Slides - Carnegie Mellon University

mlsp.cs.cmu.edu

Slides - Carnegie Mellon University

Towards Energy-Aware

Facilities Through Minimally

Intrusive Approaches

Mario Bergés

Assistant Professor

Civil and Environmental Engineering

09/23/2010 – Carnegie Mellon University

Guest Lecture for 11-755 MLSP


2

99.2 quads Primary

Energy

40% Electricity 60% Other

73% Commercial

and Residential

Buildings

24% Other


4

{Benchmarking Buildings}

35 MW / month?

Green?

Efficient?

35 MPG

Hybrid

Fuel efficient


5

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


6

Collaborators

• Faculty:

– Lucio Soibelman (CEE)

– Scott Matthews (CEE)

– Anthony Rowe (ECE)

• Students:

– Kyle Anderson (ECE), Ethan Goldman (CEE)


{Introduction}

7


9

{Motivation}

Monthly: Free

Daily: Free*

N/A

$10k -

$50k

$2k -

$10k

Real-Time:

$200 - $3k

Moore’s law doesn’t apply to labor!


10

{Vision}

Low-Cost

High-Value


{Vision}

11


{1 Sample/day}

12


{1 Sample/minute}

13


{1 Hz | Coffee Machine}

14


{Sub-second}

15


{Mining the Total Electric Power}

16

• Non-Intrusive Load Monitoring

(NILM)

– Data Acquisition

– Event Detection

– Feature Extraction

– Classification

– Energy Computation


{Challenges}

19


20

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


21

{Research Questions}

• RQ1:

– For certain type of appliances, is it possible to

obtain signatures that generalize across many

different make/models?


22

{Research Questions}

• RQ2:

– How to take advantage of user input and other

sensor data to continuously train the

algorithms?

?


23

{Research Questions}

• RQ3:

– How to evaluate the performance of different

NILM algorithms using an objective and

comparable metric (e.g., accuracy, error rate,

etc.)?


24

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


25

{Data collection}

• A, E:

• B, D:

• C:

8, 30 appliances

34 , 6 transitions

483, 195 events

17, 8 appliances

44 , 16 transitions

281, N/A events

16 appliances

30 transitions

300+ events


{Experimental Setup}

Figure courtesy of: Anthony Rowe


27

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


Data Pre-Processing

28

{NILM Block Diagram}

Signature

Database

I n t e r e s t e d P a r t i e s

Electric

Power

Data:

v(t), i(t)

Edge

Detection

L o a d D i s a g g r e g a t i o n

Feature

Extraction

Classification

Appliance

State

Transition

Appliance

State Model


29

{Data Acquisition}

A

v(t)

i(t)

B

v(t)

i(t)


{Data Pre-Processing}

30


31

{Data Pre-Processing}

Real Power

Reactive

Power


{Where is the information?}

32


{Power Spectrum}

33


{Power Spectrum}

34


{Power Spectrum}

35


{Power Spectrum}

36


{Power Spectrum}

37


{Steady-State & Transients}

38


{Detecting Events}

39


{Detecting Events}

40


{Detecting Events}

41


{Classifying events}

42


43

{Feature Extraction}

Linear regression with

Linear regression:

non-linear basis

All samples: Delta, fit a line.

functions: or




43


44

{Feature Extraction}

Polynomial Basis:

Gaussian RBF:

All samples:


Fourier Basis:

44


45

{Classification}

k-NN:

Gaussian Naïve Bayes:

45


{Training}

46


A

A

47

t 1,2

t 1,2

t 2,3

B

B

C

A

A

t 1,2

t 2,3

t 3,1

t 2,1

t 1,2

t 2,3

t 3,1

B

C

B

C


48

A

t 2,1 = -30

t 1,2 = 30

t 3,1

B

t 2,3 = 50

C


{Energy Computation}

49


50

Determining best parameter settings

EVALUATION


{Event Detection Results}

51


{Event Detection Problems}

52


53

{Classification Results}

Validation Results

(Accuracy in %)

GNB kNN, k=1

Ada

Boost

Delta 52% 67% 51% 61%

DT

Laboratory

(A)

Whole Transient 38% 73% -- 58%

Polynomial

Coefficients 58% 67% 51% 52%

Fourier Coefficients 64% 79% 2% 64%

RBF Coefficients 67% 67% ** 64%

Delta 47% 73% 36% 42%

House

(B)

