Building Energy and Comfort Management through ... - IBPSA-Boston

Building Energy and Comfort Management through Occupant Behaviour Pattern Detection Based On a LargeScale Environmental Sensor Network Bing Dong, Ph.D.

2011.05.19

IBPSA-Boston ,Seminar on Occupancy, Plug Loads, and Building Simulation, May 19, 2011

Introduction

Building Energy Consumption


Introduction

Installed units (millions)

Energy Saving Potentials for Building HVAC Control

Biggest Energy Impact

Subsystem Impact (kWh/bldg) *Transforming the Market: Energy Efficiency in Buildings by world business council for sustainable development (WBCSD), 2009.


3

Introduction

Dream Reduce energy consumption in buildings while maintaining occupant’s comfort

How to? Best achieved by understanding occupant’s behavior and patterns both in activities and energy consumption.

Objective Develop and implement algorithms for environmental sensor-based modeling and prediction of user behavior in intelligent buildings, and evaluate its energy impacts.


Methodology

Test-bed Setup

The data is collected continuously since January 29th, 2008 IBPSA-Boston ,Seminar on Occupancy, Plug Loads, and Building Simulation, May 19, 2011

Intelligent Workplace, CMU

5

Methodology

A Generalized Occupancy Model (1) -- Number of Occupants Implemented Method: Hidden Markov Model (HMM)

Ai ,n

Actual Estimated

Bay13 75% accuracy

Bo| j

Occupants ofOccupancy Number Number of

A j ,n

Ai , j

Bo|i

3

2

1

Bo|n 0 0

50

100

150

200

250

(Number of Steps) Number ofTime Time Steps (2 Minutes Interval)

300

350

3

Occupancy Number of of Occupants Number

Actual Estimated

2

1

0 0

HMM Structure

Bay10 70% accuracy

50

100

150

200

250

of Steps) Number ofTime(Number Time Steps (2 Minutes Interval)

Testing Results IBPSA-Boston ,Seminar on Occupancy, Plug Loads, and Building Simulation, May 19, 2011

300

350

Methodology 100 50 Model(2)--Occupancy Duration A Generalized Occupancy 0 06/01

Motivation: Sequential Data Mining Sensors

State Transitions Acoustics Data

2

90

06/08

Codes

80

Acoustics

70

a

1 Background Noise 60

50

40

30

b

0 06/01People Talking Off-On 20

10

0

00:00

Lighting

03:00

06:00

09:00

Lighting Data

12:00

15:00

18:00

21:00

100

06/08

c

90

80

70

60

Raw Sensor Data

d

50

On-off

40

Event Detector

30

20

2000 Motion 1500 Off-on (motion) 1000 500 On-off (no motion) 06/01 CO2 Increasing 10

06:00

09:00

12:00

15:00

Motion Data

18:00

21:00

00:00

03:00

1.8

1.6

1.4

e

1.2

1

0.8

0.6

0.4

f

0.2

0

11:10

11:20

11:30

11:40

11:50

12:00

12:10

12:20

CO2 Data

1800

1600

g

06/08

1400

1200

1000

600 09:00

Temperature 35 30 25

h

Decreasing

800

12:00

15:00

18:00

21:00

00:00

Increasing

i

Decreasing

j

Sample 06/01 events (60 minutes): agghkjhkjkkjkjilhjkkililaagbebfchdekcfeafefdllebcgbfglhibelfbebgk

7

60


06/08

Methodology

A Generalized Occupancy Model(2)--Occupancy Duration Implemented Method : Episode Discovery + Hidden Semi Markov Model (HSMM) Episode Discovery MDL

Raw Sensor Data

Event Detector

Period Detection

Schedule Prediction Occupancy Patterns

HSMM

Arranged Schedule Energy Weight

Predicted Schedule

8

MDL: Minimum Description Length HSMM: Hidden Semi Markov Model


Optimal Control Platform

Methodology

A Generalized Occupancy Model(2)--Occupancy Duration Acoustics Data 100 50

Results of Occupant Patterns

0 06/01

Example Events „bfjbe„ T „becfb„

06/08

c: light off->on b: acoustics happen f: motion on->off

06/08 06/08

j: temp decrease

06/15 06/15

06/22 06/22

Motion Data Lighting Data

2 100 1 50 0 06/01 0 06/01

06/08 06/08

06/15 06/15

06/22 06/22

CO2 Data Motion Data

06/08 06/08

06/15 06/15

06/22 06/22

Temperature Data CO2 Data

40 5000 30 0 20 06/01 -5000 06/01 100 40 50 30 0 06/01 20 06/01

06/22

Lighting Data Acoustics Data

100 100 50 50 0 06/01 0 06/01

5000 2 0 1 -5000 06/01 0 06/01

e: motion off->on

06/15

June 01 to June 22

06/08 06/08

06/15 06/15

06/22 06/22

RH Data Temperature Data

Raw sensor data from June 1st, 2009 to June 22th , 2009 06/08 06/08

06/15 06/15

06/22 06/22

RH Data 100 50 0 06/01


06/08

06/15

06/22

Methodology

A Generalized Occupancy Model(2)--Occupancy Duration i

 h  { i , Ai , j , Bo|i } Enter Room

k

Stay Room Motion On/Off Acoustics Loud Light On

  { i , Ai , j , Bo|i , Di }

HMM

Leave Room

cde cde j

cde cde i

j

k cde cde

HSMM1

Duration modelExponential distribution * 

D  Pr(m |  )  e m E[ X ]  1 i m

Example patterns X~(30)

