Supero: A Sensor System for Unsupervised Residential Power Usage - - PowerPoint PPT Presentation

A Sensor System for Unsupervised Residential Power Usage Monitoring

Supero:

Dennis E. Phillips1, Rui Tan2; Mohammad-Mahdi Moazzami1; Guoliang Xing1; Jinzhu Chen1; David K. Y. Yau2,3

1Michigan State University, USA 2Advanced Digital Sciences Center, Singapore 3Purdue University, USA

1 / 23

Outline

• Motivation & Approach
• Light Sensing
• Acoustic Sensing
• Implementation & Experiments
• Implementation & Experiments

2 / 23

Residential Electricity in U.S.

• Residential electricity

– Largest sector

• Rising cost

Industrial 25.5% Residential 36.7% Commercial 34.2% Others

• Rising cost

– Increase by 75% in 10 years

• Understanding usage

– Real-time power readings – Fine-grained usage info

Electricity retail sales in U.S. 2011 [US EIA-861, EIA-923]

Appl. Joul % When? Bed light 5% 7pm-11pm Fridge 8% Every 1h Space heater 30% Jan 1 … …. …. ….

3 / 23

Related Work

• Direct sensing

– ACme [IPSN’09] Per-appliance inline meter, intrusive

[Jiang IPSN’09]

4 / 23

Related Work

• Direct sensing

– ACme [IPSN’09] Per-appliance inline meter, intrusive

• Indirect sensing

[Jiang IPSN’09]

• Indirect sensing

– At-the-flick [UbiComp’07] High-rate ADC, in-situ training

5 / 23

Related Work

• Direct sensing

– ACme [IPSN’09] Per-appliance inline meter, intrusive

• Indirect sensing

[Jiang IPSN’09]

• Indirect sensing

– At-the-flick [UbiComp’07] High-rate ADC, in-situ training – ViridiScope [UbiComp’09] Labor-intensive sensor installation

6 / 23

[Kim UbiComp’09]

Objective & Challenge

• Fine-grained usage monitoring

– Accurate energy disaggregation – Inexpensive and easy-to-install sensors – Training-free, ad hoc system deployment (“place sensor on shelf facing light to be monitored”) sensor on shelf facing light to be monitored”)

7 / 23

Objective & Challenge

• Fine-grained usage monitoring

– Accurate energy disaggregation – Inexpensive and easy-to-install sensors – Training-free, ad hoc system deployment (“place sensor on shelf facing light to be monitored”) sensor on shelf facing light to be monitored”)

• High-degree sensing uncertainty

– Noises from environment and human activities – Source appliance identification

• A sensor can sense multiple appliances
• An appliance can be sensed by multiple sensors

8 / 23

Supero

Smart meter Base station Light and acoustic sensors 9 / 23 Light + acoustic captures 90% power consumption

Supero

Smart meter Base station

100W ‘+1’

Light and acoustic sensors

Event Correlation (remove false alarm) Light/acoustic event Power reading

10 / 23 Light + acoustic captures 90% power consumption

Supero

Smart meter Base station

100W ‘+1’

Light and acoustic sensors

Event Correlation (remove false alarm)

Event clustering

Light/acoustic event Power reading

11 / 23 Light + acoustic captures 90% power consumption

Supero

Smart meter Base station Event-Appliance Association

100W ‘+1’

Light and acoustic sensors

Event Correlation (remove false alarm)

Event clustering Association

Light/acoustic event Power reading

12 / 23 Light + acoustic captures 90% power consumption

Supero

Smart meter Base station Event-Appliance Association

100W ‘+1’

Light and acoustic sensors

Event Correlation (remove false alarm)

Event clustering Association

Light/acoustic event Power reading

13 / 23 Light + acoustic captures 90% power consumption

Outline

• Motivation & Approach
• Light Sensing
• Acoustic Sensing
• Implementation & Experiments
• Implementation & Experiments

14 / 23

Event Detection & Correlation

light sensor reading exponential diff

Time (second)

• Exponential difference filter

– Diff between long-/short-term moving averages

15 / 23

Time (second)

Event Detection & Correlation

light sensor reading exponential diff

Time (second)

Light on

Report event

Light off

Report event

• Exponential difference filter

– Diff between long-/short-term moving averages

16 / 23

Time (second)

Light on

Event Detection & Correlation

light sensor reading exponential diff

Time (second)

