Estimating fugitive emissions of airborne particulate matter ......Estimating fugitive emissions of...
Transcript of Estimating fugitive emissions of airborne particulate matter ......Estimating fugitive emissions of...
Estimating fugitive emissions of airborne particulatematter using a Gaussian plume model
Bamdad Hosseini and John M. StockieSimon Fraser University
The 2015 AMMCS-CAIMS congress
June 10, 2015
1 / 18
Fugitive emissions
Companies can place sensors on well-known sources.
Unknown sources: piles of material, moving trucks, vents andwindows, etc.
Fugitive emissions are difficult to measure directly.
Indirect measurements → source inversion
2 / 18
Fugitive emissions
Teck Resources Ltd. at Trail, British Columbia.
Integrated lead-zinc smelter.
Different pollutants: zinc, lead, arsenic, SOx etc.
During August 20 to September 19,2014.
Multiple sensors: Dust-Fall jars, Xact, PM10 and TSP.
Horizontal wind velocity and direction at a single post.
3 / 18
The forward problem
Simplifying assumptions:Ground is flat.Horizontal wind velocity changes with altitude.One type of contaminant with no chemical reactions.Fixed deposition and settling velocities.Only dry deposition (Ignore washout due to rain).All sources in an area are approximated by a single point source.
4 / 18
The forward problem
Gaussian plume (Ermak’s1) solution for a point source at (0, 0, H),x-axis pointing in the direction of wind:
c(x) =q
2πσyσz |u|exp
(−y2
2σ2y
−Wsetz
2Kz−W2
setσ2z
8K2z
)
×[exp
(−H2
2σ2z
)+ exp
(−
(H + 2z)2
2σ2z
)− exp
(Wo(H + 2z)
Kz+W2
o σ2z
2K2z
)erf
(Woσz√2Kz
+H + 2z√2σ2
z
)].
(1)
c concentration
q rate of emissions
u wind velocity
Wset settling velocity
Wdep deposition velocity
Wo = Wdep − 12Wset
Kz vertical eddy diffusioncoefficient
σy, σz horizontal and verticalplume standard deviation
Linear in the emission rates!
Commonly used in industry standard software.
1Donald L. Ermak. “An analytical model for air pollutant transport and depositionfrom a point source”. In: Atmospheric Environment 11.3 (1977), pp. 231–237.
5 / 18
The forward problem
Wind data measured at times {τ1, τ2, · · · , τNw}.Piece-wise constant wind data for t ∈ (τi, τi+1].
Piece-wise constant Ermak’s solution for each sensor.
qi(t), rate of emission of source i ∈ {1, 2, · · · , Nq}.Sensor j is actively measuring on intervals {Ij,1, Ij,2, · · · , Ij,Nj},Ij,k := {t ∈ (aj,k, bj,k]}.
Model sensor measurements as dj,k = Bj∫Ij,k
∑Nqi=1 ci,j(t)dt.
Discretize on a grid {t1, t2, · · · , tNT }, over the interval (0, T ].
Concatenate and regroup to get
d = Fq
.
6 / 18
The inverse problem
Additive, uncorrelated Gaussian noise
d = Fq + εεε, ε ∼ N(0,ΣΣΣ).
Bayes’ rule2
πpost(q|d) ∝ exp
(−1
2
∥∥∥ΣΣΣ−1/2(Fq− d)∥∥∥22
)︸ ︷︷ ︸
likelihood
πprior(q)
Three choices for the prior: constant+positivity, smoothness andsmoothness+positivity
2Jari P. Kaipio and Erkki Somersalo. Statistical and computational inverse problems.Springer, 2005.
7 / 18
The inverse problem
Constant+positivity:Take p = (p1, · · · , pNq
) and q = BpReduces to Nq dimensional problem.Maximum likelihood estimator (MLE) will suffice
qc = Bpmle, pmle := arg minp�0
∥∥∥ΣΣΣ−1/2(FBp− d)∥∥∥22.
8 / 18
The inverse problem
Smoothness:Use constant emissions as prior mean.Gaussian smoothness prior, πprior(q) = N(qc,C).C = INq
⊗ L−2, where INqis the Nq ×Nq identity matrix.
L is a discretization of α(I − γ∂tt) with homogeneous Neumann bc.
0 200 400 600 800
−2
0
2
4
6
Time (hr)
mo
l/h
r
Sample 1
q1
q2
q3
q4
q5
q6
q7
0 200 400 600 800
−2
0
2
4
6
Time (hr)
mo
l/h
r
Sample 2
Linear forward map → πpost = N(qs,Cs),
qs := qc + CF∗(ΣΣΣ + FCF∗)−1(d− Fqc),
Cs := C−CF∗(ΣΣΣ + FCF∗)−1FC.
9 / 18
The inverse problem
Smoothness+positivityIntroduce an auxiliary variable: q = max{v,0}.Pose nonlinear problem for v
π(v|d) ∝ exp
(−1
2
∥∥∥ΣΣΣ−1/2(F(max{v,0})− d)∥∥∥22
)πprior(v).
