Estimating Building Consumption
Breakdowns using ON/OFF State Sensing
and Incremental Sub-Meter Deployment
Deokwoo Jung and Andreas Savvides
Embedded Networks & Applications Lab (ENALAB)
Yale University
http://enalab.eng.yale.edu
Nov 4, 2010
Deokwoo Jung
Sensing Loads on Electricity Network
Breaker Box
How to Estimate Electrical
Loads of Appliances ?
Electric Meter
Electricity Network
Electrical Outlet
Bed Room
Nov 4, 2010
Living Room
Kitchen
Deokwoo Jung
Electricity Energy Monitoring Systems
• Indirect Monitoring
• Direct
Monitoring
: Expensive
and brute-force
– Total
Load Disaggregation
+ Load Signature
Detection method
NALM (Hart.et.al): Nonintrusive Appliance Load Monitoring
•
Kill-A-Watt EZ
–
–
$45
Data display only
•
Watts up?
–
–
$100-$130
Data Logging
•
Watts up? .Net
–
–
–
$ 230
Internet enabled
Power switching
A ElectriSense (Sidhant et.all) :Single-Point Sensing
Using EMI for Electrical Event Detection and Classification
in the Home
Nov 4, 2010
Deokwoo Jung
Load Disaggregation Data Flow
High frequency
electromagnetic
interference
Event Detection
Edge detection
Partial Load
Information
Heat
at Electrical
outlets or
Power entry
point
Vibration
Light intensity
Voltage and
current
waveforms
e.g. Total Power
consumption
ON/OFF state
Load
Disaggregation
How do we compute the load disaggregation?
Nov 4, 2010
Deokwoo Jung
The Diverse Nature of Loads
Resistive vs. Inductive -> Short-term property
Stationary vs. Non-stationary -> Long-term property
NonStationary
Hard to estimate
energy breakdown
Washing
Machine
heater
Air
Conditioner
Long-term
property
TV
Laptop
Refrigerator
Electri
c
Kettle
Stationary
Bulb
Resistive
Nov 4, 2010
DVD
Player
Short-term
property
Dehumid
ifier
Water
Pump
Hard to measure
power consumption
Inductive
Deokwoo Jung
Our Approach: Energy Breakdown
per Unit Time
Instead of instantaneous measurements, use
average consumption over a time window
Actual Average Power
Consumption
Estimated Average
Power Consumption
Actual Power
Consumption Profile
Estimation Error
Estimation Period k-1
Estimation Period k
Estimation Period k+1
Example appliance: LCD TV
Nov 4, 2010
Deokwoo Jung
Problem Setup
Three Tier Tree
Network
Time
Appliance
Consumption
fluctuation
properties
 0 1 1 0 0 1 1 1 0 1
 1 1 1 1 0 0 0. 0 1 0


 1 1 0 1 0 1 0 1 0 1
Goal: Estimate the average power consumption for a time window
Select an appropriate time window to get the best estimate of
energy consumption
Nov 4, 2010
Deokwoo Jung
Prototype System Implementation
One Energy Meter and
ON/OFF Sensors
TED 5000 Monitor
BehaviorScope Portal
Consumption measurements
Appliance ON/OFF Information
Active RFID
Dry Contact Sensor
Nov 4, 2010
Deokwoo Jung
Main Idea
ON/OFF sequence of appliances occurs between the worst
(Perfectly Synch) and the best case (Perfectly Desynch)
appliance A
appliance B
Worst Case
Observed Binary Data
Best Case
Approach – Variant of Weighted Linear Regression
– Accounting for Diversity
• Design Optimal Weight Matrix, W
– Metric Driven Data Selection
• Regression data set is adaptively chosen according to active power
consumption property, stationary vs. non-stationary
• Using Prediction Metric for Estimation Error
Nov 4, 2010
Deokwoo Jung
Problem Formulation
Objective Function: Min W TotalPower- P1 X1  P2 X 2  Pn X n 
Sample
Index
On/Off state
of TV
On/Off state
of Microwave
On/Off state
of Lamp
The Average of
Power meter
measurement
(Watt)
1
0
0
1
59.3
3
2
0
1
1
369.3
3
3
1
0
0
120
1
4
1
0
1
Y
160
1
5
1
1
1
469
1
X
Solve Opt. Problem:
Nov 4, 2010
# of
samples
observed
ˆ  arg min (XWX)P  XWY
P
P0
Deokwoo Jung
Designing Weights and Selecting
Appropriate Time Window
•
Optimal Choice of Weight Matrix, W
No Weight
W
# of Samples
Unit Sum Matrix
# of Samples
 On State
All Appliances
Estimated Variance
Sum Matrix
Exact Variance Sum
Matrix
# of Samples
 On State Varianceof ActivePower
All Appliances
• Account for (Non-) Stationary Property
– Stationary Load : larger window of measurements is better
– Non-Stationary Load: small window of measurements is better
– Automatically select to use either of the entire estimation periods
(Cumulative Data) or only the current period (Current Data)
Nov 4, 2010
Deokwoo Jung
Evaluation - Case Study
A small electricity Network with single power meter
•
•
Collecting data from 12 appliances in one-bedroom Apt from Thu-Sat
A large variation of energy load
– the heater accounts for more than 60% of the total energy consumption
– the laptop consumed the least, less than 1% of the total load.