Light on

Report event

Light off

Report event

Human movements

• Exponential difference filter

– Diff between long-/short-term moving averages

17 / 23

Time (second)

Light on movements

Event Detection & Correlation

light sensor reading exponential diff

Time (second)

Light on

Report event

Light off

Report event

Human movements

• Exponential difference filter

– Diff between long-/short-term moving averages

• Event correlation

– Simultaneous events have same source – False alarm if no power reading change

18 / 23

Time (second)

Light on movements

Light Event Clustering

Light 1 Light 3 Sensor 1 Sensor 2

• Feature: change of light intensity

Floor plan Clustering on intensity changes

19 / 23

Sensor 1 Light 2

Light Event Clustering

Cluster A Cluster B Cluster C Light 1 Light 3 Sensor 1 Sensor 2

• Feature: change of light intensity

Cluster C Floor plan Clustering on intensity changes

20 / 23

Sensor 1 Light 2

Light Event Clustering

Cluster A Cluster B Cluster C Light 1 Light 3 Sensor 1 Sensor 2

• Feature: change of light intensity
• {Cluster A, B, C} ↔ {Light 1, 2, 3}?

Cluster C Floor plan Clustering on intensity changes

21 / 23

Sensor 1 Light 2

Power Law Decay of Light

log (measurement) (intensity)

measurement = β power d-α

22 / 23

log (distance from light source) lo

Power Law Decay of Light

log (measurement) (intensity)

measurement = β power d-α

23 / 23

log (distance from light source) lo distance from light source

Power Law Decay of Light

log (measurement)

α = 3.5

(intensity)

measurement = β power d-α

24 / 23

log (distance from light source) lo Path loss exponent α ∈ [2, 5] distance from light source

Power Law Decay of Light

log (measurement)

α = 3.5

(intensity)

measurement = β power d-α

25 / 23

log (distance from light source) lo Path loss exponent α ∈ [2, 5] distance from light source Scaling factor

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

26 / 23

Discrepancy between model prediction and observation for sensor i

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m

27 / 23

Discrepancy between model prediction and observation for sensor i

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i

28 / 23

Discrepancy between model prediction and observation for sensor i

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i sensors that can detect event in cluster m

29 / 23

Discrepancy between model prediction and observation for sensor i

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i sensors that can detect event in cluster m

30 / 23

Discrepancy between model prediction and observation for sensor i Light 1 (50W) Light 2 (50W) Sensor 1

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i sensors that can detect event in cluster m

31 / 23

Discrepancy between model prediction and observation for sensor i Light 1 (50W) Light 2 (50W)

Intensity change

• f sensor 1

Sensor 1 Cluster 1 Cluster 2 P1 = 49W P2 = 52W

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i sensors that can detect event in cluster m

32 / 23

Discrepancy between model prediction and observation for sensor i Light 1 (50W) Light 2 (50W)

Intensity change

• f sensor 1

Sensor 1 Cluster 1 Cluster 2 P1 = 49W P2 = 52W

em,j

Light 1 Light 2 Cluster 1 3000 100 Cluster 2 110 4000

Cluster-Light Association

• Error of associating cluster m and light j

∑

∈ − −

⋅ ⋅ =

m

R i m i j i m j m

d P e

, , ,

µ β

α

• bserved intensity

change of sensor i in cluster m model-predicted intensity change of sensor i sensors that can detect event in cluster m

33 / 23

Discrepancy between model prediction and observation for sensor i Light 1 (50W) Light 2 (50W)

Intensity change

• f sensor 1

Sensor 1 Cluster 1 Cluster 2 P1 = 49W P2 = 52W

em,j

Light 1 Light 2 Cluster 1 3000 100 Cluster 2 110 4000

Cluster-Light Association (cont’d)

• For given light decay model, find a binary matrix [am,j]

∑ ∑ ∑

∀ ∀

⋅ =

j m j m j m

e a E ) , ( min

, , ,

β α

am,j=1: cluster m is associated with light j

34 / 42

∑ ∑

∀ ∀

= =

j j m m j m

a a 1 , 1 s.t.

, ,

Cluster-Light Association (cont’d)

• For given light decay model, find a binary matrix [am,j]

∑ ∑ ∑

∀ ∀

⋅ =

j m j m j m

e a E ) , ( min

, , ,

β α

Total association error

am,j=1: cluster m is associated with light j

35 / 42

∑ ∑

∀ ∀

= =

j j m m j m

a a 1 , 1 s.t.

