Download presentation
Presentation is loading. Please wait.
Published byPearl Howard Modified over 8 years ago
1
Accuracy of multi-parameter response surfaces generated from sensitivity coefficients Daniel Cohan and Antara Digar CMAS Conference 2009 October 19, 2009
2
Incorporating Uncertainty Analysis into Integrated Air Quality Planning U.S. EPA – Science To Achieve Results (STAR) Program Grant # R833665 DANIEL COHAN (PI) DENNIS COX ANTARA DIGAR MICHELLE BELL MENG JI ROBYN WILSON JAMES BOYLAN MICHELLE BERGIN BYEONG-UK KIM
3
Project Objectives Quantify uncertainty in the modeling that informs state-level air quality attainment plans –Control costs and emissions reductions –Pollutant responses to emission reductions –Health impacts Explore how uncertainty can be communicated and incorporated in the decision-making process
4
Multi-pollutant, multi-objective air quality planning How can we objectively evaluate disparate control options, impacting different precursors, sectors, and locations? NO x SO 2 PM 2.5 Ozone Acid deposition N deposition Human health Visibility Ecosystems and crops PM VOC NH 3 Source Emission Ambient Impact Societal Impact Attainment
5
Air Quality Modeling Pollutant sensitivities to emissions reductions Cost Assessment Control costs and emissions reductions Health Assessment Health benefits Integrated Evaluation of Control Options Attainment Evaluation Improvement at monitors Cohan et al., Environmental Management 2007 $, Tons ppb/ton ppb, Impacts
6
Causes of Uncertainty in Modeled Concentrations and Sensitivities Parametric Uncertainty: Caused by uncertainty in model input parameters –Emission inventory, reaction rates, boundary conditions, etc. –Focus of this study Structural Uncertainty: Caused by imperfections in the model’s numerical representations of atmospheric chemistry and dynamics Model/User Error
7
How parametric uncertainty affects sensitivity to ΔEmissions ΔEi*ΔEi* Digar and Cohan, manuscript in preparation
8
Probability distribution of pollutant response (ΔC) to emission control (ΔE) Dwd NOx Dwd Anth VOC Dwd Bio VOC RJs R(NO2+OH) R(NO+O3) BC (O3) BC (NOy) Impact of Emission Control Under Parametric Uncertainty Monte Carlo CMAQ impractical – Need more efficient approach ΔEΔE ΔCΔC
9
Efficient characterization of parametric uncertainty by response surface equations 1. Compute high-order sensitivities relating (Inputs) to (Pollutant Response) -- E.g.: (∂ 2 Ozone/∂E NOx ∂E BioVOC ) shows how biogenic VOC inventory affects sensitivity of ozone to NO x emissions 2. Create “surrogate model” of pollutant response to ΔEmissions as function of uncertain inputs -- ΔC actual = F(ΔEmissions, ΔInput i,j,k,… )(Taylor series) 3. Apply Monte Carlo sampled inputs in surrogate model to generate probability distribution for ΔC ΔCΔC ΔEΔE
10
Emissions Conc. C base How to compute sensitivities: Brute Force or Decoupled Direct Method C E-10% -10% C E+10% +10% E base CMAQ-HDDM (in base case) Brute Force (3 runs; finite difference)
11
Computing concentration response to emission reduction under uncertainty (2 nd -order Taylor expansion) Impact of ΔE if no uncertainty in inputs: ΔE j = -ε j E j ΔC = ε j S j (1) - 0.5ε j 2 S j (2) + … Impact if Φ k error in each input parameter P k : P k * = (1+Φ k )P k P j * = (1+Φ j )P j ΔE j * = -ε j (1+Φ j )E j ΔC* = (1+Φ j )ε j S j (1) - 0.5(1+Φ j ) 2 ε j 2 S j (2) + (1+Φ j )ε j Σ (Φ k S j,k (2) ) Previously shown accurate for ozone response to +/- 50% emissions ??? How accurate for big changes in multiple inputs ??? We’ve assumed: (1) Accurate sensitivity coefficients; (2) 2nd-order sensitivities sufficient; (3) Additive impacts of input uncertainties
12
??? Does it Work ??? Accuracy Testing of Surrogate Model 3-day air pollution episode from Georgia SIP –CMAQ v. 4.5 with CB-IV chemistry, 12-km grid –Year 2002 meteorology with Year 2009 emissions Evaluate surrogate model up to 50% E and Inputs Pollutant Impact Emission Controlled Uncertain Inputs Sensitivity Method Ozone (8-hr) Atlanta NO x E_NOx; E_VOC; R_photolysis CMAQ-HDDM PM Sulfate (24-hour) Atlanta SO 2 E_SO 2 ; E_NH 3 ; R_photolysis Brute Force
13
Highly accurate predictions of impact under extreme uncertainty
14
8-hour Ozone Performance Actual impact of -50% Atlanta NO x, if +50% E_NO x, E_VOC and R_photolysis Prediction Neglecting Uncertainty Surrogate Model Prediction
15
24-hour PM Sulfate Performance Actual impact of -50% Atlanta SO 2, if +50% E_SO2, E_NH3, and R_photolysis Surrogate Model Prediction Prediction Neglecting Uncertainty
16
Special Case: Discrete control options at coal-fired power plants Largest point sources of NO x and SO 2 Major focus of Georgia control efforts Discrete control options –SCR ~85% NO x reduction –Scrubber ~95% SO 2 reduction –Replace w/natural gas 85% NO x, 99.8% SO 2 cut Amount of emission reduction independent of domain-wide inventory
17
“Discrete Model” to predict impact of known emission reduction Compute ΔC impact of power plant control under: –Base case conditions –P k reduced by 10% Response coefficient: F k = 10(ΔC base -ΔC -10% ) Predict impact when multiple P k are uncertain: ΔC* = ΔC base + ΣΦ k F k Power Plant Emis Input P k P k,-10% 4 runs to determine each F k Targeted Reduction Base Case
18
Accuracy of Discrete Model Pollutant Impact Emission Controlled Uncertain Inputs Bias (NMB) Error (NME) R2R2 Ozone (8-hr) -85% McDonough NO x +50% all other E_NO x, E_VOC, and R_photolysis 3.3%13.1%0.993 PM Sulfate (24-hour) -99.8% McDonough SO 2 +50% all other E_SO 2, E_NH 3, and R_photolysis -0.7%3.9%0.998 Impact of Plant McDonough natural gas, if input parameters +50% Even more accurate than continuum model, because targeted at predetermined emission reduction
19
Discrete model performance for +50% change in input parameters Actual PM_SO4 Impact of -99.8% McDonough SO 2 Predicted PM_SO4 Impact of -99.8% McDonough SO 2
20
Conclusions New methods efficiently characterize impact of emission reductions under parametric uncertainty –Can use HDDM or brute force –Continuum model well suited to flexible % controls –Discrete model applicable when % control is known High confidence that both methods accurately represent relationships in underlying model –Caveat: Methods only as good as underlying model Next talk: Applying surrogate model to estimate likelihood of attainment for SIP strategy
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.