Whole Transient 9% 73% -- 47%

Polynomial

Coefficients 61% 80% 61% 57%

Fourier Coefficients 50% 80% 55% 54%

RBF Coefficients 47% 76% 35% 54%


{Classification Results: C.0}

54


{Overall Performance}

56


57

{Overall Performance}

- Measured

- Estimated with models, based on ground-truth

events

- Estimated with models, based on NILM predictions


58

{Overall Performance: C.0}

watts

watts


59

{Overall Performance: C.0}

watts

watts


{Classification Error}

60


{Event Detection Error}

61


{Energy Identification Ratios}

62


{Overall Performance: C.0}

63


64

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


65

{Data Collection}

Appliance Type # of Models Manufacturers

Turn-on

Transients

Turn-off

Transients

Refrigerators 13

Microwave Ovens 9

Compact

Fluorescent Lights

(CFLs)

8

Frigidaire, Americana,

Whirlpool, Absocold,

General Electric, Haier,

Quasar, Sanyo

Emerson, General

Electric, Samsung,

Westinghouse, Oster,

Amana

Sylvania, Bright Effects,

General Electric

19 18

58 59

14 19


{Transients: Refrigerator}

66


{Transients: CFL}

67


{Transients: Microwave}

68


{Transients: Refrigerator}

69


{General Models}

70


{Unwrapping}

71


72

{Eigen-analysis}

< >


{U 1 : Refrigerator Turn-on }

73


{U 2 : Refrigerator Turn-on }

74


{Cross-correlation u 1 F u [n]}

75


76

{Correlation u i u j }

Harmonics used: [1]


77

{Correlation u i u j }

Harmonics used: [1, 3, 5, 7]


{Cross-correlation Results: C.0} 78


{Cross-corr. Results: C.2 & C.3}

79


80

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


{Experimental Setup}

Figure courtesy of: Anthony Rowe


{Week-long datasets}

82


{Refrigerator (D)}

83


{Television (D)}

84


{Television (C.4)}

85


{Refrigerator (C.4)}

86


87

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


{Assumptions}

88


{Limitations}

89


90

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


91

{Conclusions}

• RQ1:


92

{Research Questions}

• RQ2:

– How to take advantage of user input and other

sensor data to continuously train the

algorithms?

A

B

C


93

{Research Questions}

• RQ3:

– How to evaluate the performance of different

NILM algorithms using an objective and

comparable metric (e.g., accuracy, error rate,

etc.)?

EIR(t)


94

{Outline}

• Introduction

• Research Scope

• Research Method

• User-centric NILM

• Robust Appliance Signatures

• Validation

• Discussion

• Conclusions

• Future Work


{Lower-Resolution Data}

95


{Latent Variable Models}

96


{Other data-sources}

97

Disaggregat

f s1

f or

motion

s2

power


{Sensing}

Historical

Data

App.

Database

Adapter

SOX lib

Web Service

Sensor

Gateway

Sensor

Registry

Meta-Sensor

Domain Data

Handler

Discovery

Historical

Data

App.

App.

App.

Event Node(s)


101

Testbed demonstration

http://undisclosed.location.com


{New insights}

102


{New insights}

103


{Other sources}

104


105

{Preliminary Results}

Correlations Best correlated with Correlation Coefficient P-value

bathroom light bathroom motion 0.59 0

bathroom motion bathroom light 0.59 0

bed lamp 1 bedroom motion 0.09 7.33E-19

bed lamp 2 bedroom motion 0.09 3.97E-17

bedroom light media laptop 0.28 2.21E-176

bedroom motion bathroom motion 0.24 5.29E-132

refrigerator kitchen motion 0.09 2.64E-19

kitchen motion living room motion 0.34 2.69E-269

laptop lcd monitor 0.77 0

lcd monitor laptop 1 0.77 0

living room audio tv 0.36 4.51E-300

living room motion kitchen motion 0.34 2.69E-269

media laptop bedroom light 0.28 2.21E-176

tall lamp laptop 1 0.15 3.06E-50

toaster kitchen motion 0.13 2.48E-35

tv living room audio 0.36 4.51E-300


106

Waste Detection

Automatically detecting waste.


{Higher level decisions}

107


108

{Broader Concept of Energy}

Automated continuous commissioning


Synchronization and

Pre-Processing

109

{Multi-Modal Distributed NILM}

Signature

Database

I n t e r e s t e d P a r t i e s

Electric

Power Data:

v(t), i(t)

Non-

Electricity

Data

Edge

Detection

Edge

Detection

L o a d D i s a g g r e g a t i o n

Feature

Extraction

Feature

Extraction

Local

Classification

A p p l i a n c e O p e r a t i o n I d e n t i f i c a t i o n

Appliance

State

Transition

Appliance State

Model


110

Questions?

THE END

More magazines by this user
Similar magazines