, ,

Cluster-Light Association (cont’d)

• For given light decay model, find a binary matrix [am,j]

∑ ∑ ∑

∀ ∀

⋅ =

j m j m j m

e a E ) , ( min

, , ,

β α

Total association error

am,j=1: cluster m is associated with light j

36 / 42

∑ ∑

∀ ∀

= =

j j m m j m

a a 1 , 1 s.t.

, ,

One-to-one mapping

Cluster-Light Association (cont’d)

• For given light decay model, find a binary matrix [am,j]

∑ ∑ ∑

∀ ∀

⋅ =

j m j m j m

e a E ) , ( min

, , ,

β α

Total association error

am,j=1: cluster m is associated with light j

– Hungarian algorithm

37 / 42

∑ ∑

∀ ∀

= =

j j m m j m

a a 1 , 1 s.t.

, ,

One-to-one mapping

Cluster-Light Association (cont’d)

• For given light decay model, find a binary matrix [am,j]

∑ ∑ ∑

∀ ∀

⋅ =

j m j m j m

e a E ) , ( min

, , ,

β α

Total association error

am,j=1: cluster m is associated with light j

– Hungarian algorithm

• Iterate α and β to further minimize E(α, β)

– Adaptively calibrate environment-dependent α and β

38 / 42

∑ ∑

∀ ∀

= =

j j m m j m

a a 1 , 1 s.t.

, ,

One-to-one mapping

Outline

• Motivation & Approach
• Light Sensing
• Acoustic Sensing
• Implementation & Experiments
• Implementation & Experiments

39 / 23

Adaptive Acoustic Sampling

Signal energy

Slow sampling (1 KHz)

Signal energy > T1 Signal energy

Signal energy # zero crossings

base station

Fast sampling (12 KHz)

Signal energy < T2

40 / 23

Low pass Band pass High pass

Clustering-based Event Detection

• Multiple phases (fan, microwave)

– Unknown and unpredictable

• K-means clustering

– Automatically identify K

scatter cluster within scatter cluster between max

41 / 23

Clustering-based Event Detection

• Multiple phases (fan, microwave)

– Unknown and unpredictable

• K-means clustering

– Automatically identify K

scatter cluster within scatter cluster between max

Detect the phase changes of 3-speed fan

42 / 23

Acoustic feature

Clustering-based Event Detection

• Multiple phases (fan, microwave)

– Unknown and unpredictable

• K-means clustering

– Automatically identify K

scatter cluster within scatter cluster between max

Detect the phase changes of 3-speed fan

43 / 23

Acoustic feature metric

Clustering-based Event Detection

• Multiple phases (fan, microwave)

– Unknown and unpredictable

• K-means clustering

– Automatically identify K

scatter cluster within scatter cluster between max

Detect the phase changes of 3-speed fan

44 / 23

Acoustic feature metric K=3

Outline

• Motivation & Approach
• Light Sensing
• Acoustic Sensing
• Implementation & Experiments
• Implementation & Experiments

45 / 23

Implementation & Deployments

TelosB (light) Iris (acoustic) Kill-A-Watt Apartment-1 deployment

• System

– TelosB/Iris + TED5000 + KAW ground truth meters

• Five deployments

– Three apartments (40~150 m2), two houses – 9 ~ 22 sensors

Iris (acoustic) Kill-A-Watt Apartment-1 deployment

46 / 23

Supero in Action

• Video demonstrating installation and setup

47 / 23

Evaluation

• 10 days experiment in Apartment-1
• Impact of sensor deployment in Apartment-2
• Compare with ViridiScope [UbiComp’09]

(Regression on appliance states + power readings)

– Oracle: ground truth appliance states – Oracle: ground truth appliance states – Baseline: closest appliance is source

48 / 23

10-day Results

Appliance Supero Oracle Baseline kWh Error (%) kWh Error (%) kWh Error (%) Light 1 4.17 0.5 4.11 0.9 4.11 0.9 Light 2 4.96 0.1 4.92 0.8 4.92 0.8 Light 3 6.24 1.4 6.25 1.7 6.25 1.7 Light 4 1.45 0.1 1.45 0.1 1.48 1.7 Light 5 0.39 0.2 0.39 0.7 0.41 5.5 Water boiler 0.48 0.5 0.48 0.5 100 Tower fan 0.21 50 0.17 17.9 0.24 66.2

