Problem Statement. Pressure Drop Constraint. Pressure Drop Function: Ïd : R+ â R+ Ïd (Q) â Q1.85 d4.87. Hazen-Williams Ïd (Q) â Q2 d5. Darcy-Weisbach.
Global Optimization of Sizing Problem in Pipe Networks Arvind U. Raghunathan System Dynamics & Optimization United Technologies Research Center East Hartford, CT, USA
The International Conference on Continuous Optimization, 2010
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
1 / 32
Introduction
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
2 / 32
Introduction
Motivation
Motivation Networks with Nonlinear Resistances Resistances associated with transmission lines Cost and resistance are inversely correlated Minimize cost of distribution network
(a) Water Distribution A. U. Raghunathan (UTRC)
(b) Gas Pipeline Global Opt. of Sizing in Pipe Networks
(c) Electrical Transmission ICCOPT 2010
3 / 32
Introduction
Problem Statement
Problem Statement: Minimum Cost Network
0
2
1
3
4
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
4 / 32
Introduction
Problem Statement
Problem Statement: Minimum Cost Network Pump nodes (N pump ) Fixed pressure
pump (Pi )
Pressure Drop Function φd (Q)
Elevation (Hi )
0 Edges (E) Length (Lij ) Diameter choices (D := {D1 , . . . , Dnd }) max Flow bound (Qij,d ∀d ∈ D)
2
1
3 Q Cost Function cd * *
4 Demand nodes (N \ N pump )
*
Bounds on pressure (Pimin , Pimax ) Elevation (Hi ) Demand (Qidem ) A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
D ICCOPT 2010
4 / 32
Introduction
Problem Statement
Pressure Drop Constraint Pressure Drop Function: φd : R+ → R+ 1.85
φd (Q) ∝ Q d 4.87 2 φd (Q) ∝ Q d5
Hazen-Williams Darcy-Weisbach
j i
j i
Flow from i to j
Flow from j to i
(Hi +Pi )−(Hj +Pj ) = φd (Qij )Lij
(Hi + Pi ) − (Hj + Pj ) = −φd (−Qij )Lij
Non-smooth & Non-convex (Hi + Pi ) − (Hj + Pj ) = sgn(Qij )φd (|Qij |)Lij A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
5 / 32
Introduction
Problem Statement
Non-Smooth, Non-Convex MINLP Formulation Objective Function
nd P P
cd Lij xij,d
d=1 (i,j)∈E
Choice of one diameter
nd P
xij,d = 1∀(i, j) ∈ E
d=1
xij,d ∈ {0, 1} Flows on edges
−Qijmax xij,d ≤ Qij,d ≤ Qijmax xij,d ∀(i, j) ∈ E, d = 1, . . . , n
Flow conservation
nd P P
Qij,d −
d=1 (j,i)∈E
∀i ∈ N \ N Pressure drop constraints
nd P P
Qij,d = Qidem
d=1 (i,j)∈E pump
(Hi + Pi ) − (Hj + Pj ) =
nd P
sgn(Qij )φd (|Qij,d |)Lij
d=1
∀(i, j) ∈ E Pimin ≤ Pi ≤ Pimax ∀i ∈ N A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
6 / 32
Introduction
Previous Work
Previous Work Split-pipe Formulation: Shamir & co-workers [2, 4] Allow pipes of different diameters xij,d ∈ {0, 1} → xij,d ∈ [0, 1] Qij,d = Qij
Two-stage iterative algorithm Fix flows, Qij to satisy flow demands Solve resulting Linear Program (LP) Modify flows based on sub-gradients
MILP Formulation: Artina & Walker Piecewise linear inner approximation of pressure drop constraint Posed as SOS2 constraint
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
7 / 32
Introduction
Previous Work
Previous Work (contd.)