X~(30)

0.8

0.6 Motion On/Off Acoustics Loud CO2 Increase

X~(120) 0.8 Light Off Motion Off

Motion On/Off Acoustics Loud Light On

X~(5) 0.2 1Duong,

T.V. Phung, D.Q. Bui, H.H. and Venkatesh, S. Human. 2006. Behaviour Recognition with Generic Exponential Family Duration Modelling in the Hidden Semi-Markov Model. In Proceedings of the 18th international Conference on Pattern Recognition. Vol. 3.


Methodology

Energy and Comfort Management Type of patterns 1. long-term patterns „cfbgefbfeefefefhibgefged’ (need n hours to finish) 2. short-term patterns such as „cedf’. Light onmotion onlight offmotion off (finish in 5 minutes)

Ventilation Rate Supply Air Temp

HVAC Controller

No action

Sensor data

Raw Sensor Data

Short-term Pattern Occupancy Behavior Pattern Discovery

Action with argmax{P(d)}

Long-term Pattern with duration


DP 160 Results—Part 1: Occupant Behavior Patterns 140

Number of Occupants

5 100

Actual Estimated

Number Of Occupants

Temperature(C)

120

4 80

3 60 2 40

180 1 20

DP

160 00 Mar/27

Mar/28

Mar/29

140

Mar/30 Mar/31 Time (minutes)

Apr/01

Apr/02

Apr/03

5 100

Actual

Number Of Occupants

Temperature(C)

Occupancy estimation results for Bay 10 from March 27th to April 3rd with GMM-HMM showing 82% accuracy 120

Estimated

4 80

3 60 2 40 1 20

00 Mar/27

Mar/28

Mar/29

Mar/30 Mar/31 Time (minutes)

Apr/01

Apr/02

Apr/03

Occupancy estimation results for Bay 13 from March 27th to April 3rd with GMM-HMM showing 85% accuracy IBPSA-Boston ,Seminar on Occupancy, Plug Loads, and Building Simulation, May 19, 2011

Results—Part 1: Occupant Behavior Patterns

Occupancy Duration # of Patterns

Longest Pattern

Most Compressed

Other Patterns

Pattern bfe, bff, efefe,

MDL (Bay 10)

7

begef (12)

aebf (22)

PD (Bay 10)

5

ebbfe(18)

be (51)

gfb, fge, beh

MDL (Bay 13)

9

efefbbe (11)

bf (15)

cef, beg, bbgdf, efef, aa, ghg,gge

PD (Bay 13)

6

bgbefe(12)

aeb (40)

egf, ff, efe, abc, fhd

bb, ghg

Discovered patterns on 10 minutes maximal time window in Bay 10

X~(45)

X~(55) 0.3

X~(10)

X~(30) Discovered patterns on 10 minutes maximal time window in Bay 13

aebf (22)

0.2

bf (15)

X~(55) 0.3

0.5

0.2

aeb(20)

bff (15)

a) Actual: 105(min); Predicted: 118(min)

0.25

X~(60)

0.3

X~(10)

X~(30) X~(40)

be (21)

ghg (15)

fhd (8)

X~(10) 0.2

b) Actual: 165(min); Predicted: 155(min) IBPSA-Boston ,Seminar on Occupancy, Plug Loads, and Building Simulation, May 19, 2011

efefe (7)

X~(78)

0.4

0.25

0.2

bef (8)

Results—Part 2: Energy Consumption

E+ Simulation Setup: 1)

Running period: from March 27th to April 3rd

2)

Assume a standard VAV system for IW

Control Strategies: 1)

Cooling set point of 24 °C and heating set point of 22°C from 7:00am to 6:00pm, and night setbacks for cooling of 30°C and heating of 15°C . (Fixed)

2)

When there is no occupant in the day time, the HVAC set point is set to 27 °C for cooling and 18 °C for heating. (Occupancy-based)

3)

The zone ventilation rate is designed to be 17 cfm/person (ASHRAE, 2004) Bay 10 13

Fixed Temp. Set-point (kWh) 64 48

Occupancy Based (kWh)

Savings (%)

Comfort Not Met (%)

51 40

20 17

14 11


Conclusion

Findings The most important sensors for the accurate occupant behavior pattern prediction are CO2, acoustics and motion. The developed occupant behavior pattern method can achieve 83% accuracy in terms of number of occupants prediction The maximum observed number of occupants is 4. In case of higher number of occupants, lower prediction accuracy may be expected. The energy savings are mainly from the dynamic control of HVAC system based on the real-time occupancy schedules. The energy savings from this study is based on simulation results with EnergyPlus’s “perfect” control. In reality, with a local controller, the temperature responses maybe different and the savings could be affected as well.


Thank you! Thank You!