• Supero

– All 146 light events detected, no false alarm, no miss 49 / 23

Rice cooker 0.98 2.2 1.01 1.2 1.01 0.8 Hair dryer 0.07 19.2 0.09 0.4 0.02 73.2 Fridge 11.8 3.7 11.8 3.2 11.8 3.2 Bath fan 0.12 N/A 0.17 N/A N/A Router 2.03 4.3 3.04 43.3 3.04 43.3 Average error 7.5 6.5 27.0

10-day Results

Appliance Supero Oracle Baseline kWh Error (%) kWh Error (%) kWh Error (%) Light 1 4.17 0.5 4.11 0.9 4.11 0.9 Light 2 4.96 0.1 4.92 0.8 4.92 0.8 Light 3 6.24 1.4 6.25 1.7 6.25 1.7 Light 4 1.45 0.1 1.45 0.1 1.48 1.7 Light 5 0.39 0.2 0.39 0.7 0.41 5.5 Water boiler 0.48 0.5 0.48 0.5 100 Tower fan 0.21 50 0.17 17.9 0.24 66.2

• Supero

– All 146 light events detected, no false alarm, no miss – Comparable to Oracle 50 / 23

Rice cooker 0.98 2.2 1.01 1.2 1.01 0.8 Hair dryer 0.07 19.2 0.09 0.4 0.02 73.2 Fridge 11.8 3.7 11.8 3.2 11.8 3.2 Bath fan 0.12 N/A 0.17 N/A N/A Router 2.03 4.3 3.04 43.3 3.04 43.3 Average error 7.5 6.5 27.0

10-day Results

Appliance Supero Oracle Baseline kWh Error (%) kWh Error (%) kWh Error (%) Light 1 4.17 0.5 4.11 0.9 4.11 0.9 Light 2 4.96 0.1 4.92 0.8 4.92 0.8 Light 3 6.24 1.4 6.25 1.7 6.25 1.7 Light 4 1.45 0.1 1.45 0.1 1.48 1.7 Light 5 0.39 0.2 0.39 0.7 0.41 5.5 Water boiler 0.48 0.5 0.48 0.5 100 Tower fan 0.21 50 0.17 17.9 0.24 66.2

• Supero

– All 146 light events detected, no false alarm, no miss – Comparable to Oracle

• Baseline: False alarms caused by hair dryer and bath fan

51 / 23

Rice cooker 0.98 2.2 1.01 1.2 1.01 0.8 Hair dryer 0.07 19.2 0.09 0.4 0.02 73.2 Fridge 11.8 3.7 11.8 3.2 11.8 3.2 Bath fan 0.12 N/A 0.17 N/A N/A Router 2.03 4.3 3.04 43.3 3.04 43.3 Average error 7.5 6.5 27.0

Impact of Sensor Placement

52 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 53 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 54 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 55 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 56 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 57 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 58 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 59 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris Deployment Sensor # result Red 11

• Green

7

• Blue

4

• Black

2

• Acoustic sensing

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 60 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris Deployment Sensor # result Red 11

• Green

7

• Blue

4

• Black

2

• Acoustic sensing

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 61 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris Deployment Sensor # result Red 11

• Green

7

• Blue

4

• Black

2

• Acoustic sensing

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 62 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris Deployment Sensor # result Red 11

• Green

7

• Blue

4

• Black

2

• Acoustic sensing

Impact of Sensor Placement

Deployment Sensor # result Red 6

• Green

6

• Blue

4

• Yellow

4

• Black

3

• Light sensing

Black 3

• 63 / 23

6 lights, exhaust fan, waste disposer dish washer, vacuum cleaner 6 TelosB, 11 Iris Deployment Sensor # result Red 11

• Green

7

• Blue

4

• Black

2

• Acoustic sensing

Conclusion

• Supero

– Multi-sensor fusion – Unsupervised event clustering – Autonomous appliance association

• Easy to install

– Considerable flexibility in sensor placement

• Real Implementation/Evaluation

– 5 environments (3 apartments, 2 houses) – Accurate, 7.5% average error

64 / 23

Q & A

65 / 23