Model predictive control of energy flow andthermal comfort in buildings

Roy S. Smith

Automatic Control LaboratoryETH Zurich

rsmith@control.ee.ethz.ch

Thursday 23rd May, 2013

Collaborators: David Sturzenegger, Dimitrios Gyalistras, MarkusGwerder, Carina Sagerschnig, Frauke Oldewurtel, Manfred Morari.

Financial support: swisselectric research, Siemens, Gruner AG, Actelion

Pharmaceuticals, EMPA, CCEM.

2013-5-24 1

Introduction

Motivation: energy

I In Europe, 41% of energy is used within buildings;

I Also accounts for 36% of CO2 emissions;

I Large thermal energy storage gives energy shifting options.

Industry)26%)

Transport)33%)

Residential6Buildings)27%)

Commercial6Buildings)14%)

EU6Energy6Consumption)

2013-5-24 2

Introduction

Building sector

I The refurbishing rate in Switzerland is approx. 2% per annum;

I As much as 25% of the total energy is in the concrete;

I Control and commissioning occurs at building project end;

I Lack of integrated design and control;

I Currently little expertise in advanced control.

Potential MPC capabilities

I Potential use of weather forecasts;

I Potential use of occupancy forecasts;

I Dynamic electricity prices;

I Potential peak power reduction;

I Flexibility in upgrade modeling.

2013-5-24 3

Introduction

Model predictive control

I Constraints determine operational modes;I Most of the disturbances can be forecast:

I Ambient temperature (typ. 3 days ahead);I Solar radiation;I Occupancy and energy use patterns.

I The dynamics are slow and relatively benign(time constants: 30 minutes to days);

I The sampling rate is also slow (typ. 15 minutes);

I Adequate computational power is readily available;

I Sensing is (relatively) inexpensive;

I Actuation is typically expensive;

I Modeling can be time-consuming.

2013-5-24 4

MPC for buildings

Objectives

I Minimisation of energy (NRPE) or cost.

I Potentially time-varying energy/operational costs.

Constraints

I Comfort:I Room temperatures within limits;I CO2 level limit;

I HVAC:I TABS water supply temperature limits;I Minimum ventilation airflow;I Ventilation operational limits;I Air heating/cooling temperature limits;I Energy recovery system operational bounds.

2013-5-24 5

MPC for buildings

Modeling

I Building thermal dynamics;I Room node dynamics;I Solar radiation;I Ambient air and heat exchange;I Energy exchange within the building;

I HVAC dynamics;

I Forecasting error dynamics.

Disturbances (partially predictable)

I Ambient temperature;

I Solar radiation;

I Occupancy and use.

2013-5-24 6

Opticontrol project

Opticontrol I: (2007–2010)

I Integrated room automation studies;I MPC potential using:

I weather forecasting;I dynamic electricity pricing;I occupancy information.

I IfA, EMPA Dubendorf, Meteo Swiss, Siemens, Gruner AG.

Opticontrol II: (2011–2013)

I MPC demonstration;

I IfA, Siemens, Roschi + Partner AG (Gruner),Actelion Pharmaceuticals.

2013-5-24 7

Opticontrol I

Key contributors:F. Oldewurtel, D. Sturzenegger, C.N. Jones, A. Parisio, M. Morari, D.Gyalistras, A. Ulbig, G. Andersson, OptiControl Team.

Case studies

Building/HVAC

5 HVAC systems2 building standards2 construction types2 window area fractions

Weather

10 locations

Control

Rule-basedDeterministic MPCStochastic MPCPerformance bound

960 building cases × 10 locations × 4 orientations = 38’400 cases.

2013-5-24 8

Opticontrol II

Theoretical savings potential

Assessment with idealized cases: – perfect predictions;– perfect modeling.

Example:

60% of cases have < 100Kh/a violations and> 30% NRPE use withrespect to the performancebound.

