Rui HuangResearch Engineer

United technologies Research Centerhuangr@utrc.utc.com

860-610-7655

Challenges and Opportunities foroptimization-based workflow in

Industry

IMA Workshop Research Collaboration Workshop: Optimization and UncertaintyQuantification in Energy and Industrial Applications

University of MinnesotaFeb 23th 2016

Team and Acknowledgement

2

Johan Åkesson ModelonMiroslav Barić United Technologies Research Center (former)Lorenz Biegler Carnegie Mellon UniversityJohn Burns Virginia TechJohn Cassidy United Technologies CorporationNai-Yuang Chiang United Technologies Research CenterEugene Cliff Virginia TechClas Jacobson United Technologies CorporationCarl Laird Purdue University/ModelonManfred Morari ETH ZurichSlaven Peleš United Technologies Research Center (former)Drăguna Vrabie United Technologies Research Center (former)Victor Zavala University of Wisconsin

Key Points

3

Energy efficient systems (situation)Market conditions cost, labeling and regulations require more energy efficientand environmental-friendly buildings. They pose challenges and opportunitiesfor new workflows for system design, operation and etc.

Takeaway:• System level engineering (not limited to component level) DOES benefit• Optimization-based approaches provide additional capabilities• Uncertainty analysis and fault handling need more work• Mobile computing platform is the desired state

Open Problems:• Modeling issues: industry-accepted, system compatibility• Diagnostic for users: workflows for engineers to identify modeling issues• Computational issues: solving optimization problems online• Mobile computing platform: customizing optimization formulation with limited

computational resources

Agenda

4

Context and problem definitionWhy is it important to revisit the design and operation?How do we define system engineering?

System level design and operation workflows based on optimizationProviding diagnostic information for users.

Uncertainty analysis and fault operationPresenting the state-of-the-art and challenges

Optimization on the Embedded PlatformProviding flexibility for optimization-based applications

Summaries and Future Work

5

Loads

Electrical

InformationManagement

Distribution

Weather

Heating,Ventilation,Air Conditioning

Lighting Lights &Fixtures

EnvelopeStructure

BuildingInsulation

Building Operating Conditions

Safety &Security

Information Thermal Power

The System Level View for Buildings

Building ManagementSystem

Cost Utilities

Thermostat

MotionSensors

IT Network

BuildingGeometry

Grid

On-Site Gen

Distribution(Fans, Pumps)

Heating & ACEquipment

Other LoadsOffice

Equipment

Facilityaccess

Fire / SmokeDetection and Alarm

Video

Interface that is exploited

Water Heating

Building Energy Challenges

6

Buildings consume• 39% of total U.S. energy• 71% of U.S. electricity• 54% of U.S. natural gas

Buildings produce 48% of U.S. Carbon emissionsAnnual energy costs for U.S. commercial buildings & industrial facilities: $400 BPortion of energy in buildings used inefficiently or unnecessarily: 30%The only energy end-use sector showing growth in energy intensity

• 17% growth 1985 – 2000• 1.7% growth projected through 2025

European Union Requirements:Buildings:

1. From 2019 all new buildings produce as much energy as they consume2. Member states set minimum targets for zero-energy buildings in 2020.

Residential1. After 2018 must generate as much as consumed via solar, heat pump and

conservation2. Member states set energy targets for existing buildings by 2015.

The energy challenges need innovative design and operation

The Implications

7

A global leader in building systems and aerospace industries

UTC is the largestbuilding supplier anddelivers the smart,sustainable solutionsthe world needs.