MINLP Formulation: Bragalli & others [6] Smooth pressure drop constraint to obtain MINLP Fit cost as smooth function of diameter Use cost-fit for NLP relaxation of MINLP Non-convex MINLP solved by Bonmin’s Branch & Bound
Convergence to Global Optimum not guaranteed
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
8 / 32
Convex MINLP Formulation
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
9 / 32
Convex MINLP Formulation
Hydraulic Calculation as Convex Program
Conventional Formulation
Nonlinear Equations (NLE) (Pi + Hi ) − (Pj + Hj ) = sgn(Qij )φdij (Qij )Lij Pi = Pipump P P Qji − Qij = Qidem j:(j,i)∈E
∀(i, j) ∈ E ∀i ∈ N pump ∀i ∈ N \ N pump
j:(i,j)∈E
max ≤ Q ≤ Q max −Qij,d ij ij,dij ij
∀(i, j) ∈ E
Assumption 1
φd (·) is strictly monotonically increasing
2
Pressure drop inversely proportional to diameter φd (Q) < φd 0 (Q) ∀d > d 0 , Q > 0
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
10 / 32
Convex MINLP Formulation
Hydraulic Calculation as Convex Program
Convex Formulation Definition Φd (Q) =
RQ
φd (Q 0 )dQ 0
0
Variational Formulation (VF) [3] min
P (i,j)∈E
(Φdij (Qij+ ) + Φdij (Qij− ))Lij
−
P i∈N pump ,(i,j)∈E
s.t. πij − λ+ ij , λij − Λ+ ij , Λij
P
(Hipump + Pipump )(Qij+ − Qij− )
(Qji+ − Qji− ) −
j:(j,i)∈E Qij+ , Qij− Qij+ , Qij−
P j:(i,j)∈E
(Qij+ − Qij− ) = Qidem ∀i ∈ N \ N pump
≥ 0∀(i, j) ∈ E max ∀(i, j) ∈ E ≤ Qij,d ij
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
11 / 32
Convex MINLP Formulation
Hydraulic Calculation as Convex Program
Sketch of the Proof Theorem (Q, P) solves (NLE) iff (Q + , Q − ,π, λ,Λ) with Λ = 0. Proof. Only If Qij+ = max(0, Qij ); Qij− = min(0, Qij ); πi = Hi + Pi ; λ+ ij = max(0, (H + P = max(0, (Hj + Pj ) − (Hi + Pi )); λ− i i ) − (Hj + Pj )) ij Φd (·) is strictly convex =⇒ Solves (VF) If Solves (VF) =⇒ Qij+ > 0 or Qij− > 0 πj − πi +φdij (Qij+ )Lij − λ+ ij = 0 πi − πj +φdij (Qij− )Lij − λ− ij = 0 Choose, Pi = πi −Hi ; Qij = Qij+ − Qij− A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
12 / 32
Convex MINLP Formulation
Hydraulic Calculation as Convex Program
Sketch of the Proof Theorem (Q, P) solves (NLE) iff (Q + , Q − ,π, λ,Λ) with Λ = 0. Proof. Only If Qij+ = max(0, Qij ); Qij− = min(0, Qij ); πi = Hi + Pi ; λ+ ij = max(0, (H + P = max(0, (Hj + Pj ) − (Hi + Pi )); λ− i i ) − (Hj + Pj )) ij Φd (·) is strictly convex =⇒ Solves (VF) If Solves (VF) =⇒ Qij+ > 0 or Qij− > 0 πj − πi +φdij (Qij+ )Lij − λ+ ij = 0 πi − πj +φdij (Qij− )Lij − λ− ij = 0 Choose, Pi = πi −Hi ; Qij = Qij+ − Qij− A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
12 / 32
Convex MINLP Formulation
Hydraulic Calculation as Convex Program
Bound Multiplier as Flow Limiting Valve max (NLE) is ill-posed if Qij > Qij,d ij
(Hi + Pi ) − (Hj + Pj ) = φdij (Qij )Lij
Λ+ ij > 0 - Pressure loss in valve πi − πj = φdij (Qij+ )Lij + Λ+ ij Smooth model for flow limiters
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
13 / 32
Convex MINLP Formulation
Convex MINLP Formulation
Convex MINLP Formulation Additional Variables
Bounds on Flows
Direction variables: xijdir ∈ + − Flow variables: Qij,d , Qij,d
{0, 1}
+ max x dir Qij,d ≤ Qij,d ij − max (1 − x dir ) Qij,d ≤ Qij,d ij
Convexify the Pressure Drop Constraint (Hi + Pi ) − (Hj + Pj ) ≥ (Hj + Pj ) − (Hi + Pi ) ≥
