Development and Case Study - IPEC

Long Term Monitoring Using
Evolutionary Multiobjective
Optimization
Dennis D. Beckmann, P.E.
IPEC
October 13, 2004

Authors
• Barbara Minsker, , National Center for
Supercomputing Applications at the
University of Illinois
• Peter Grove, Moire, , Inc.
• Jon Greetis, , Delta Environmental
• Megna Babbar, , UIUC
• Dennis Beckmann, Atlantic Richfield/BP

Agenda
• Introduction
• Background on method
• Case Study 1 – Michigan site
• Detailed analysis
• Case Study 2 – New Jersey site
• Brief discussion
• Conclusions

Perspective/Objective
• Environmental business grew up with many rule
of thumb approaches.
• Now becoming more quantitative and sophisticated
• Difference between delineation and monitoring
• Likely have some redundancy in the data
• Cost of monitoring is very significant
• Waste money collecting redundant data
• Optimize monitoring in a sound manner
• Considering both science and math
• Eliminate data redundancies

Background
• What’s important here?
• Statistical approach
• Use of Genetic Algorithm
• Optimize among several variables

LTM Optimization Process
1 2 3
Define objectives
and constraints
Create interpolation
model
Create population
of monitoring
desgns
4
Evaluate design using objectives and
constraints
5
Repeat until
convergence to
optimal solution
Apply Genetic Algorithm operations
to create new population

Optimization Process
• Use genetic algorithms to search for monitoring
designs that best meet the objective functions
and constraints
• When more than one objective exists, find optimal
tradeoffs among objectives (e.g., cost vs. errors)
• Optimization process was implemented in Multi-
objective Long Term Monitoring Optimizer
Software (M-LTMO) developed at University of
Illinois and Moire
• Automated interpolation model fitting and selection
• Multiobjective optimization to find monitoring designs
that best meet objectives

Objectives and Constraints
• Identify Decision Variables
Is each well sampled or not?
Optimization problem is to identify which wells are “in”
2 x possible sampling plans:
for 36 wells that is 7x10 10 different plans
• Identify Objective Functions
• Minimize cost: here, that is the just the number of wells
• Minimize the maximum error between actual concentrations and
those estimated with the subset of wells being considered
• Identify Constraints
• None in this formulation

Error Objective Functions
• Error objective for benzene
• 5 ppb maximum acceptable error
• Error objective for BTEX
• 100 ppb maximum acceptable error
• Locations for measuring error are important
• At monitoring wells only
• Don’t measure error at grid nodes because no data
support (actual values), only modeled values

Interpolation Modeling Process
• Identify key contaminants of concern
• Create spatial grid for interpolating COC
concentrations
• Fit interpolation models
• Such as ordinary kriging, quantile kriging,
inverse distance weighting, etc.
• Test interpolation model fit and choose
model with best performance
• Quantile kriging performed best

Interpolation Model Evaluation
• Use cross validation
• Eliminate one well from network
• Interpolate concentration at well using data
from other wells
• Compare actual vs interpolated concentration
• If the error is small, then well is redundant
• Repeat for all wells

Case Study 1 - Michigan
• Remedial Actions began in 1987 when a leaking pipeline
gasket was discovered
• Catastrophic Release - estimates of the volume released
are in the range of 350K gallons
• 14 years of monitoring data support plume stability
• BTEX concentrations reduced 80-80%
80%
Remediation History
No LNAPL observed after 1993
Groundwater Treatment / Product Skimming * 118,000 gals of LNAPL recovered
No LNAPL observed after 1993
Air Sparging
Groundwater Treatment / Product Skimming * 118,000 gals of LNAPL recovered
Air Sparging
MNA
MNA
1987 1989 1991 1993 1995 1997 1999 2001 2003

LTM Scenario / Drivers
• 36 wells currently being monitored
• 30 years of Post Closure Monitoring
required
• Optimization can be used to identify
redundant data points
• There is a cost penalty for collecting
redundant data
• Spatial and temporal analyses are possible
• Only spatial redundancy addressed at this site

N
26
MW-63
25
MW-60
36
MW-82D
MW-61
27
MW-74
24
MW-51
MW-75D
35
MW-81
21
MW-54
20
MW-53
19
28
18
MW-49
17
MW-48
16
MW-46
MW-44
23
15
MW-56
34
MW-80
12
MW-41
7
MW-28B
MW-28A
6
8
MW-27
22
MW-55
9
MW-31
MW-42
MW-19
4
MW-21
11
13
MW-36
3
5
MW-25
2
MW-16B
14
MW-43
10
MW-34
29
33 MW-76 30
31
32
MW-79
MW-77
MW-78A
MW-78B 1
MW-12A
Legend:
Current Monitoring Network - 7-Jan-04
14 Interpolation ID
MW-21
Monitoring Well ID
0 100 200 Feet
Site Map
Date:
16 Jan 2004

• Of the 36 wells, the following numbers of wells were
predicted sufficiently accurately during cross-validation:
• Benzene: 17 (within 5 ppb)
• Toluene: 32 (within 100 ppb)
• EthylBenzene: : 28 (within 100 ppb)
• Xylene: 23 (within 100 ppb)
• BTEX: 19 (within 100 ppb)
• Benzene performs quite well, but has a much stricter
acceptability threshold.
• Summing the predictions of the components of BTEX
gives a small boost in accuracy over predicting it directly.

