masoud asadzadeh bryan a. tolson university of waterloo
DESCRIPTION
Multi-Objective Calibration of a Real Water Distribution Network. Masoud Asadzadeh Bryan A. Tolson University of Waterloo. Genevieve Pelletier François-Julien Delisle Manuel J. Rodriguez Laval University. Outline. Problem Definition (Single- vs. Multi-Objective Optimization). - PowerPoint PPT PresentationTRANSCRIPT
Masoud Asadzadeh
Bryan A. Tolson
University of Waterloo
Multi-Objective Calibration of a Multi-Objective Calibration of a Real Water Distribution NetworkReal Water Distribution Network
Genevieve Pelletier
François-Julien Delisle
Manuel J. Rodriguez
Laval University
• Problem Definition (Single- vs. Multi-Objective Optimization)Problem Definition (Single- vs. Multi-Objective Optimization)
OutlineOutlineOutlineOutline
2WATER 2010QC July 5-7
• Optimization AlgorithmOptimization Algorithm
• Case Study (WDN Calibration)Case Study (WDN Calibration)
• Discussion of ResultsDiscussion of Results
• Future WorkFuture Work
WATER 2010QC July 5-7
Single-Objective OptimizationSingle-Objective OptimizationSingle-Objective OptimizationSingle-Objective Optimization
3
• Minimize: f (x)
We are looking for a single best value of the objective function f(x) and the corresponding solution
WATER 2010QC July 5-7
Optimization Algorithm: DDSOptimization Algorithm: DDSOptimization Algorithm: DDSOptimization Algorithm: DDS
4
Perturb the current best solution
Initialize starting solution
Continue?STO
P
– Globally search at the startstart of the search by perturbing allall decision variables (DV) from their current best values
– Perturb each DV from a normal probability normal probability distribution centered on the current value of DV
– Locally search at the endend of the search by perturbing typically only oneonly one DV from its current best value
NN
YY
WATER 2010QC July 5-7
Multi-Objective OptimizationMulti-Objective OptimizationMulti-Objective OptimizationMulti-Objective Optimization
5
• Minimize: F(x)=[f1(x),f2(x),…,fN(x)]
f1
f2
f1
f2
Non-Conflicting Objectives
Non-Conflicting Objectives Conflicting ObjectivesConflicting Objectives
WATER 2010QC July 5-7 6
Optimization Algorithm: PA-DDSOptimization Algorithm: PA-DDSOptimization Algorithm: PA-DDSOptimization Algorithm: PA-DDS
Perturb the current ND solution
Update the set of ND solutions if
necessary
Continue?STOP
New solution is ND?
Pick the New solution
Pick a ND solution based on
crowding distance
Initialize starting
solutions
YYNN
Create the non-dominated (ND)
solutions set
YYNN
7WATER 2010QC July 5-7
Case Study: Problem DefinitionCase Study: Problem DefinitionCase Study: Problem DefinitionCase Study: Problem Definition
Delisle, 2009Delisle, 2009 Determine proper pipe diameterDetermine proper pipe diameter
Adequately simulate observationsAdequately simulate observations
Have better understanding of the systemHave better understanding of the system
230,000 People230,000 People
Modeled in EPANET2 (Université Laval)Modeled in EPANET2 (Université Laval)
4700 pipes, 3691 Nodes, 379 Km, 34.2 Km4700 pipes, 3691 Nodes, 379 Km, 34.2 Km22
15 Flow Rate Measurements15 Flow Rate Measurements
19 Pressure Measurements19 Pressure Measurements
8WATER 2010QC July 5-7
Case Study: Objective FunctionsCase Study: Objective FunctionsCase Study: Objective FunctionsCase Study: Objective Functions
Mean Absolute Error: MAE =
Σ |Hi - hi(xx)|i = 1
# obs
# obs
Hi : Observed data point
hi (xx) : Simulated data point
xx : x1, x2, …, x4700: Vector of decision variables
9WATER 2010QC July 5-7
Previous Results: Single Objective Previous Results: Single Objective Calibration, Flow OR PressureCalibration, Flow OR Pressure
Previous Results: Single Objective Previous Results: Single Objective Calibration, Flow OR PressureCalibration, Flow OR Pressure
Which Solution?
10WATER 2010QC July 5-7
New ResultsNew ResultsNew ResultsNew Results
Bi-Objective Optimization with PA-DDS can be more Effective than Single-Objective Bi-Objective Optimization with PA-DDS can be more Effective than Single-Objective Optimization with DDSOptimization with DDS
11WATER 2010QC July 5-7
Discussion of ResultsDiscussion of ResultsDiscussion of ResultsDiscussion of Results
Some Data Points are Hard to MatchSome Data Points are Hard to Match
12WATER 2010QC July 5-7
Future Work and DiscussionFuture Work and DiscussionImprove the Case StudyImprove the Case Study
Future Work and DiscussionFuture Work and DiscussionImprove the Case StudyImprove the Case Study
• Decrease the problem size by decision variable groupingDecrease the problem size by decision variable grouping
4700 Decision Variables to Fit 34 Data Points4700 Decision Variables to Fit 34 Data Points
Why some data points are hard to match?Why some data points are hard to match?
• Check the data qualityCheck the data quality
• Collect more measurementsCollect more measurements
• Check the model in the vicinity of the data pointCheck the model in the vicinity of the data point
13WATER 2010QC July 5-7
Future WorkFuture WorkImprove the Optimization AlgorithmImprove the Optimization Algorithm
Future WorkFuture WorkImprove the Optimization AlgorithmImprove the Optimization Algorithm
PA-DDS has Comparable Results with NSGAII and SPEA2PA-DDS has Comparable Results with NSGAII and SPEA2
14