AN ADAPTIVE CYBERINFRASTRUCTURE FOR THREAT MANAGEMENT IN URBAN WATER DISTRIBUTION SYSTEMS
Kumar Mahinthakumar
North Carolina State University
DDDAS Workshop, ICCS 2006, Reading, UKMay 28 - 31, 2006
2
Our Team
North Carolina State University Mahinthakumar, Brill, Ranji (PI’s) Sreepathi, Liu, Vankalaya (Grad Students) Zechman (Post-Doc)
University of Chicago Von Laszewski (PI) Haetgen (Post-Doc)
University of Cincinnati Uber (PI) Feng (Post-Doc)
University of South Carolina Harrison (PI)
Greater Cincinnati Water Works
3
Outline
Overall Scope of Project Test problem Preliminary Results
4
Many security threat problems in civil infrastructure systems
5
Water Distribution Security ProblemDiameter
6.00
12.00
24.00
36.00
in
6
Water Distribution Problem
7
Why is this an important problem? Potentially lethal and public health hazard Cause short term chaos and long term issues Diversionary action to cause service outage
Reduction in fire fighting capacity Distract public & system managers
8
What needs to be done?
Determine Location of the contaminant source(s)Contamination release history
Identify threat management optionsSections of the network to be shut downFlow controls to
Limit spread of contamination Flush contamination
9
DDDAS Aspects
Dynamic Data Optimization Simulation Workflow Computer Resources
Data Driven and Vice Versa Water Demand Data Water Quality Data
10
Grid Resource Broker and SchedulerGrid Resource Broker and Scheduler
Adaptive Simulation ControllerAdaptive Simulation Controller
Adaptive Optimization Controller
Sensors& Data
Our Cyberinfrastructure
Mobile RF AMR SensorsStatic RF AMR
Sensor NetworkStatic Water Quality
Sensor Network
Resource Needs
Resource Availability
Bayesian Monte-Carlo Engine
Optimization Engine
Simulation Model
Grid Computing Resources
Adaptive Wireless Data Receptor and Controller
Deci
sio
ns
Data
Adaptive Workflow
Portal
Algorithms & Models
Middleware & Resources
Mod
el
Para
me
ters
Mod
el
Ou
tpu
ts
11
Key DDDAS Developments Algorithm and Model Development
Dynamic Optimization Bayesian Data Sampling and Probabilistic Assessment Model Auto Calibration Model Skeletonization Network Assessment using Back Tracking
Middleware Development Adaptive Workflow Engine Adaptive Resource Management Controller Designs
Cincinnati Application Scenario Development Source Identification Sensor Network Design Flow control design
12
Water Distribution Network Modeling Solve for network hydraulics (i.e., pressure, flow)
Depends on Water demand/usage Properties of network components
Uncertainty/variability Dynamic system
Solve for contamination transport Depends on existing hydraulic conditions Spatial/temporal variation
time series of contamination concentration
13
Source Identification Problem
Find: L(x,y), {Mt}, T0
Minimize Prediction Error ∑i,t || Ci
t(obs) – Cit(L(x,y), {Mt}, T0) ||
where L(x,y) – contamination source location (x,y) Mt – contaminant mass loading at time t T0 – contamination start time Ci
t(obs) – observed concentration at sensors Ci
t(L(x,y), {Mt}, T0) – concentration from system simulation model i – observation (sensor) location t – time of observation
• unsteady• nonlinear• uncertainty/error
14
Interesting challenges
Non-unique solutions Due to limited observations (in space & time)
Resolve non-uniqueness Incrementally adaptive search
Due to dynamically updated information stream
Optimization under dynamic environments Search under noisy conditions
Due to data errors & model uncertainty
Optimization under uncertain environments
15
Resolving non-uniqueness Underlying premise
In addition to the “optimal” solution, identify other “good” solutions that fit the observations
Are there different solutions with similar performance in objective space?
Search for alternative solutions
16
Resolving non-uniqueness
Search for different solutions that are far apart in decision space
x
f(x)
17
Resolving non-uniqueness…
x
f(x)
Effects of uncertainty
18
Resolving non-uniqueness…
x
f(x)
Search for solutions that are far apart in decision space and are within an objective threshold of best solution
19
Resolving non-uniqueness…EAs for Generating Alternatives (EAGA)
Create n subpopulations
Sub Pop 1
Evaluate objfunction values
Best solution (X*, Z*)
Evaluate pop centroid(C1) in decision space
Selection(obj fn values)& EA operators
STOP?