2013-5-24 9

Stochastic MPC

Building control formulation

J(x0) = minµ(w)

E

{N−1∑k=0

cTuk

}Energy use

subject to:

Su ≤ s, HVAC inputs

P {Gjx ≤ gj} ≥ 1− α, Thermal comfort

x = Ax0 +Bu+ Ew, Building dynamics

w ∈ N (0, I), Weather prediction errors

uk = µk(w0, . . . , wk−1). Control policy

2013-5-24 10

Stochastic MPC

Affine disturbance feedback approximation

µk(w0, . . . , wk−1) = hk +

k−1∑p=0

Mk,pwp,=Mw + h

Input optimisation vs. policy optimisation

Optimisation over uk, uk+1, . . . Optimisation over µk(w)

2013-5-24 11

Stochastic MPC

Deterministic reformulation of chance constraintsState constraints:

P {Gjx ≤ gj} ≥ 1− α

With affine disturbance feedback,

P{Gj(Ax0 +Bh+BMw + Ew) ≤ gj} ≥ 1− α

This is reformulated as,

Gj(Ax0 +Bh) ≤ gj − φ−1(1− α)‖Gj(BM + E)‖2

(second order cone constraint)

2013-5-24 12

Stochastic MPC

Reformulation of chance constraints as an SOCP

J(x0) = minM,h

N−1∑k=0

cThk

subject to:

Sih ≤ si − φ−1(1− α)‖SiM‖2,Gj(Ax0 +Bh) ≤ gj − φ−1(1− α)‖Gj(BM + E)‖2,

Evaluating the expectation makes the cost function deterministic.

2013-5-24 13

Stochastic MPC

Rule-based controller/Stochastic MPC comparison

Simulation over 24 hours (with a 1 hour sampling time)

Six representative building/location cases:

SMPC RBC SMPC RBC

2013-5-24 14

Stochastic MPC

Tunability

P{Gjx ≤ gj} ≥ 1− α

The energy/comfort trade-off can be tuned by α:

2013-5-24 15

Dynamic electricity tariffs

Normalized daily maximum peak

0

0.1

0.2

0.3

0.4

0.5

0.6

Passive House Struct heavy WF high IG high

Passive House Struct heavy WF high IG low

Passive House Struct heavy WF low IG high

Passive House Struct heavy WF low IG low

Passive House Struct light WF high IG high

Passive House Struct light WF high IG low

Passive House Struct light WF low IG high

Passive House Struct light WF low IG low

Swiss Average Struct heavy WF high IG high

Swiss Average Struct heavy WF high IG low

Swiss Average Struct heavy WF low IG high

Swiss Average Struct heavy WF low IG low

Swiss Average Struct light WF high IG high

Swiss Average Struct light WF high IG low

Swiss Average Struct light WF low IG high

Swiss Average Struct light WF low IG low

Constant Tariff Day/Night Tariff Dynamic Tariff 2013-5-24 16

Opticontrol II

Key contributors:D. Sturzenegger, D. Gyalistras, M. Gwerder, C. Sagerschnig, M. Morari,R. Smith.

Demonstrator building

OptiControl Phase II – Demonstrator Gebäude

Gebäudedaten

Neubau 2007

Büro (1.-5. OG), Restaurant (EG), Einstellhalle (1.-2. UG)

6000 m2 klimatisierte Fläche

Gemessener Wärmeverbrauch 46 kWh/m2 pro Jahr

Gemessener Elektrizitätsverbrauch 83 kWh/m2 pro Jahr

Gebäudetechnik

TABS (Heizen/Kühlen)

Ventilation mit Wärmerückgewinnung, adiabater Kühlung und Heizung

Radiatoren in Eckbüros

Wärmegenerierung: Gasboiler

Kältegenerierung TABS: Kühlturm

Storen

Located in Basel.6 storeys + 2 underground.Offices in upper 5 floors.2nd floor instrumented.Siemens building control system.

2013-5-24 17

Actuation

TABS for heating/cooling

Water piping within floor.

Using building thermalstorage.

Heat: gas boiler.

Cooling: dry water tower.

Cooling restricted to night.

Efficiency depends onambient humidity andtemperature.

2013-5-24 18

Actuation

Ventilation

Heater on incoming air.

Cooling on return air.

Energy recovery via heatexchanger.

Minimum air flow requirements.

Significant losses in ducting.

2013-5-24 19

Actuation

Blinds

Control only by facade.

Limited settings: open, closed,shaded (2 levels).