Use smoothness prior on v
πprior(v) = N(qc,C)
Use Markov Chain Monte Carlo to compute vCM (posterior mean),qsp := max{vCM,0}, standard deviations, etc.
10 / 18
Application: setup
During August 20 to September 19, 2014.
PbO particulates.
Density [kgm−3] Diameter [m] Deposition vel. [ms−1] Settling vel. [ms−1]
ρ d Wdep Wset
9530 5× 10−6 0.005 0.0026
Wind data at 10 minute intervals.
0 5 10 15 20 25 30 35−5
0
5
10
time (days)
win
d v
elo
city
(m
/s)
raw data
regularization
0 5 10 15 20 25 30 350
100
200
300
400
time (days)
win
d d
irec
tion (
deg
)
raw data
regularization
11 / 18
Application: setup
Type Dust-fall jar TSP PM10 Xact
Sched. entire month every two days every week every hour
SNR 11 100 100 100
# meas. 1 16 6 761
12 / 18
Application: synthetic data set
0 200 400 600 800
0
1
2
3
4
5
6
Time (hr)
mol/
hr
Artificial
q1
q2
q3
q4
q5
q6
q7
0 200 400 600 800
0
1
2
3
4
5
6
Time (hr)
mol/
hr
qs
q1
q2
q3
q4
q5
q6
q7
0 200 400 600 800
0
1
2
3
4
5
6
Time (hr)
mol/
hr
qsp
q1
q2
q3
q4
q5
q6
q7
q1 q2 q3 q4 q5 q6 q70
1
2
3
4
5Average emissions
mol/
hr
Artificialq
c
qs
qsp
0 200 400 600 8000
0.5
1
1.5
2
qs standard deviation
Time (hr)
mol/
hr
0 200 400 600 8000
0.5
1
1.5
qsp
standard deviation
Time (hr)
mol/
hr
13 / 18
Application: actual data
[tn/yr]
qc 54.0
qs 57.5± 5
qsp 63.0± 5
Prev. study 25− 40
Indep. 15− 500 200 400 600 800
0
5
10
15
20
Time (hr)m
ol/
hr
qs
q1
q2
q3
q4
q5
q6
q7
0 200 400 600 8000
5
10
15
20
Time (hr)
mo
l/h
r
qsp
q1
q2
q3
q4
q5
q6
q7
q1 q2 q3 q4 q5 q6 q70
5
10
15Average emissions
mo
l/h
r
q
c
qs
qsp
0 200 400 600 8000
0.5
1
1.5
2
2.5
3
qs standard deviation
Time (hr)
mo
l/h
r
0 200 400 600 8000
0.5
1
1.5
2
2.5
3
qsp
standard deviation
Time (hr)
mo
l/h
r
14 / 18
Application: impact assessment
0.05
0.05
0.05
0.0
5
0.05
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.5
0.5
0.5
0.5
0.5
1
1
1
1
2
0.5
1
12
2 2
2
2
2
2
x [m]
y [
m]
Monthly total deposition at ground level [mg/m2]
1000 1500 2000 2500
0
500
1000
1500
2000
0.5
1
1.5
2
2.5
3
15 / 18
Application: uncertainty propagation
Approximate posterior by N(qsp, C̃sp).
Forward problem is linear
πdeposition = N(F̃qsp, F̃C̃spF̃∗).
0.05
0.05
0.05
0.0
50.05
0.1
0.1
0.1
0.1
0.1
0.1
0.1
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.5
0.5
0.5
0.5
0.5
1
1
1
1
2
0.5
1
12
2 2
2
2
2
2
x [m]
y [
m]
Monthly total deposition at ground level [mg/m2]
1000 1500 2000 2500
0
500
1000
1500
2000
0.5
1
1.5
2
2.5
3
0.01
0.01
0.0
1
0.0
1
0.01
0.01
0.0
1
0.020.02
0.0
2
0.0
2
0.02
0.02
0.05
0.05
0.05
0.05
0.1
0.1
0.1
0.05
0.05
0.1
0.1
0.02
0.1
0.2
x [m]
y [
m]
Monthly total deposition std [mg/m2]
1000 1500 2000 2500
0
500
1000
1500
2000
0
0.05
0.1
0.15
0.2
16 / 18
Closing remarks
Estimating fugitive emissions is difficult.
We solve three inverse problems.
Constant and positive emissions.Smoothness.Smoothness and positivitiy.
Insufficient data.
Room for improvement:
Better prior.Faster algorithms.Improve uncertainty propagation.Include wind uncertainty.Design of experiments.
17 / 18
Thank you!
Ermak, Donald L. “An analytical model for air pollutant transport and depositionfrom a point source”. In: Atmospheric Environment 11.3 (1977), pp. 231–237.Kaipio, Jari P. and Erkki Somersalo. Statistical and computational inverseproblems. Springer, 2005.
18 / 18