nd P d=1 nd P d=1
A. U. Raghunathan (UTRC)
+ φd (Qij,d )Lij − ∆P max (1 − xijdir ) − φd (Qij,d )Lij − ∆P max xijdir
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
14 / 32
Linearizations-based MINLP Algorithm
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
15 / 32
Linearizations-based MINLP Algorithm
Sketch of MINLP Algorithm Root Node
Converges to Global Optimum Integer Infeasible Check for violation of constraints Add linearization of violated constraint
Integer Feasible
Integer Feasible
Solve (VF) for given diameters
Solve (VF) for given diameters
Feasible w.r.t. pressure requirement
Infeasible w.r.t. pressure requirement
Add linearization of violated constraint
Fathom node(s)
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
16 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Linearizations at Integer Infeasible Nodes Local: Relaxation of Pressure Drop (RelPD) Let (x , x dir , P, Q + , Q − ) be solution to node LP. For each edge (i, j) ∈ E, Pick direction based on xijdir If xijdir > 0.5, evaluate pressure drop constraint from i → j cij = −(Hi + Pi ) + (Hj + Pj ) +
nd P d=1
+ φd (Qij,d )Lij + ∆P max (1 − xijdir )
If xijdir ≤ 0.5, evaluate pressure drop constraint from j → i cij = (Hi + Pi ) − (Hj + Pj ) +
nd P d=1
− φd (Qij,d )Lij + ∆P max xijdir
If cij > , add linearization
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
17 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Linearizations at Integer Infeasible Nodes (contd.)
0
2
1
Global: Directed Cycles (DirCyc)
3
Let (x , x dir , P, Q + , Q − ) be solution to node LP. For each edge (i, j) ∈ E, If xijdir > 0.5, then flow from i → j If xijdir ≤ 0.5, then flow from j → i
4
A. U. Raghunathan (UTRC)
Check for directed cycle. If directed cycle exists, add constraint to invalidate the choice of edge directions.
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
18 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Linearizations at Integer Infeasible Nodes (contd.) Maximum Pressure Drop Infeasibility + (Pi + Hi ) − (Pj + Hj ) ≥ max φd (Qij,d )Lij − ∆P max (1 − xijdir ) d=1,...,nd
Linearization provides info on a single diameter only ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d −∆P max (1 − xijdir )
φd (Q)
ˆ + ) − ∇φd (Q ˆ + )Q + = φd (Q ij,d ij,d ij,d eq eq eq φd 0 (Qij,d 0 ) − ∇φd 0 (Qij,d 0 )Qij,d 0 eq ˆ+ Q Qij,d 0Q ij,d
A. U. Raghunathan (UTRC)
Global: Max. Pressure Drop (MaxPD) ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d eq + max +∇φd 0 (Qij,d (1 − xijdir ) 0 )Lij Qij,d 0 −∆P
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
19 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Linearizations at Integer Infeasible Nodes (contd.) Maximum Pressure Drop Infeasibility + (Pi + Hi ) − (Pj + Hj ) ≥ max φd (Qij,d )Lij − ∆P max (1 − xijdir ) d=1,...,nd
Linearization provides info on a single diameter only ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d −∆P max (1 − xijdir )
φd (Q)
ˆ + ) − ∇φd (Q ˆ + )Q + = φd (Q ij,d ij,d ij,d eq eq eq φd 0 (Qij,d 0 ) − ∇φd 0 (Qij,d 0 )Qij,d 0 eq ˆ+ Q Qij,d 0Q ij,d
A. U. Raghunathan (UTRC)
Global: Max. Pressure Drop (MaxPD) ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d eq + max +∇φd 0 (Qij,d (1 − xijdir ) 0 )Lij Qij,d 0 −∆P
