EFFICIENT CHARACTERIZATION OF UNCERTAINTY IN CONTROL STRATEGY IMPACT PREDICTIONS

1
EFFICIENT CHARACTERIZATION OF UNCERTAINTY IN
CONTROL STRATEGY IMPACT PREDICTIONS

• 8th Annual CMAS Conference
• 19-21th October, 2009

Antara Digar and Daniel S. Cohan Rice University
2
AIR QUALITY PROBLEMS

• Non-attainment of multiple pollutants (ozone
  PM2.5) in multiple regions across US

3
CHALLENGES IN PLANNING ATTAINMENT
PM2.5
Secondary Pollutants
O3
NOx
VOC
NH3
PM
CO
Pb
SOx
Measure Control Emission
Issues Controlling Multiple Pollutants ?
Nonlinear Chemistry How Much to Control ? Which
Measures are most Effective?
4
The Attainment Limbo
Does (DVF Base DV RRF) attain EPA standard?
5
WHAT IF ADDITIONAL CONTROLS NEEDED TO ATTAIN
States need to target additional pollutant
reduction by adding more emission controls
Therefore, in order to attain ? target ?Cextra
DVF - NAAQS
Model
?E
?C
CHECK ?C ? Cextra
Yes
Implement Control Strategy
Add more controls
Selection based on feasibility
No
Repeat
6
DRAWBACKS OF CURRENT PRACTICE
UNCERTAINTY
7
CAUSES OF UNCERTAINTY IN PAQM
Due to imperfections in the models numerical
representations of atmospheric chemistry and
dynamics
Due to error in model input parameters

• Emission and Reaction Rates
• Boundary Conditions
• Meteorology

8
PHOTOCHEMICAL AIR QUALITY MODELS
E or ?E
Output Pollutant Concentration (e.g. O3) or
Impact (e.g. ?O3)
9
EFFECT OF PARAMETRIC UNCERTAINTY
Uncertain Boundary Conditions
Uncertain Chemistry
Uncertain Emission
Range of Output Pollutant Concentration (e.g.
O3) or Impact (e.g. ?O3)
Uncertain Model Output
10
METHODOLOGY FOR PREDICTING ?C IMPACT OF
EMISSION REDUCTION

• Pick an emission reduction scenario
• Characterize probability distributions of
  uncertain input parameters
• Compute sensitivity coefficients to emissions
  and uncertain inputs to create surrogate model
  equations
• Apply randomly sampled (Monte Carlo) input
  parameters in surrogate model to yield
  probability distribution of ?C

Uncertainties of Input Parameter
EMISSION REDUCTION
MONTE CARLO
Output ?C
Sensitivity coefficients from HDDM or finite
difference
11
UNCERTAINTY IN INPUT PARAMETERS
Parameter Uncertainty Sigma Reference
Domain-wide NOx ? 40 (1?) 0.336 a
Domain-wide Anthropogenic VOC ? 40 (1?) 0.336 a
Domain-wide Biogenic VOC ? 50 (1?) 0.405 a
All Photolysis Rates Factor of 2 (2?) 0.347 b
R(All VOCsOH) ? 10 (1?) 0.095 a, b
R(OHNO2) ? 30 (2?) 0.131 c
R(NOO3) ? 10 (1?) 0.095 b
Boundary Cond. O3 ? 50 (2?) 0.203 a
Boundary Cond. NOy Factor of 3 (2?) 0.549 a
References aDeguillaume et al. 2007 bHanna et
al. 2001 cJPL 2006

• Note
• Based on literature review All distributions
  are assumed to be log-normal

12
UNCERTAINTY IN PREDICTING IMPACT OF CONTROL
STRATEGY
Uncertainty In Atlanta Ozone Attainment
Modeling Summer Ozone Episode May 29 June
16, 2002 meteorology Year 2009 emissions
12km grid resolution
13
ATTAINMENT PLANNING OPTIONS
CASE STUDY Ozone attainment at worst Atlanta
monitor (Confederate Avenue), accounting for
parametric uncertainty
Likelihood of Attainment when
Targeted Ozone Reduction is Perfectly Known
Targeted Ozone Reduction is Uncertain
Option 1
Option 2
Choose your own adventure
14
ATTAINMENT LIKELIHOOD FUNCTIONS

• Option 1
• Targeted Ozone Reduction is Perfectly Known
• IF ?O3 Targeted Reduction,
• THEN Attainment,
• ELSE Non-Attainment

• Option 2
• Targeted Ozone Reduction Uncertain (due to
  uncertain weather/meteorology)
• Suppose, future weather causes
• Actual Target Target 3 ppb
• (assume normally distributed)

15
FINAL LIKELIHOOD OF ATTAINMENT

• Hypothetical Emission Reduction Implement all
  identified Atlanta region NOx control options,
  and replace Plant McDonough with natural gas
• Uncertainties Considered Domain-wide emission
  rates, reaction rates, and boundary conditions
• Output Probability distribution of ?C for
  8-hour ozone at Confederate Avenue monitor, for
  days exceeding ozone threshold

Ozone Impacts From Monte Carlo / Surrogate Model
75 considering fixed target
Attainment Likelihood Function A
Probability Density
68 considering variable target
Attainment Likelihood Function B
Ozone Reduction (ppb)
COMPARISON OF TWO SCENARIOS
16
LIKELIHOOD OF ATTAINMENT AS A FUNCTION OF CONTROL
STRATEGY

• ASSUMING TARGET IS UNCERTAIN

• ASSUMING TARGET IS KNOWN

Probability Plots for Different Scenarios
17
SUMMARY

• Uncertainty is typically neglected in modeling
  impact of SIP control measures
• Efficient new method to characterize
  probabilistic impact of controls under parametric
  uncertainty
• Demonstration for Atlanta ozone case study
• Can flexibly apply with alternate control amounts
  and input uncertainties
• Can compute likelihood of attaining a known or
  uncertain pollution reduction target
• Likelihood of attainment is far more responsive
  to amount of emission control if the target is
  known (fixed)

18
FUTURE PLAN OF ACTION

• Explore the likelihood of ozone attainment under
  different available control scenarios
• Extend to winter episode for PM2.5
• Assess which controls are most effective at
  improving attainment likelihood health
• Jointly consider uncertainty in cost, AQ
  sensitivity, and health estimates

19
ACKNOWLEDGEMENT

• U.S. EPA
• For funding our project (STAR Grant R833665)
• GA EPD
• For providing emission data and baseline modeling
• CMAS

20
For further information updates of our project

• Contact antara_at_rice.edu
• Log on to http//uncertainty.rice.edu/