decomposition method for the multiperiod blending...
TRANSCRIPT
Decomposition method for the Multiperiod Blending Problem
Irene Lotero, Francisco Trespalacios and Ignacio Grossmann September 17-18, 2014
Center for Advanced Process Decision-making Department of Chemical Engineering Carnegie Mellon University
Pittsburgh, PA 15213
2
Outline
• Problem description • Applications • Mathematical Formulation
Background
• State-of-the-art commercial MINLP solvers struggle • Generate “good” solutions fast Motivation
Approach • Alternative mathematical formulation based on GDP • Decomposition algorithm
Supply Tanks (s) Blending Tanks (b) Demand Tanks (d)
Multiperiod blending problem is defined over supply, blending and demand tanks
Assumptions: • Supply concentrations are constant • Perfect mixing • No simultaneous input/output to blending tanks
Given: • Topology, initial conditions and flow profit/costs • Supply tank flow and concentration • Demand tank flow and concentration limits
Determine • Flows between which tanks in which time periods • Inventories/concentrations for tanks in each period • Maximum total profit of blending operation
Fs1,t Cs1
Fs3,t Cs3
Fs2,t Cs2
Fd1,t CL
d1-CUd1
Fd2,t CL
d2-CUd2
Fd3,t CL
d3-CUd3
Fs,b,t
Fb,d,t
Ib,t
Is,t
Id,t
Fs,d,t
3
ys,b,t
yb,d,t
Cq,s,t
General model for many applications
4
• Gasoline and crude oil blending • Raw material feed scheduling • Storage of intermediate streams • Water treatment • Emissions regulation
Supply Tanks (s) Blending Tanks (b) Demand Tanks (d)
Fs1,t Cs1
Fs3,t Cs3
Fs2,t Cs2
Fd1,t CL
d1-CUd1
Fd2,t CL
d2-CUd2
Fd3,t CL
d3-CUd3
Fs,b,t
Fb,d,t
Ib,t
Is,t
Id,t
Fs,d,t
Applications Nomenclature
ys,b,t
yb,d,t
Cq,s,t
MINLP formulation contains bilinear terms
[ [ ] ] ⋁
Type
max profit - network flow cots
linear
mixed-integer linear
mixed-integer linear
linear
nonlinear (nonconvex)
integer linear
linear
s.t for flows into blending tanks: [flow within bounds] ⋁ [flow = 0] for flows into demand tanks: flow within bounds flow = 0 concentrations within demand spec. “no bounds” on concentration
total inventory mass balance in tanks inventory mass balance by component in blending tanks no simultaneous in/out flow variable bounds
s.t
Contains bilinear terms, for each contaminant, each blending tank
and each time period
5
6
Outline
• Problem description • Applications • Mathematical Formulation
Background
• State-of-the-art commercial MINLP solvers struggle • Generate “good” solutions fast Motivation
Approach • Alternative mathematical formulation based on GDP • Decomposition algorithm
Commercial solvers not able to find feasible solutions
Commercial Solvers Performance Objectives
Increase size of the problems • Current problems: 4 time periods 8 qualities 4 blending tanks • Target problems: 12 time periods 10 qualities 16 blending tanks
• After 5 minutes of computational time,
Average normalized best feasible solution
Optimal 1.0
GloMIQO 0.57
BARON 0.19
SCIP 0.35
• After 30 minutes of computational time,
Found Feasible Solution Large Instances
GloMIQO 54%
BARON 0%
SCIP 15%
• 36 randomly generated instances
7
Generate “good solutions” fast • Guaranteeing global optimality not a priority • Obtain good feasible solutions
8
Outline
• Problem description • Applications • Mathematical Formulation
Background
• State-of-the-art commercial MINLP solvers struggle • Generate “good” solutions fast Motivation
Approach • Alternative mathematical formulation based on GDP • Decomposition algorithm
1. Raman R. and Grossmann I.E., “Modelling and Computational Techniques for Logic-Based Integer Programming”, Computers and Chemical Engineering, 18, 563, 1994.