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
19 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Linearizations at Integer Infeasible Nodes (contd.) Maximum Pressure Drop Infeasibility + (Pi + Hi ) − (Pj + Hj ) ≥ max φd (Qij,d )Lij − ∆P max (1 − xijdir ) d=1,...,nd
Linearization provides info on a single diameter only ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d −∆P max (1 − xijdir )
φd (Q)
ˆ + ) − ∇φd (Q ˆ + )Q + = φd (Q ij,d ij,d ij,d eq eq eq φd 0 (Qij,d 0 ) − ∇φd 0 (Qij,d 0 )Qij,d 0 eq ˆ+ Q Qij,d 0Q ij,d
A. U. Raghunathan (UTRC)
Global: Max. Pressure Drop (MaxPD) ˆ+ ) − (Hi + Pi ) − (Hj + Pj ) ≥ (φd (Q ij,d ˆ + )Q ˆ + )Lij + ∇φd (Q ˆ + )Lij Q + ∇φd (Q ij,d ij,d ij,d ij,d eq + max +∇φd 0 (Qij,d (1 − xijdir ) 0 )Lij Qij,d 0 −∆P
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
19 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
Cuts at Integer Feasible Nodes 0
Global: Infeasible Configurations (InfCfg) Solve (VF) to check for pressure satisfaction.
2
1
4
3
If not satisfied Add cut to invalidate this configuration. Invalidate configurations with same flow direction and smaller diamaters If satisfied Add cut to invalidate other flow directions in this configuration.
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
20 / 32
Linearizations-based MINLP Algorithm
Deriving Linearizations
MINLP Algorithm Data: nrml = 10−2 ; slwp = 10−6 ; N RelPD = 151; N MaxPD = 40 begin Start Branch-and-Bound algorithm. while Gapintegral > 0 do = (Gapintegral reduction over 100 nodes > 10−5 ) ? nrml : slwp for each node in tree do if Directed-Cycle then Add global cut (DirCyc) to invalidate choice of directions on edges Continue if integer infeasible and mod(N node , N MaxPD ) == 0 then Add global cut (MaxPD) if MaxDP-Infeas(i, j) > ∀(i, j) ∈ E if integer infeasible and mod(N node , N RelPD ) == 0 then Add local cut (RelPD) if RelDP-Infeas(i, j) > ∀(i, j) ∈ E if integer feasible then Solve the (VF) for given configuration. Add global cut to invalidate flow direction choices for given configuration if PressDrop-Infeas or MaxFlow-Active then Add global to invalidate current configuration if PressDrop-Infeas then Add global cut (InfCfg) to invalidate other configurations with same flow direction and smaller diameter sizes
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
21 / 32
Results
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
22 / 32
Results
Water Distribution Networks
Example Problems (Bragalli & others [6])
Problem name shamir hanoi new york blacksburg foss_poly_0 foss_iron foss_poly_1 pescara modena
A. U. Raghunathan (UTRC)
# nodes 8 33 21 32 38 38 38 74 276
# edges 8 34 21 25 58 58 58 99 317
Global Opt. of Sizing in Pipe Networks
# diameters 14 6 12 11 7 13 22 13 13
ICCOPT 2010
23 / 32
Results
Water Distribution Networks
Computational Results - Moderate Instances Implementation CPLEX using CPXcutcallback IPOPT - NLP solver Pentium Duo Core - 2.63 GHz, 3 GB RAM, 1 thread Automatic variable priority Integrality Gap =0.0 % Problem name shamir hanoi new york blacksburg
Optimal Objective 419, 000 6, 109, 620.09 3, 9307, 799.72 118, 251.2
A. U. Raghunathan (UTRC)
# NLPs 34 135 53 252
# cuts 203 2072 1783 810
Global Opt. of Sizing in Pipe Networks