Cross Validation Results of BTEX
(summed from constituents)
Summed BTEX Error
Zoomed In
200
150
Summed BTEX Error (ppb)
100
50
0
36
-50
-100
22
13
25
30
29
1
26
20
15
23
9
3
10
18
7
2500
31
12
Summed BTEX Error
Zoomed Out
-150
2000
-200
Interpolation ID
Interpolation ID (Well ID) sorted by
true concentration
Summed BTEX Error (ppb)
1500
1000
500
0
36
22
13
25
30
29
-500
-1000
1
26
20
15
23
Interpolation ID
9
3
10
18
7
31
12

Optimal Tradeoffs Between Errors
& Sampling Levels
Max Benzene Error (ppb)
5
4.5
4
3.5
3
2.5
2
1.5
1
0.5
0
24 26 28 30 32 34 36
Number of Wells Sampled
Benzene
BTEX
100
90
80
70
60
50
40
30
20
10
0
Max BTEX Error (ppb)

Benzene Concentrations for 30
Well Design
30-Well Predictions
All-Well Predictions
+ = locations that are not sampled
O = locations that are sampled

BTEX Concentrations for 30 Well
Design
30-Well Predictions
All-Well Predictions
+ = locations that are not sampled
O = locations that are sampled

Benzene – 28 Well Design
28-Well Predictions
All-Well Predictions
+ = locations that are not sampled
O = locations that are sampled

BTEX – 28 Well Design
28-Well Predictions
All-Well Predictions
+ = locations that are not sampled
O = locations that are sampled

BTEX Cross Validation
Comparisons
2500
Actual - Predicted (ppb)
2000
1500
1000
500
0
-500
-1000
0
All 36 Wells
28-Well Design
30-Well Design
10
13
16
21
24
26
29
34
14
22
8
2
9
17
6
30
11
Well ID

Optimization Findings
• Found good predictions at all well locations
using 28-30 wells
• 17 to 22% reduction in sampling costs possible
• 28-well solution has more difficulty interpolating
correctly in the southeast corner, although this
area is of much less concern than the leading
edge of the plume
• M-LTMO software is useful tool for identifying
data redundancies
• Further testing at a New Jersey terminal site with
more wells is underway

Case Study 2 – New Jersey
• Terminal Facility, now decommissioned
• In the process of being completely removed
• Located on the Delaware River
• Multiple aquifer system
• Only one aquifer considered in this analysis
• You could say this was a PHD project…
• (Push here, dummy)
• That is, it was an automated optimization requiring
no user knowledge of genetic algorithms or
geostatistics…load the data, push the start button.

Interpolated Optimal 80 Well
Solution for Benzene

Interpolated Optimal 66 Well
Solution for Benzene

Conclusions
• Have we approached any regulators with
this technique?
• Not yet…still in development
• But…there appears to be a lot of interest,
especially at the Federal level

• Contact info:
Thank You
Dennis D. Beckmann, P.E.
Atlantic Richfield/ a BP Affilliate
509 S. Boston, N 352
Tulsa, OK 74103
Phone 918.581.3048
Email
beckmadd@bp.com

References
• ASCE Task Committee on the state of the Art in Long Term Groundwater Monitoring Design,
Long Term Groundwater Monitoring: The State of the Art, , ed by Barbara Minsker, , American
Society of Civil Engineers, 2003. Available at
http://www.pubs.asce.org/bookdisplay.cgi?9991614
• Deb, K. Multiobjective Optimization Using Evolutionary Algorithms, , John Wiley& Sons, Ltd., NY,
2001.
• Deb, K., A Fast and Elistist Multiobjective Genetic Algorithm: NSGA-II,
IEEE Trans. Evol.
Computation, , 6(2), 182-197, 197, 2002.
• Goldberg, D.E., Genetic Algorithms in Search, Optimization & Machine Learning, , Addison Wesley,
Reading, MA, 1989.
• Goovaerts, , P., Geostatistics for Natural Resources Evaluation, , Oxford University Press, NY, 1997.
• Holland, J.H., Adaptation in Natural and Artificial Systems, , University of Michigan Press, 1975.
• Kitanidis, , P.K., Introduction to Geostatistics: : Applications in Hydrogeology, , Cambridge University
Press, NY, 1997.
• Michael, W., Integrating Data Sources to Improve Long Term Monitoring and Management: agement: A
hierarchical Machine Learning Approach, , M.S. Thesis, University of Illinois, 2002.