GDP is a higher level of representation for MILP/MINLP Optimization problem with algebraic expressions, disjunctions & logic propositions
General form of GDP1 Illustration – Process network
9
Objective Function
Global Constraints
Disjunctions
Logic Propositions
Objective function
Global Constraints
Logic
R1
R2 S2
F1
F3
F2
F4
F6
F7
F5
S1
Disjunctions
1. The new binary variable can in fact be continuous due to the problem structure. Furthermore, using it as binary variable and giving it priority over the network variables can increase the solution time of a branch-and-bound method
Remark: Possible to reduce bilinear terms using a disjunction
10
If some of the tanks are fixed as “split tanks”, the resulting MINLP becomes easier to solve
Blending tank Split tank
⋁
Requires 0-1 variable, but it does not increase combinatorial
complexity (it might even reduce it!)1
Blend + Split tank
Tank
Mass balance
Comp. mass
balance Fewer or none bilinear terms
Original MINLP formulation Type
linear
mixed-integer linear
mixed-integer linear
linear
nonlinear (nonconvex)
integer linear
linear
s.t
11
Alternative MINLP formulation based on GDP Type
linear
mixed-integer linear
mixed-integer linear
linear
nonlinear (nonconvex)
integer linear
linear
s.t
12
Implies decomposition
Decomposition algorithm seeks to find feasible solutions in short periods of time
13
Algorithm Decisions in the algorithm
Master Problem (MILP) Provides Upper Bound (UB)
Add
cut
s: o
ptim
ality
or f
easi
bilit
y
Stop
No
Yes
Master MILP (MINLP relaxation) • Multiparametric disaggregation[1]
Subproblem (MINLP) Provides Lower Bound (LB)
Fix split tanks
If feasible, it provides feasible solution for the original MINLP
[2] Kolodziej, S.P., Castro, P.M., Grossmann, I.E. Global optimization of bilinear programs with a multiparametric disaggregation technique. Journal of Global Optimization (2013)
Solving MINLP • Smaller nonconvex MINLP • Global solution through specialized technique,
e.g. multiparametric disaggregation
Other considerations • “no-good” are included in the subproblem,
eliminating regions already evaluated in previous subproblems
Multiparametric disaggregation provides a lower bound
Discretization Technique Notes and examples
Multiperiod blending problem
We can use other bases for more/fewer binary variables
• Base two takes fewer binary variables
Precision
Range 0 – 15 0 – 9
0 – 12.7 0 – 9.9
0 – 10.23 0 – 9.99
Increment 1 0.1 0.01
Binary Variables
Base 2 8 14 20
Base 10 10 20 30
Binary variables
Scales up better If approx. MILP is feasible,
provides solution for original MINLP 14
Slack variables are introduced for upper bound
Discretization Technique + Slack Notes
Is a relaxation problem because it includes at least the entire feasible region of the problem
The relaxed MILP provides a valid upper bound for original MINLP
15
Subproblem algorithm solves the discretized and relaxed MILPs successively increasing the precision
16
Decomposition Algorithm
Master Problem (MILP) Provides Upper Bound (UB)
Add
cut
s: o
ptim
ality
or f
easi
bilit
y
Stop?
Stop
No
Yes
Subproblem (MINLP) Provides Lower Bound (LB)
Fix split tanks
Subproblem Algorithm[2]
[2] Kolodziej, S.P., Grossmann, I.E., Furman, K.C. and Sawaya, N.W. A discretization-based approach for the optimization of the multiperiod blend scheduling problem, Computers & Chemical Engineering, (2013).