Achieving Persistent OptimalityAchieving superior performance across life cycle

Reduce

demand/loads

Efficient components,systems and sub-

systems

Monitorand optimize operation

Optimized Design

Demand

Selection

DemandMinimization –

System/EquipmentSelection

SystemOperation

Match demand withsupply

(microgrid/smartgrid/storage)

Demand andDemand andSupply

Integration

Diagnostics and continuous monitoring and commissioning

Optimized operation

Audits and Retrofits

Op

era

tio

ns

8

Successful Examples

Energy Retrofit10-30% Reduction

Very Low Energy>50% Reduction

LEED Design

20-50% Reduction

Tulane Lavin BernieNew Orleans LA150K ft2, 150 kWhr/m2

1513 HDD, 6910 CDDPorous Radiant Ceiling, Humidity ControlZoning, Efficient Lighting, Shading

Cityfront SheratonChicago IL1.2M ft2, 300 kWhr/m2

5753 HDD, 3391 CDDVS chiller, VFD fans, VFD pumpsCondensing boilers & DHW

Deutsche PostBonn Germany1M ft2, 75 kWhr/m2

6331 HDD, 1820 CDDNo fans or DuctsSlab coolingFaçade preheatNight cool

• Different types ofequipment for spaceconditioning &ventilation

• Increasing designintegration ofsubsystems & control

Highly efficient buildings exist

9

SYSTEM ENGINEERING

System engineering is a methodology for product systemlevel design, optimization and verification that:

Provides guarantees of performance and reliabilityagainst customer requirements

Produces modular, extensible architectures for products

Exploits model-based analytical tools and techniques

Coordinated execution of a prescriptive, repeatable andmeasurable design process (engineering standardworkflows)

Definition & scope

10

Agenda

11

Context and problem definitionWhy is it important to revisit the design and operation?How do we define system engineering?

System level design and operation workflows based on optimizationProviding diagnostic information for users.

Uncertainty analysis and fault operationPresenting the state-of-the-art and challenges

Optimization on the Embedded PlatformProviding flexibility for optimization-based applications

Summaries and Future Work

Engineering Standard Work

12

Definition & case study

Pratt & Whitney: Engineering Standard Work,H. KentBowen and C. Purrington, 2005, Harvard Business

School Case Study

Example: Vapor Compression System

13

13 state variables,3 control variables2 boundary conditions

A typical system for design

The Current System Design Workflow

14

It limits to solving steady state simulation problemsThe system models are assembled from componentsAdditional requirements further complicate the design flow

Inefficient and time consuming

Guess design parameters

Update theparameter

No

Done

System design workflow

Yes

System simulation

checkrequirement

Source: Emerson climate

x state variabley control variable

boundary condition

The Current System Design Workflow (2)

15

The optimal design workflow is based on the running multiple simulations.Additional objectives increases the difficulty (….cost, weight, noise, emissions, sales,productivity, ….)

Lack of interaction and challenging for complex systems

Optimal design based on multiple simulations

Yes

Initial design parameter

Update theparameter

No

Select theOptimal solution

System design workflow

Check designparam range

More

energ

y

Bestdesign

All designs

Lesscomfort

Optimization Based Workflow – feasibility formulation

16

Seeks a feasible solution for system if exists

If weights are large enough and system is feasible, q ≈ r ≈ 0

Diagnose problematic system equation when q or r ≠ 0

Workflow with diagnostic capability

Reference valueControl variables

Slack variables

Feasibility Constraint

Optimization based workflow- targeting formulation

17

Define targeting constraints (c) to specify performance targets

Seek a value of the controls for which the feasibility constraints and targetsare satisfied

Targeting constraints relaxed, penalized with L1-norm

The formulation verifies the system reachability

It directly extends to system optimization.

Solve the set point reachability problem

Slack variables

Targeting constraints

Feasible solution

Optimization based workflow

18

Provide diagnostic information & system verification for users

Can the system operate at current operatingrange? If not, identify the potential modelingerror and identify the feasible region. If yes, thesimulation formulation can be used to solve theproblem. Diagnostic information for users.

The system is feasible at the given boundarycondition, can it reach the specified performancetarget? If not, identify & modify the limitingtargeting constraint. System verificationinformation for users.