Operation limited to 3 times perday for user acceptance.

Manual override possible (andundetectable).

OptiControl Phase II – Demonstrator Gebäude

Gebäudedaten

Neubau 2007

Büro (1.-5. OG), Restaurant (EG), Einstellhalle (1.-2. UG)

6000 m2 klimatisierte Fläche

Gemessener Wärmeverbrauch 46 kWh/m2 pro Jahr

Gemessener Elektrizitätsverbrauch 83 kWh/m2 pro Jahr

Gebäudetechnik

TABS (Heizen/Kühlen)

Ventilation mit Wärmerückgewinnung, adiabater Kühlung und Heizung

Radiatoren in Eckbüros

Wärmegenerierung: Gasboiler

Kältegenerierung TABS: Kühlturm

Storen

2013-5-24 20

Sensing

Additional sensors installed in 22 rooms on the 2nd floor.

Window sensor Room temperature Illumination/presence

2013-5-24 21

Sensing

Weather station on roof Measurements:

Ambient temperature;Solar radiation (4 directions);

Additional radiation sensor oneach facade;

Existing building control sensing(HVAC, cooling tower, etc.).

2013-5-24 22

Building control system

Automation level(Actelion G03)

Remote access

Field level(Actelion G03)

Management level(Actelion research center)

BACnet-RouterBACnet over LON

LONMARK over LON

BACnet over IP

Industry PCBACnet OPC serverOptiControl-II high-level control Matlab (OPC toolbox)OPC client

PCDESIGO INSIGHTDatabase, used by Facility Management

Internet

PXE-CRSIntegrationof M-Bus meter

PXC64-UVentilation, kitchen and restaurant

PXC128-UHeating, misc. electrical devices

PXC64-UVentilation, offices, auxiliary rooms, parking garage

PXC64-UCooling, exhaust air,blinds

PXR11Integrationof single-room control

RXC31.1Conference room B26

RXC31.1Conference room A26

RXC31.1Conference room D26

RXC31.1Conference room C26

RXC31.1Conference room E26

PXC00/PXXL11Integration of EnOcean components

PXC100/TX-OpenIntegration of M-Bus meter

2013-5-24 23

Modeling

Energy Plus: “truth model” for simulation studies

2013-5-24 24

Modeling

Modular components

Building geometry,location and orientationConstruction material data sheets Actuation components

Thermal model

TABS heating model

Solar radiation

TABS cooling model

Ventilation model

Blinds model

External temp.

Buildingtemperatures

Energy usage

Heatfluxes

2013-5-24 25

Modeling

EnergyPlus: Physics based; construction/material specifications; 2nd floor (725 sq.m.) divided into 24 zones; Simulation studies (via BCVTB) possible; Unsuitable for MPC implementation.

Large-scale RC model: Semi-automated derivation from EP; 2nd floor thermal model has 294 states; Too large for MPC implementation;

Convenient test environment for MPC.

Reduced-order RC model: Hankel-norm reduction of thermal model; Thermal model: 15 states (with 0.1% error); Actuation adds 10 more states; Time constants: 20, 8 and 5 hours. Majority < 30 minutes.

2013-5-24 26

Modeling

RC modeling

Thermal dynamics model:

xk+1 = Axk +Bqqk(x,uk, vk) (linear thermal dynamics)

External heatflux model:

qk(xk, uk, vk) = Aqxk+Bq,uuk+Bq,vvk+

nu∑i=1

(Bq,vuvk+Bq,xuxk)ui,k

Bilinear actuation/heatflux dynamics.

Thermal model reduced to approximately 25 states in total.

Another 10 states are used in modeling the actuation.

2013-5-24 27

MPC annual simulation configuration

EnergyPlusmodelRoom temps.

Stored energy Solar radiation / ambient temp.(simulated)

Energy use

Simplified RCbuilding model

TABS/ventilation/blinds

Weather forecasts(simulated)Room temps.

(averaged)

Stored energyEnergy use

Optimization

Random disturbances

Matlab environment (periodic operation)

Simulated OPC interface

Matlab environment

Run-timedata storage

2013-5-24 28

MPC testing and development structure

Large-scaleRC modelRoom temps.

