EPA CMAQv4.6
12-km 00z May 2008
SBU* MM5/WRF
NYSDEC CMAQv4.6
36-km, 12-km
00z June 2008
NYSDEC_3x WRF-NMM
NYSDEC CMAQv4.6
12-km 00z November 2008
ASRC** WRF-ARW
NYSDEC CAMxv4.5.1
12-km 00z March 2009
Synoptic Setup8-hr Max. Ozone Daily Total PM 2.5
150100
7040302010
0
10080604020
0
1-hr max PM 2.5AQI Categories1-hr Max. Ozone
50
40
30
20
10
0
130
115
100
85
70
55
30
Operational AQM Example - 8/18/2009ASRC Member - http://asrc.albany.edu/research/aqf/aqfms/camx/mfb.php
Data and Methods Retrospective simulations of particulate matter 2.5 and ozone were verified over following time periods:
- June 4, 2008 – July 22, 2008- December 1, 2008 – February 28, 2009
Regions 1, 2, and 7 were selected to represent coastal, urban and inland New York, respectively.
AQM output was compared against daily 8-hr maximum ozone and 24-hr average PM 2.5 model predictions from the AIRNOW database and official NYSDEC forecasts.
To elucidate potential error sources in the AQM ensemble, the SBU 10-m wind speed and 2-m temperature were verified with ASOS observations over the same time period.
NYSERDA Regions
SBU Ensemble Domain
AQM Retrospective Simulations SBU Ensemble Members
• F2 and F9 were used to drive CMAQ forecasts each day since June 1, 2008. They were selected based on temperature and wind verification results for summer 2007 and operational considerations.
• Two additional SBU members use the Ferrier microphysics scheme that is currently not compatible with CMAQ.
Name Model Cloud PBL Radiation Microphysics Initialization
F1 MM5 BM MY CCM2 Simple Ice GFS
F2 MM5 Grell MRF CloudRad Simple Ice WRF-NMM
F3 MM5 Grell MY CloudRad Reisner2 WRF-NMM
F5 MM5 Grell Blackadar CCM2 Simple Ice NOGAPS
F6 MM5 KF2 MY CCM2 Simple Ice CMC
F7 MM5 KF2 MRF CloudRad Reisner2 GFS
F8 WRF KF2 MY RRTM WSM3 CMC
F9 WRF BM MY RRTM WSM3 WRF-NMM
F10 WRF KF2 MY RRTM WSM3 GFS
F13 MM5 Grell Blackadar CCM2 Simple Ice GFS
F14 WRF BM YSU RRTM WSM3 NOGAPS
F15 WRF KFE MY RRTM Thompson GFS
Ozone Retrospective SimulationsTime Series – 6/4/08 to 7/22/08
•Model simulations generally track observations (in red) well.
10080604020
0
10080604020
0
10080604020
0
NCEP NAM 12zNCEP NAM 00zNYSDEC 3x SBU MM5 BMMY-GFSSBU MM5 GRMRF-NAMSBU MM5 GRMY-NAMSBU MM5 GRBK-NGPSSBU MM5 KFMY-CMCSBU MM5 KFMRF-GFS
•Ozone and PM forecasts are “L” shaped (biased) or “U” shaped (underdispersed).
•Biases and dispersion issues have also been noted in the SBU ensemble and may be negatively affecting the AQM.
•Therefore it is important to verify the SBU ensemble in juxtaposition with the AQM.
Retrospective Simulations - Rank Histograms
Winter Particulate Matter Summer Ozone
SBU/AQM Ensemble Comparison – TemperatureOzone and Bias – 6/4/08 to 7/22/08
SBU Ensemble AQI Ensemble •The cooler, shallower and cloudier simulated PBL in the MM5 MY scheme is likely resulting in lower model ozone.
•This affect may be offset in one MY member by the KF convective scheme, which has been shown to decrease cloudiness and increase simulated ozone. (Tao et al. 2008). •The MYJ WRF members have greater ozone concentrations than MY MM5, which could be the result of a higher PBL growth within the MYJ scheme. (Zielonka et al. 2008).
oC
oC
oC
SBU/AQM Ensemble Comparison – TemperaturePM 2.5 and Bias – 12/1/08 to 2/28/09
SBU Ensemble AQI Ensemble•The MM5 members using the Reisner microphysics have more PM than those using Simple Ice. PM sensitivity to cloud microphysics schemes have also been noted in Meij et al. 2009.
•Lower WRF PM concentrations have been noted compared to MM5 (Meij et al 2009) due to the increase of vertical mixing within WRF caused by warmer surface temperatures.
oC
oC
oC
SBU/AQM Ensemble Comparison – Rank Histogram
Summer Ozone AQMRegion 7
Winter PM 2.5 AQMRegion 7
Winter Wind SBURegion 7
Summer Temp. SBURegion 7
•After bias correction, the SBU ensemble is underdispersed for temperature and wind speed in all regions.
•The AQI ensemble also appears to be underdispersed in the absence of biases, suggesting that a lack of variability in atmospheric forecasts could affect the air quality models.
•Post-processing techniques, such as Bayesian Model Averaging (BMA), could help correct this lack of variability in ensemble forecasting.
Post-Processing - Bayesian Model Averaging Bayesian Model Averaging (BMA, Raftery et al. 2005) has been shown to correct some model deficiencies associated with reliability and dispersion.
BMA creates a probability density function (PDF) for each ensemble member depending on the uncertainty in the model forecast and weights the result based on its performance and uniqueness in the recent past.
The main advantages of BMA appear to be with probabilistic skill, although deterministic skill is also increased.
An example using the 24 hour temperature forecast from the SBU ensemble will be presented.
PDF for Temperature PDF for Wind Speed
BMA weights each member based on past performance and assigns an uncertainty.
BMARegion 1
Region 7
Region 2
Region 1
Region 7
Region 2
Bias Corrected BMA
BMA Example – Temperature Hour 24Rank Histogram- Warm Season 2007-2009
BMA Example – Temperature Hour 24Reliability > 295 K- Warm Season 2007-2009
Region 1 Region 7Region 2
Conclusions•An operational air quality forecast ensemble is currently being run using a variety of
atmospheric models, air quality models (AQM) and pollutant emission inventories.
•Particulate matter and ozone simulations track observations reasonably well in the warm and
cool seasons, although the ensemble exhibits systematic biases and underdispersion.
•Ensemble biases may be sensitive to the PBL parameterization, with the decreased (increased)
vertical mixing within the MY (YSU) scheme resulting in lower (higher) ozone and higher (lower)
PM forecasts.
•Bayesian model averaging (BMA) has been shown to correct dispersion and improve reliability
for 2-m temperature and 10-m wind speed within the SBU ensemble. Therefore BMA could
improve AQM forecasts through direct application or insertion of the post-processed SBU