The system can reach the target at the givenboundary condition, the process optimizationfinds the optimal solution w.r.t. system objectivefunction.

Boundary condition

Simulationproblem

Optimization problem

Feasibility problem

Targeting problem

19

0.5 1 1.5 2 2.5100

200

300

400

500

600

700

V_des (m3/s)

Pw

r(k

w)

Feasibility set boundary

Data points (Simulation sweep)

Solutions: capacity 2000 kW (Targeting sweep)

Example: Vapor Compression System

Feasibility, targeting and optimization

• Feasibility workflow identified the nonsmooth model• The targeting problem verifies the system capacity of 2000W.• Optimal solution is on the left boundary.

Agenda

20

Context and problem definitionWhy is it important to revisit the design and operation?How do we define system engineering?

System level design and operation workflows based on optimizationProviding diagnostic information for users.

Uncertainty analysis and fault operationPresenting the state-of-the-art and challenges

Optimization on the Embedded PlatformProviding flexibility for optimization-based applications

Summaries and Future Work

Uncertainty – State of the Art

21

Situation:• In a given system design, modeling and prediction of normal operation is

challenging but typically straight forward.• For failure analysis, a different mindset is needed, hypothesizing what can break

is not as straight forward.

An approach: uncertainty is (largely) a computational issue, need to attack fromhow to model uncertainty, how to get data, how to compute, how to exploit systemstructure…

How to put in a process or design flow & use control to mitigate uncertainty

Challenges: Fault handling (and detection) are a consequence of uncertainty…Critical for operations and critical to get at the complexity of softwareintensive systems.

Optimization formulation with uncertainty

22

Uses multi-scenario uncertainty parameter (ui).

The formulation chooses the finite amount of sampling points in the uncertaintyset .

Covering the entire uncertainty sets leads to infinite sampling points.

It leads to large scale optimization problems, thus computational challenges.

The approach with computational challenges

Agenda

23

Context and problem definitionWhy is it important to revisit the design and operation?How do we define system engineering?

System level design and operation workflows based on optimizationProviding diagnostic information for users.

Uncertainty analysis and fault operationPresenting the state-of-the-art and challenges

Optimization on the Embedded PlatformProviding flexibility for optimization-based applications

Motivation: Why Embedded Optimization

24

Distributed computational platform and user interface

State-of-the-Art

Challenges

Desired State

Optimization-based

supervisorycontrol

There are some challenges:

Guarantee of service,

Security, ….

But these are not our concern!

Objective

Optimization-based supervisory controlon embedded controller platforms as atechnology option.

Embedded Controller

Computational Platform

25

Limitations and Challenges?

On your PC

• 1GHz ARM Cortex-A8 Processor• 32-bit Dual-core architecture• 512MB DDR3 RAM• 4GB flash storage• Linux operating system• GNU development environment.

Embedded controller

• 2.8 GHz Intel® Core™ i7 Processor• 64-bit Quad-core architecture• 32 GB DDR3 RAM• 512GB flash storage• Variety operating system• Variety development/building

environment

Algorithms & Solvers

26

Algorithms evaluated for embedded optimization

Requirements on the algorithm:• Solves the relevant class of problems

(continuous, non-linear optimization problems),• Can be tailored to the specific problem and the computational platform.

Two practical algorithms:• Augmented Lagrangian Method (ALM)

• Use penalty parameter to handle the equalities or/and inequalities.• Solve unconstrained or bound-constrained sub-problems• State-of-the-art: MINOS and LANCELOT

• Primal-Dual Interior-Point Method (PDIPM)• Use logarithm-barrier parameter to handle inequalities• Solve equality-constrained sub-problems• State-of-the-art: IPOPT and PIPS-NLP

Algorithms & Solvers

27

Primal-Dual Interior Point Method (PDIPM)

Efficient and Customizable :

STEP 1: Fix μ, solve the sub-problems:

Newton-like step on the constraints promotessteady progress toward feasibility.