Stored energy Solar radiation / ambient temp.(simulated)

Energy use

Simplified RCbuilding model

TABS/ventilation/blinds

Weather forecasts(simulated)Room temps.

(averaged)

Stored energyEnergy use

Optimization

Random disturbances

Matlab environment (periodic operation)

Simulated OPC interface

Matlab environment

Run-timedata storage

2013-5-24 29

MPC building operation structure

BuildingRoom temps.Stored energy Solar radiation / ambient temp.

Energy use

Simplified RCbuilding model

TABS/ventilation/blinds

Weather forecasts(MeteoSwiss)Room temps.

(averaged)

Stored energyEnergy use

Optimization

Occupancy/internal gains

Matlab environment (periodic operation)

OPC interface

Run-timedata storage

Building

2013-5-24 30

Basic MPC structure

y(t)

u(t)

Kalmanfilter

Building measurements

Prediction:low order

model

Weather predictions

TABS/ventilation/blinds

Room temps.(averaged)

Stored energyEnergy use

ConstrainedEnergy

Optimization

Estimated statesActuation

Kalmanfilter

w(t)

x(t)

2013-5-24 31

Software structure

MeteoSwiss)server,!  weather'forecasts,!  updated'every'

12h,

Building)database)server,

•  measurement'data'storage,

Control)PC,!  Runs'Matlab,!  Write'interface'

to'building,

Desigo)PC,!  reads'sensors,!  visualizes'

measurements,!  feeds'database,

Building)incl.)Siemens)low)level)control)system),

External)PC,

on>site,

measurements,

commands,

remote)control,

weather)predictions,

measurements,

2013-5-24 32

MPC algorithm

Problem framework

I Sampling period is 15 minutes.

I Prediction horizon is 54 hours.

I Three weather forecasts per day (each valid for 72 hours).

Optimisation

I Mixed integer problem (quantized blind positions)I Two step sub-optimal solution:

I Solve with continuous blind positions.I Round blind controls to allowed values.I Resolve over remaining control variables.

2013-5-24 33

Operation

Jun.09 00:00 Jun.10 00:00 Jun.11 00:00 Jun.12 00:00 Jun.13 00:0012

14

16

18

20

22

24Outside Air Temperature [degrees C]

Filtered Prediction

Measurement

“Raw” Prediction

Current time

w(t)

Weather prediction filtering Filter dynamics are based on OptiControl I analyses.

2013-5-24 34

Operation

Disturbances (measured and filtered predictions)

Jun.18 00:00

Jun.18 12:00

Jun.19 00:00

Jun.19 12:00

Jun.20 00:00

Jun.20 12:00

Jun.21 00:00

Jun.21 12:00

Jun.22 00:00

Jun.22 12:00

Jun.23 00:00

0

50

100

150

200

250

300

Jun.18 00:00 Jun.19 00:00 Jun.20 00:00 Jun.21 00:00 Jun.22 00:00 Jun.23 00:00

10

15

20

25

30

35Temperature [degC]Solar Heatflux [W/m2]Historical Prediction

Maximum Solar

Heatflux (SE) [W/m2]

Ambient Temperature [degC]

Maximum Solar

Heatflux (NW) [W/m2]

2013-5-24 35

Operation

Actuation (TABS and blinds)

Jun.18 00:00 Jun.19 00:00 Jun.20 00:00 Jun.21 00:00 Jun.22 00:00 Jun.23 00:00

0

50

100

150

200

250

300

Jun.18 00:00

Jun.18 12:00

Jun.19 00:00

Jun.19 12:00

Jun.20 00:00

Jun.20 12:00

Jun.21 00:00

Jun.21 12:00

Jun.22 00:00

Jun.22 12:00

Jun.23 00:00

0

5

10

15

Solar Heatflux [W/m2]

HistoricalCooling Power [kW]

Prediction

Solar Heatflux

(SE) [W/m2]

Solar Heatflux (NW) [W/m2]

TABS Cooling

Power [kW]

2013-5-24 36

Operation

Room temperatures (2 facades)

18

20

22

24

26

28

30

18

20

22

24

26

28

30

Historical Prediction

Average Room Temperature (SE)Average Room Temperature (NW)