GLOMIQO BARON SCIP
• 18 randomly generated instances • 3 – 4 time periods, 2 – 10 qualities, 4 – 12 blending tanks
0 – 300 s
Results Small Instances
17
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 60 120 180 240 300
Aver
age
norm
aliz
ed
best
feas
ible
sol
utio
n
TIME (s)
GLOMIQO BARON SCIP DECOMPOSITION
• 18 randomly generated instances • 3 – 4 time periods, 2 – 10 qualities, 4 – 12 blending tanks
0 – 1800 s 0 – 300 s
Results Small Instances
17
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 60 120 180 240 300
Aver
age
norm
aliz
ed
best
feas
ible
sol
utio
n
TIME (s)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 600 1200 1800
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
GLOMIQO BARON SCIP DECOMPOSITION
• 18 randomly generated instances • 3 – 4 time periods, 2 – 10 qualities, 4 – 12 blending tanks
0 – 1800 s 0 – 300 s
Results Small Instances
17
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 60 120 180 240 300
Aver
age
norm
aliz
ed
best
feas
ible
sol
utio
n
TIME (s)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 600 1200 1800
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
• 18 randomly generated instances • 4 – 12 time periods, 6 – 10 qualities, 4 – 16 blending tanks
0 – 300 s
Results Large Instances
18
GLOMIQO BARON SCIP DECOMPOSITION
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 50 100 150 200 250 300
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
• 18 randomly generated instances • 4 – 12 time periods, 6 – 10 qualities, 4 – 16 blending tanks
0 – 1800 s 0 – 300 s
Results Large Instances
18
GLOMIQO BARON SCIP DECOMPOSITION
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 50 100 150 200 250 300
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 600 1200 1800
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
• 18 randomly generated instances • 4 – 12 time periods, 6 – 10 qualities, 4 – 16 blending tanks
0 – 1800 s 0 – 300 s
Results Large Instances
18
GLOMIQO BARON SCIP DECOMPOSITION
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 50 100 150 200 250 300
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1.0
0 600 1200 1800
Ave
rage
nor
mal
ized
be
st fe
asib
le s
olut
ion
TIME (s)
After 10 minutes the average normalized best feasible solution is twice the average normalized
best feasible solution of the following best commercial solver
• 36 randomly generated instances • 3 – 12 time periods, 2 – 10 qualities, 4 – 16 blending tanks
0 – 300 s
Results All Instances
19
GLOMIQO BARON SCIP
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 50 100 150 200 250 300
Aver
age
norm
alize
d be
st fe
asib
le so
lutio
n
TIME (s)
• 36 randomly generated instances • 3 – 12 time periods, 2 – 10 qualities, 4 – 16 blending tanks
0 – 1800 s 0 – 300 s
Results All Instances
19
GLOMIQO BARON SCIP DECOMPOSITION
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 50 100 150 200 250 300
Aver
age
norm
alize
d be
st fe
asib
le so
lutio
n
TIME (s)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 600 1200 1800
Aver
age
norm
alize
d be
st fe
asib
le so
lutio
n
TIME (s)
• 36 randomly generated instances • 3 – 12 time periods, 2 – 10 qualities, 4 – 16 blending tanks
0 – 1800 s 0 – 300 s
Results All Instances
19
GLOMIQO BARON SCIP DECOMPOSITION
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 50 100 150 200 250 300
Aver
age
norm
alize
d be
st fe
asib
le so
lutio
n
TIME (s)
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 600 1200 1800
Aver
age
norm
alize
d be
st fe
asib
le so
lutio
n
TIME (s)
After 5 minutes the average normalized best feasible solution is >0.7
The algorithm performs better than state-of-the-art MINLP commercial solvers
Good MILP approximation and decomposition allow finding good solutions in first iteration
20
Instance
1. Normalized to best known solution
Incr
ease
siz
e
Group 1
Group 2
Group 3
Group 4
Group 5
Group 6
Instance
0-1 vars
Bilinear terms
114
235
539
988
1524
2736
195
428
1071
1335
1940
3200
First MINLP
0-1 vars
Bilinear terms
30
63
152
223
305
494
72
132
516
580
446
1089
UB
First iteration (normalized1)
LB Time (s)
1.07
1.04
1.10
1.06
1.06
1.08
0.94
0.62
0.69
0.80
0.50
0.48
14.5
60.9
95.1
79.7
95.4
213.7
70%-80% fewer 0-1 variables
~55% fewer bilinear terms
~65% fewer bilinear terms
~75% fewer bilinear terms
Very good solution after 1st iteration
Good solution after 1st iteration
Not as good but at least feasible
solution found after 1st iteration
Feasible solution found in less than
a minute
Feasible solution found in less than a
two minutes
Almost 5 minutes for large instances
To wrap up
21
Propose an alternative formulation based on GDP • Blending/Splitting tanks
Decomposition algorithm that simplifies the search for feasible solutions • Solving smaller MINLPs with fewer 0-1 variables and bilinear terms • “Guided” by an MILP relaxation of the problem
Tested the approach in randomly generated instances • Increase problem sizes to match industrial applications
Generate “good” solutions fast
Thank you
Dimitri J. Papageorgiou and Myun-Seok Cheon from ExxonMobil
Acknowledgments
22