STEP 2: Update ߤ using different approaches.

CONVERGED ?

STEP 1

STEP 2

NO

YES

END

Instead of solving the original problem ܲܮܰ) ), PDIPM solves a sequence of log-barrierconstrained sub-problems ܲܮܰ) ఓ).

௫

ா

UPDATE ߤ�

(ఒ,ఘ)௫

ா

SOLVE ܲܮܰ ఓ

Pros:

Fast local convergence (superliner)

Cons:

Sub-problems are constrained optimization. Limitedselection of solving strategies.

Numerical Results on Embedded System

IPM(PIPS-NLP)(PC)PIM(PIPS-NLP)

(Embedded)Prob 1 Prob 2 Prob 1 Prob 2

InitPoint time (s) Iter time (s) Iter time (s) Iter time (s) Iter

setting 1 0.009 23 0.008 16 0.131 23 0.100 16

setting 2 0.009 23 0.009 16 0.13 23 0.102 16

setting 3 0.009 23 0.009 21 0.165 23 0.118 21

setting 4 0.009 22 0.010 21 0.12 22 0.117 21

setting 5 0.009 22 0.009 22 0.12 22 0.121 22

setting 6 0.008 21 0.010 24 0.118 21 0.129 24

setting 7 0.009 21 0.010 25 0.112 21 0.131 25

setting 8 0.009 21 0.009 23 0.117 21 0.130 23

setting 9 0.009 21 0.010 26 0.117 21 0.135 26

setting 10 0.009 22 0.011 30 0.118 22 0.148 30

Average 0.009 23 0.009 19 0.133 23 0.112 19

• The embedded platform solutions are the same as those on PC• The NLP problem has ~50 variables.• The average computational time on the embedded platforms is ~ 0.1 sec.• Memory usage: PC: 1.9 MB and BeagleBone: 1.4 MB

The optimization CAN be solved in the BeagleBone in real time

28

Embedded Optimization Customize

29

Interior Point Method Improvements:• Adaptive strategy to update the barrier parameter• Primal-dual regularization to improve numerical property• Use advanced linear algebra packages:

1. Customized LU solver for small dense linear system.2. Optimized Blas/Lapack routine for dense system3. Acceleration technique based on the utilizing problem structure

Further improvements to reduce the computation time

Constraint Control (Model Predictive control) Applications:• Systems with dynamics (problem structure).• Parameter estimation problem

Verification:• Feasibility guarantee with limited computational resources

Key Points

30

Energy efficient systems (situation)Market conditions cost, labeling and regulations require more energy efficientand environmental-friendly buildings. They pose challenges and opportunitiesfor new workflows for system design, operation and etc.

Takeaway:• System level engineering (not limited to component level) DOES benefit• Optimization-based approaches provide additional capabilities• Uncertainty analysis and fault handling need more work• Mobile computing platform is the desired state

Open Problems:• Modeling issues: industry-accepted, system compatibility• Diagnostic for users: workflows for engineers to identify modeling issues• Computational issues: solving large optimization problems online• Mobile computational platform: solving optimization problems on embedded

system

Slide Number 1Team and AcknowledgementKey PointsAgendaThe System Level View for BuildingsBuilding Energy ChallengesThe ImplicationsSlide Number 8Slide Number 9Slide Number 10AgendaEngineering Standard WorkExample: Vapor Compression SystemThe Current System Design WorkflowThe Current System Design Workflow (2)Optimization Based Workflow – feasibility formulationOptimization based workflow- targeting formulationOptimization based workflowExample: Vapor Compression SystemAgendaUncertainty – State of the ArtOptimization formulation with uncertaintyAgendaMotivation: Why Embedded OptimizationComputational PlatformAlgorithms & SolversAlgorithms & SolversNumerical Results on Embedded SystemEmbedded Optimization CustomizeKey Points