Survey of Industrial Applications of �Embedded Model Predictive Control
Alexander Domahidi �Collaborators: Joachim Ferreau & Stefan Almér (ABB),�Juan Jerez (embotech), Tobias Gybel Hovgaard (Vestas)
European Control Conference�Aalborg, Denmark�June 29, 2016
Is embedded optimization just a bubble?▶ Is it used at all in industry?▶ In which applications?▶ Why? How much better is it?▶ Which type of problems are important?▶ How fast is “fast”?▶ What are the main challenges?
▶ Carried out using SurveyMonkey.com▶ Addressed ~1000 people + CSS E-letter 333▶ 160 responses, 134 complete
2
Approach: Ask people in an online survey
en.wikipedia.org/wiki/Double_bubble_conjecture
Outline▶ Results of the first survey on �
embedded optimization
▶ Application examples
• Compressor drive control (ABB)�by courtesy of Thomas Besselmann, �Stefan Almér and Joachim Ferreau
• Wind turbine control (Vestas)�by courtesy of Tobias Gybel Hovgaard
3
Applications (158)
Automotive 25%
Energy 19%
Robotics 18%Aerospace 11%
Power Electronics
8%
Health 3%
Manufacturing tools 4%
Chemical, Oil, Gas 3%
Computing 1%
Food processing 2%
Pricing & Marketing
1% Other 5%
4
Technological readiness (158)
5
Simulation study 24%
Academic prototype
40%
Industrial prototype
21%
Product beta 4%
Product 10%
Other 1%
Automotive43%
Energy9%
Robotics12%
Aerospace9%
Medical9%
Other18%
Next-gen automotive applications likely to use optimization
Reason for using embedded optimization (159)
6
System performance
53%
Novel features 30%
Development time 9%
Marketing appeal 5%
Competitor has it 3%
>100%4
51-100%10
31-50%6
21-30%9
11-20%6
6-10%11
0-5%3
>50%1
26-50%7
10-25%0
0-9%1
Quantifiable performance improvement (49)
Quantifiable reduction �of development time (9)
>10% improvement in KPIs common
Problem Types & Sampling Times (138)
7
02468
10
<10µs 10µs - 999µs
1ms - 9ms
10ms - 99ms
100ms - 999ms
1s - 10s >10s
Convex Problems (27%)
02468
10
<10µs 10µs - 999µs
1ms - 9ms
10ms - 99ms
100ms - 999ms
1s - 10s >10s
NLPs (30%)
0
2
4
6
8
<10µs 10µs - 999µs
1ms - 9ms
10ms - 99ms
100ms - 999ms
1s - 10s >10s
MI-LP/QPs (27%)
0123456
<10µs 10µs - 999µs
1ms - 9ms
10ms - 99ms
100ms - 999ms
1s - 10s >10s
MI-NLPs (16%)
otherIndustrial prototype or product
3 out of 4 applications solve non-convex problems
Major Challenges (125)
Problem formulation
34%
Software configuration
23%
Deployment on hardware
22%
Convincing the customer 8%
Convincing the management
6%
Training of engineers 5%
Other 2%
8
Half of problems could be solved by improved software
«Kollsnes accounts for more than 40% of all Norwegian gas deliveries» (Gassco)
© ABB Group Slide 9
Illus
tratio
n: S
tato
il
«Kollsnes accounts for more than 40% of all Norwegian gas deliveries» (Gassco)
Embedded MPC!
© ABB Group Slide 10
NMPC for Load Commutated Inverters Controlling 48 MW at 1 kHz sampling rate
June 28, 2016 Slide 11 © ABB Group
Load commutated inverters (LCIs) play an important role in powering electrically-driven compressor stations
Goal: Enable LCIs to ride through partial loss of grid voltage
Solution: § Auto-generated NMPC algorithm
(ACADO/qpOASES) § Running at 1 kHz on AC 800PEC
Results: § Successfully tested on a 48 MW
pilot plant installation § Works where PID solution fails to
satisfy the constraints
see Besselmann, Van de moortel, Almer, Jörg, Ferreau (2016)
MPC PID
Violates the 1.35 pu current limit
USD 7M* / day
In Kollsnes, the 2 MPC-controlled compressor drives deliver natural gas to Europe worth
Being fully operational after emergency shutdown may take up to half a day
*EU average gas price from June 28, 2016
MPC can increase system robustness
Case: Wind turbine control
Tobias Gybel Hovgaard, Control Specialist, PhD
Vestas: The global leader in wind technology Innovating to lower the cost of energy
• Profitably bringing market-driven, innovative solutions to our customers.
