Click to edit Master title styleAPOPT and BPOPT:
New NLP and MINLP Solvers for Differential Algebraic Systems
with Benchmark Testing
John D. HedengrenJose L. Mojica
Brigham Young University
Kody M. PowellWesley Cole
Thomas F. EdgarThe University of Texas at Austin
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Presentation Overview
Overview of Benchmarking Testing. NLP Benchmarking
Hock-Schittkowski Dynamic optimization Biological models
MINLP Benchmarking MacMINLP
Case studies Gravity Drained tank (MIDAE control problem) Energy Storage for smart grid
Future Developments
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Overview of Benchmark Testing
NLP Benchmark Testing APOPT1, BPOPT1, IPOPT2, SNOPT3, MINOS3
Problem characteristics: Hock Schittkowski, Dynamic Opt, SBML Nonlinear Programming (NLP) Differential Algebraic Equations (DAEs) APMonitor Modeling Language
MINLP Benchmark Testing APOPT1, BPOPT1, BONMIN2, DICOPT4
Problem characteristics: MacMINLP, Industrial Test Set Mixed Integer Nonlinear Programming (MINLP) Mixed Integer Differential Algebraic Equations (MIDAEs) APMonitor & AMPL Modeling Language
mn uyxuyxhuyxg
uyxtxfts
uyxJ
,),,(0),,(0
,,,0..
),,(min
1–APS, LLC 2–EPL, 3–SBS, Inc., 4-CMU
mmn zuyxzuyxhzuyxg
zuyxtxfts
zuyxJ
,),,,(0),,,(0
,,,,0..
),,,(min
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
NLP Benchmark – Hock-Schittkowski (116)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
10
20
30
40
50
60
70
80
90
100
Not worse than 2 times slower than the best solver ()
Per
cent
age
(%)
APOPT+BPOPTAPOPT1.0
BPOPT1.0
IPOPT3.10
IPOPT2.3
SNOPT6.1MINOS5.5
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
NLP Benchmark - Dynamic Optimization (37)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
10
20
30
40
50
60
70
80
90
100
Not worse than 2 times slower than the best solver ()
Per
cent
age
(%)
APOPT+BPOPTAPOPT1.0
BPOPT1.0
IPOPT3.10
IPOPT2.3
SNOPT6.1MINOS5.5
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
NLP Benchmarking – SBML (341)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 510
20
30
40
50
60
70
80
90
100
Not worse than 2 times slower than the best solver ()
Per
cent
age
(%)
APOPT+BPOPTAPOPT1.0
BPOPT1.0
IPOPT3.10
IPOPT2.3
SNOPT6.1MINOS5.5
0
20
40
60
80
100
120
# of M
odels
# of Physical Entities
Click to edit Master title styleExample: ErbB Signal Transduction Cascade
Large ErbB signallingmodel (~504 physical entities)*
Parameter estimation (simulated annealing) took “24 hours on a 100-node cluster computer”
*Chen et al. Mol Syst Biol. 2009;5:239 .
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
NLP Benchmark – Summary (494)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 510
20
30
40
50
60
70
80
90
100
Not worse than 2 times slower than the best solver ()
Per
cent
age
(%)
APOPT+BPOPTAPOPT1.0
BPOPT1.0IPOPT3.10
IPOPT2.3
SNOPT6.1
MINOS5.5
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
NLP Benchmark – Combo Solvers (494)
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 510
20
30
40
50
60
70
80
90
100
Not worse than 2 times slower than the best solver ()
Per
cent
age
(%)
APOPT+BPOPTAPOPT+IPOPT3
APOPT+IPOPT2
APOPT+SNOPTAPOPT+MINOSBPOPT+IPOPT3
BPOPT+IPOPT2
BPOPT+SNOPTBPOPT+MINOSIPOPT3+IPOPT2
IPOPT3+SNOPT
IPOPT3+MINOS
IPOPT2+SNOPT
IPOPT2+MINOS
SNOPT+MINOSAPOPT1.0BPOPT1.0
IPOPT3.10
IPOPT2.3
SNOPT6.1MINOS5.5
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
MacMINLP Benchmark Summary Results
Variables Integers Equations NLP Obj MINLP Obj Solution Time Status10 4 4 1.55E+02 1.74E+02 0.031 Success an_integer_test107 24 73 2.59E+05 2.86E+05 0.25 Success batch4 4 0 6.30E‐13 7.78E‐07 0.016 Success geartrain127 16 7 5.02E+00 1.30E+01 0.094 Success mittelman12 3 6 7.59E‐01 6.01E+00 0.031 Success synthes123 5 14 4.18E+00 7.30E+01 0.047 Success synthes238 8 24 1.65E+01 7.57E+01 0.062 Success synthes36 3 5 1.18E‐01 2.54E‐01 0.031 Success wind‐fac
510 168 308 ‐1.01E+00 ‐1.01E+00 8.865 Success c‐reload‐14a510 168 308 ‐1.02E+00 ‐1.02E+00 13.728 Success c‐reload‐14b510 168 308 ‐9.91E‐01 ‐9.56E‐01 48.344 Failure c‐reload‐14c510 168 308 ‐1.03E+00 ‐1.03E+00 6.724 Success c‐reload‐14d510 168 308 ‐1.03E+00 ‐1.03E+00 4.852 Success c‐reload‐14e510 168 308 ‐1.01E+00 ‐1.01E+00 6.614 Success c‐reload‐14f