# CPU time (sec) 3.7 822.4 258.3 113
ICCOPT 2010
24 / 32
Results
Water Distribution Networks
Computational Results - Large Instances
Difficulty in solving Unable to close gap after 2 hrs of computation. Unable to find feasible solution. Problem name foss_iron foss_poly_0 foss_poly_1 pescara modena
A. U. Raghunathan (UTRC)
Lower Bound 175, 423 6.74902 · 107 25, 799.4 1.63103 · 106 2.1108 · 106
Best Objective 221, 155 8.91541 · 107 −−− −−− −−−
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
25 / 32
Results
Water Distribution Networks
Extension to Layout & Sizing Optimization Choice of edge & diameter
nd P
xij,d ≤ 1
∀(i, j) ∈ E
d=1
Afshar and Jabbari [5] Solved using Genetic Algorithm Problem name # Nodes # Edges # Diameters Reported Optimal Obj. Optimal Obj. CPU time (sec) Gap(%)
A. U. Raghunathan (UTRC)
Network I 9 12 13 39400 38600 4140 4.15
Global Opt. of Sizing in Pipe Networks
Network II 20 37 13 1783086 — 7200 —
ICCOPT 2010
26 / 32
Summary
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
27 / 32
Summary
Summary
Hydraulic calculation can be cast as Convex Program Sizing optimization problem is Convex MINLP Tailored linearizations-based algorithm (a la FILMINT) Computationally efficient on moderate-size problems Outlook Large-scale problems Dynamic optimization problems Extension to electrical grid
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
28 / 32
Appendix
Outline 1
Introduction Motivation Problem Statement Previous Work
2
Convex MINLP Formulation Hydraulic Calculation as Convex Program Convex MINLP Formulation
3
Linearizations-based MINLP Algorithm Deriving Linearizations
4
Results Water Distribution Networks
5
Summary
6
Appendix A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
29 / 32
Appendix
References I E. C. Cherry. Some general theorems for non-linear systems possessing reactance. Philosophical Magazine, (Ser. 7)42:1161–1177, 1951. E. Alperovits and U.Shamir. Design of optimal water distribution systems. Water Resources Research 13(6):885–900, 1977. M. Collins, L. Cooper, R. Legason, J. Kennington and L. LeBlanc. Solving the pipe network analysis problem using optimization techniques. Management Science, 24(7):747–760, 1978. G. Eiger, U. Shamir and A. Ben-Tal. Optimal design of water distribution networks. Water Resources Research, 30(9):2637–2646, 1994. A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
30 / 32
Appendix
References II
M. H. Afshar and E. Jabbari. Simultaneous layout and pipe size optimization of pipe networks using genetic algorithm. The Arabian Journal for Science and Engineering, 33(2B):391–409, 2008. C. Bragalli, C. D’Ambrosio, J. Lee, A. Lodi and P. Toth. Water Network Design by MINLP, IBM Research Report RC24495, 2008.
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
31 / 32
Appendix
Computational Results - Large Instances II Λ+ ij > 0 - Pressure loss in valve πi − πj = φdij (Qij+ )Lij + Λ+ ij Assume flow bounds are imposed through flow limiters
Problem name foss_iron foss_poly_0 foss_poly_1
Best Objective 175, 922 6.95878 · 107 28, 569.4
Gap (%) 0.0 4.15 10.7
CPU time (sec) 166 10800 10800
# Flow Infeas. 4 5 15
Modeling of flow bounds
A. U. Raghunathan (UTRC)
Global Opt. of Sizing in Pipe Networks
ICCOPT 2010
32 / 32