Temperature [deg C]

Jun.18 00:00

Jun.18 12:00

Jun.19 00:00

Jun.19 12:00

Jun.20 00:00

Jun.20 12:00

Jun.21 00:00

Jun.21 12:00

Jun.22 00:00

Jun.22 12:00

Jun.23 00:00

Temperature [deg C]

Comfort constraint

Comfort constraint

2013-5-24 37

Room temperatures (summer operation)

May 01 May 06 May 11 May 16 May 21 May 26 May 31 Jun 05 Jun 10 Jun 15 Jun 2020

21

22

23

24

25

26

27

28Te

mpe

ratu

re [d

egC

]

Average Measured Room Temperature

Maximum and Minimum Measured Room Temperature

Target Comfort Range

Room Temperatures [degrees C]

2013-5-24 38

Room temperatures (winter operation)

19 Nov 26 Nov 03 Dec 10 Dec 17 Dec 24 Dec 31 Dec 07 Jan 14 Jan 21 Jan 28 Jan 04 Feb 11 Feb 18 Feb21

21.5

22

22.5

23

23.5

24

24.5

25

25.5

26

Average Roomtemperature

MPC controller operation (92 days)

2013-5-24 39

Comfort violations

11/11 18/11 25/11 02/12 09/12 16/12 23/12 30/12 06/010

2

4

6

8

10

11/11 18/11 25/11 02/12 09/12 16/12 23/12 30/12 06/0121

22

23

24

25

26

Room B4

Room B9

Room B21

Daily Kh Violations Room B4

Daily Kh Violations Room B21Daily Kh Violations Room B9

2013-5-24 40

Validation experiments (step responses)

23 Dec−12AM 25 Dec−12AM 27 Dec−12AM20

21

22

23

24

25

26

27

23 Dec−12AM 25 Dec−12AM 27 Dec−12AM20

21

22

23

24

25

26

27

Simulated Operative Temperature South

Measured Operative Temperature South

Simulated Operative Temperature West

Measured Operative Temperature West

Step response experiments: 22/12/2012 to 26/12/2012

Experiment comparison with simulation (reduced order model).

Best and worst façade cases shown.

Simulated TABS inputs are the heat fluxes.

2013-5-24 41

Validation experiments (errors)

23 Dec−12AM 25 Dec−12AM 27 Dec−12AM

−0.5

0

0.5

1

1.5

Meas. − Sim. Operative Temperature East

Meas. − Sim. Operative Temperature South

Meas. − Sim. Operative Temperature WestMeas. − Sim. Operative Temperature North

Step response experiments: 22/12/2012 to 26/12/2012

Experiment/simulation errors: all four façades

2013-5-24 42

Energy balance (simulated operation)

Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec���

���

���

���

�

����

����

����

Ne

t E

ne

rgy [

kW

h]

TABS

AHU

Internal gains

Window (solar gains)

Radiators

Window (conduction)

Infiltration

Building hull (excluding windows)

Energy balance within the controlled building

The data is averaged by month

Closed blinds explain lower solar gains during summer.

TABS and AHU act as both heating and cooling actuators.

TABS is used more for heating and AHU is used more for cooling.

2013-5-24 43

Demand reduction experiments

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 230

5

10

15

20

25

30

35

Price Signal (Qualitative)

TABSVentilation

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 230

5

10

15

20

25

30

35

TABS

Price Signal (Qualitative)

Ventilation

MPC can use constraints and/or price signals to shift electrical load or gas consumption

Measured average hourly power consumption.(18 Nov 2012 - 01 Feb 2013)

Measured average hourly power consumption.05 Feb 2013 - 14 Feb 2013)

2013-5-24 44

Conclusions

I MPC is effective at managing energy and thermal comfort inthe building.

I MPC gives 10 to 15% reduction in heating/cooling energy.

I Thermal comfort is significantly improved over originalbuilding RBC.

I Tuning is relatively straightforward.

I Good user acceptance (6 months of operation).

I Modeling is a significant effort.

I Potential for load shifting within the grid.

I Flexibility is one of the most significant benefits.

2013-5-24 45

References

There are many...

See http://www.opticontrol.ethz.ch.

2013-5-24 46