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Case Study- MPC with MINLP
PCT40 is a basic process control unit produced by Armfield Main Devices: Large Process Vessel, CSTR, Hot water Tank Sensors: temperature, pressure, and level sensors Devices: valves, heating coils, and pumps
How does this work for MINLP Problems?Objective: Keep the tank level at 100 mm
Continuous: Proportional Solenoid Valve (PSV)Discrete: On/Off Valve (SOL1)
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
MINLP Model Predictive Control
0 50 100 150 200 250 300 350 400 450 5000
50
100
150
200
250
Hei
ght (
mm
)
Target HeightActual Height
0 50 100 150 200 250 300 350 400 450 50010
-2
10-1
100
101
Cycle
CP
U T
ime
(sec
)
Total TimeSolver Time
MIDAE SystemVariables: 576Integers: 16Equations: 544
FeaturesWarm Start Solutions
Discrete (16) and Continuous (16) Decisions
CPU Increases with Setpoint Changes
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 50
100
200
300
400
500
CPU Time (sec)
Num
ber o
f Cyc
les
Total CPU Time
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Results –set point changes
Set Point ChangeSet Point Change
Set Point Changes Large (50) – ON Medium(20) – ON Small (5) – ON Very Small – OFF
Needs another constraint to limit SOL1 use.
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Results-disturbances
Disturbance – Opened up an exit valve which is not part of the model
Reduced Solenoid fluctuations, but still a problem, add cost to opening SOL 1
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Smart Grid Energy Systems
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Thermal Energy Storage (TES)
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Chiller-TES System
40°F 53°F
85°F
95°F
53°F
MINLP SystemVariables: 672Integers: 24Equations: 576
FeaturesCold Start Solution
Discrete (24) and Continuous (24) Decisions
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Problem Formulation (MINLP)
,0( ) ( )ft
elec chiller ii
J c t P t dt
,
, ,
,
, is nonlinear
0 80002000 2000
175 13120,1
i chiller i buildingi
chiller i i chiller i
TES
TES
chiller i
i
y L L TES
P y f L WBT f
dE TESdt
ETESL
y
Subject to:
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Problem Formulation (Relaxed)
24
, 1, 2,1
min elec i chiller i chiller ii
J c P P
1, 2, ,
1, 1,
2, 2,
, , 1
,
, is nonlinear
,
0 80002000 2000
chiller i chiller i building i i
chiller i chiller i i
chiller i chiller i i
TES i TES i i
TES i
i
L L L TES
P f L WBT f
P f L WBT
E E TESETES
Subject to:
Solved using SQP algorithm
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Relaxing Binary Constraints
Mixed integer problem doubles # variables
Branch and bound technique significantly increases computation time
Continuous formulation?
20
0
500
1000
1500
2000
2500
3000
3500
4000
0 1000 2000 3000 4000 5000 6000 7000
Pow
er C
onsu
med
(kW
)
Load (tons)
0
1
2
3
4
5
0 1000 2000 3000 4000 5000 6000 7000
Pow
er
Con
sum
ed/L
oad
(kW
/ton)
Load (tons)
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Results
BONMIN APOPT Excel MATLAB
Cost ($/day) $ 807.53 $ 804.65 $ 801.93* $ 802.13*
CPU Time (sec) 11.6 45.0 58.4* 0.1*
-2500-2000-1500-1000-500
0500
1000150020002500
0 5 10 15 20 25
TE
S (t
ons)
Time (hours)
Excel SolutionAPM Solution
* = Relaxed Formulation (NLP)
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Optimization of investment decision in energy systems
Can we apply a similar MPC with MINLP approach to optimize investment decisions for next generation energy systems?
APMonitor.com | APOPT.com BYU – UT AustinOct 17, 2012
Future Development Most flexible and powerful platform for DAEs
Need collaborators to create a new interfaces.
Development influenced by multiple domains: Computational Biology Aerospace Engineering (UAVs) Petrochemical Industry Smart Grid Modeling and Optimization
Solver Development Mixed Integer Nonlinear Programming for DAEs (MIDAEs) Combine strengths of Active Set (APOPT) and Interior Point
(BPOPT) methods