Best Solutions
Sub Pop 2
Evaluate objfunction values
Feasible/Infeasible?
Evaluate distance in decisionspace to other populations
Selection(feasibility, dist)& EA operators
STOP?N Y NY
.
.
.
.
.
.
20
Non-uniqueness
Concentration for node 115
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 24 48 72 96 120 144 168 192 216 240 264 288Time step
Co
nce
ntr
atio
n(m
g/L
)
Concentration for node 265
0
0.0002
0.0004
0.0006
0.0008
0.001
0.0012
0.0014
0 24 48 72 96 120 144 168 192 216 240 264 288Time step
Co
ncen
trati
on
(mg
/L)
Contaminant source profile
0
500
1000
1500
2000
2500
3000
3500
1 26 51 76 101 126 151 176 201 226 251 276Time step
So
urc
e m
ass
(mg
/min
)
21
Simulation Code
EPANET from USEPA Originally for Windows environments Ported to Unix platforms including
TeraGrid architectures
22
Current Search Algorithms
Classical Random Search (CRS) Easy to implement Massively parallel Tight efficient coupling with EPANET (C)
Evolutionary Algorithms (EA) Stand alone optimization toolkit (Java)
Loose coupling with EPANET through files Specific Algorithms
Standard GA Dynamic Optimization (ADOPT)
23
Dynamic optimization Minimize Prediction Error
∑i,t || Cit(obs) – Ci
t(L(x,y), {Mt}, T0) ||
Cit(obs) – streaming data
Objective function is dynamically updated Dynamically update estimate of source
characteristics
24
Dynamic optimization…
Underlying premisePredict solutions using available information
at any time stepSearch for a diverse set of solutions (EAGA)Current solutions are good starting points for
search in the next time step
25
Preliminary Architecture
MPI
EPANET
EPANET
EPANET
Optimization Toolkit
Sensor Data
Karigen Workflow Engine Grid Resources
26
Example Test Problem
Single source contamination Unknowns
Source Location Start Time Release History
Algorithms Classical Random Search Evolutionary Computation
27
Single Source Contamination ProblemContaminant source profile
0
500
1000
1500
2000
2500
3000
3500
1 26 51 76 101 126 151 176 201 226 251 276Time step
So
urc
e m
ass
(mg
/min
)
28
Single source contamination problemContaminant source profile
0
500
1000
1500
2000
2500
3000
3500
1 26 51 76 101 126 151 176 201 226 251 276Time step
So
urc
e m
ass
(mg
/min
)
29
Single source contamination problem
Explain the contamination issues
Show animation
30
Solution Comparison
0
500
1000
1500
2000
2500
3000
3500
source start c1 c2 c3 c4 c5 c6
exact
CRS
ADOPT
GA
ERRORCRS = 6.5%, ADOPT = 5.7%, GA = 10.0 %
31
Preliminary Architecture
Parallel EPANET(MPI)
EPANET-Driver
Optimization Toolkit
Sensor Data
Karigen Workflow Engine Grid Resources
EPANET EPANET EPANET
32
Speedups on local clusterEPANET File based Vs. CRS
0500
10001500200025003000350040004500
1 2 4 8 16 22
Num of Processors
Tim
e(s
ec
)
Parallel EPANET-FileBased
Parallel EPANET-CRS Version
33
Java Toolkit Overhead
EA + Parallel EPANET
0500
10001500200025003000350040004500
1 2 4 8 16
Number of Processors
Tim
e(se
c) Total time
Epanet Time
EAGA time
34
Java Toolkit Overhead
EA + Parallel EPANET
0%10%20%30%40%50%60%70%80%90%
100%
1 2 4 8 16
Number of Processors
% o
f T
ota
l T
ime
EAGA time
Epanet Time
35
Speedups on local clusterParallel EPANET
0500
10001500200025003000350040004500
1 2 4 8 16 22Num of Processors
Tim
e(se
c)
File based coupling (EA)
Tight coupling (CRS)
36
Summary
A cyberinfrastructure is being developed for water distribution threat management
Adaptive dynamic optimization algorithm shows promise for detecting contamination sources
37
What’s Next?
Dynamic optimization for determining optimal location of sensors and optimal sampling frequency
True integration of workflow engine into the cyberinfrastructure
Backtracking to improve source identification search efficiency
38
Questions?