• Custom configurations based on modularised building blocks.
• Broad and flexible product portfolio to precisely meet the unique needs of every site.
• Collaboration with external partners to develop innovative solutions and integrate external technologies in new ways.
Basic concept Wind turbine production controller
Wind speed
Wind direction
Generatorspeed
Power
Controller
Pitch
Controller
Pitchreference
Powerreference
ElectricalConverter
Generator
Gearbox
Pitchangle
Yaw Controller
Production Controller(Main controller)
• Power extracted from the wind:
• Basic objectives: ᅳ Keep speed and pitch optimal for
maximum power extraction until point of mechanical/electrical saturation.
ᅳ Keep speed and power at rated levels for wind speeds above point of mechanical/electrical saturation.
ᅳ Mitigate structural loads (fatigue and extreme)
Basic concept Wind turbine production controller
),(½ 3 λθρ pCAvP =
Model Predictive Control (MPC) From How to What – in an Optimal Way
Generatorspeed
Windspeed
Toweracceleration
Pitchanglereference
Converterpowerreference
S
×
k1
k2
S
Generatorspeed
Windspeed
Toweracceleration
Pitchanglereference
Converterpowerreference
Large number of tuneable parameters
Generatorspeed
Windspeed
Toweracceleration
Pitchanglereference
Converterpowerreference
OptimizerModel
dx=f(x(t),u(t))dt
Weights in a cost function, directly targeting e.g. tower loads, power production, or pitch activity
Cost function and tuning
Ideally: • The controller solves a problem like the following:
maximize ( Power – λ1 Fatigue – λ2 Noise – λ3 Pitch rate – … ) subject to: System dynamics and constraints (equalities and inequalities)
Tuneables
e.g: over-speed maximum torque extreme loads
Achievements Embedded optimization successfully utilized in Vestas turbine software
• Model predictive control and numerical
optimization embedded in turbine software release package.
• Custom, code-generated solver from FORCES Pro.
• Operating flawlessly on turbines in the field.
• +5000h of safe operation
• A step-change in control technology with proven complexity reductions.
• Proven field performance.
+208d of operation
Failure probability 3x lower than winning the 6/49 lottery
* both during aggressive tuning, the solver ran into maximum number of iterations
100M solver calls
2 failures*
Results Performance as expected
Wind speed
Tow
er fa
tigue
load
s
• Great performance measured on power production as well as on actuator activity.
• Significant potential for load reductions (site/turbine specific, depending on tuning)
• Solver reliability close to 100 % (real-time requirement, feasibility, etc.)
Example field data: tower fatigue load
Conclusions▶ Embedded optimization
• is a technology (99.999998% reliability)• successfully used in a number of fields (automotive, robotics, energy)
▶ Challenges:• Problem formulation – tools, languages, examples etc.?
- Convex vs non-convex: how to approximate the problem well?• Not all solvers/methods/implementations are equally reliable• Technical integration: what to optimize, what to leave out? Interfaces?• Research: handle even more complex problems
- LTL specifications- large MI-NLPs etc.
22
Purpose of optimization (156)
Predictive Control 76%
Control (other) 8%
Estimation 9%
Scheduling/Planning 4%
Other 3%
23
3 out of 4 applications use MPC
June 28, 2016
NMPC for Load Commutated Inverters From concept to product
§ Kollsnes has a capacity of 143,000,000 cubic meters (3.8×1010 US gal) of natural gas per day.
§ Two out of six 41.2 MW compressor strings for gas export are now powered by MPC-controlled LCIs.
§ Kårstø is Europe's biggest export port for natural gas liquids and the third largest in the world.
§ Three 7.5 MW booster compressors are now powered by MPC-controlled LCI.
§ First successful ride-through (2015-11-29)
Aug 2015: Kollsnes gas processing plant, Norway
Sept 2015: Kårstø gas processing plant, Norway
see Besselmann, Jörg, Knutsen, Lunde, Stava, Van de moortel (2016)