Dehumedifier
Ceramic Heater
Laptop
Dehumedifier
Ceramic
Top Freezer Refrigerator
Halogen40
Desk
LampHeater Microwave Oven
10
10 LaptopFluorescent Desk Lamp
LCD
TV(32")
Top Freezer Refrigerator
Halogen
Desk Lamp
Desktop+Monitor
Standard Bulb Table lamp
Compact Refrigerator
Two-Way Floor Lamp
Microwave
Oven
5
20Fluorescent Desk Lamp
5 Desktop+Monitor
2.5 Standard Bulb Table lamp
Compact Refrigerator
Two-Way Floor Lamp
0
0
0
4
Friday
Saturday
440 Thursday
460
480
500
1000
1500
20
30
40
25
LCD TV(32")
Top Freezer Refrigerator
Halogen Desk Lamp
20
20
20
2
Compact Refrigerator only
Energy Consumption, (kWh)
320
Energy Consumption, (kWh)
The number of samples, nk
5
Daily The
energy
consumption
ground truth
in one-bedroom
apartment
from an
hourly
energy consumption
ground
truth in one-bedroom
apartment
Histogram
of power
consumption
appliances
during
their Onbinary
state states
experiment
from
Thursday
to Saturday
4
The number
of meter
samples
observed
given
composite
from
an
experiment
fromof
Thursday
to
Saturday
x 10Dehumedifier
Ceramic Heater
Laptop
LCD TV(32")
10
0
60
1.5
2
15
1
10
0.5
5
Nov 4, 2010
6
0
38
40
42
Desktop+Monitor
20 + Ceramic Heater
Top Freezer Refigerator
10
10
10
0
0
500
1000
Standard Bulb Table lamp
40
0
32
33
34
Compact Refrigerator
40
0
100
120
140
Two-W ay Floor Lamp
20
20
20
10
0
30
0
0
0
0
100
120
140
Fluorescent Desk Lamp
20
80
100
Microwave Oven
20
1
0
All appiances are off
10
Compact
Refigerator +10Ceramic Heater
5035
8 10 12 14 16 18 20 22 0
100
40
2
4
6
0
65
70
150
0
200 100
75
8 10 12 14 16 18 20 22 0
2
4
6
200
250
300
350
300
8 10 12 14 16 18 20 22 0
2
4
6
Decimal
of(hour)
composite binary
states, xk
Thursday representation
Friday
Saturday
Time,
Deokwoo Jung
Evaluation - Case Study :
A small electricity Network with single power meter
•
Estimated hourly energy consumption profile of each appliance
– Average 10% of relative error
Nov 4, 2010
Deokwoo Jung
Performance over Estimation Periods
Average Active Power Relatve Error, (%)
• With different weight matrix
100
Opt. Data.Sel + Est.Var.Sum.Wgt (Algorithm Performace)
Opt. Data.Sel + Unit.Sum.Wgt
No Weight
Opt. Data.Sel + No.Wgt.
Oracle. Data.Sel + Exact.Var.Sum.Wgt (Lower Bound)
80
60
40
Algorithm
performance
Unit Sum
Weight
20
0
0
0.5
Lower bound
Nov 4, 2010
1
1.5
2
Estimation Periods, T est ( hour)
2.5
3
Deokwoo Jung
Performance over Estimation Periods
Average Active Power Relatve Error, (%)
• With different data selection schemes
200
Est.Var.Sum.Wgt + Opt. Data.Sel (Algorithm Performace)
Est.Var.Sum.Wgt + Cur. Data.Sel
Est.Var.Sum.Wgt + Cma. Data.Sel
Exact.Var.Sum.Wgt + Oracle. Data.Sel (Lower Bound)
150
100
Algorithm
performance
50
0
0
0.5
Lower bound
Nov 4, 2010
Current Data
Selection
Cumulative
Data Selection
1
1.5
2
Estimation Periods, T est ( hour)
2.5
3
Deokwoo Jung
Performance by Data Selection,
Weight Matrix, and Estimation Period
Average Active Power Relatve Error, (%)
The maximum, minimum, and average value of relative error of
active power consumption for all estimation periods with various
combination of weighted matrix and data selection schemes
250
Est.Var.Wgt
+ Cur.Data.Sel
200
No.Wgt +
Opt.Data.Sel
150
Est.Var.Wgt + Opt.Data.Sel
(Algorithm Performace)
Unit.Sum.Wgt
+ Opt.Data.Sel
Exact.Var.Wgt
+ Opt.Data.Sel
100
Est.Var.Wgt +
Cma.Data.Sel
50
Est.Var.Wgt +
Oracle.Data.Sel
Exact.Var.Wgt +
Oracle.Data.Sel
(Lower Bound)
0
Nov 4, 2010
Deokwoo Jung
Increasing Accuracy on Larger
Networks with Additional Meters
Topology 1
• How many power meters we need
and where should place them?
?
– Tree Decomposition Problem
• Depending on sensor duty cycles
y1
y0
y1
y0  y1
x1 x2
x3 x4 x5
0
y2
– Combinatorial Optimization Problem
• Use Stochastic Search Algorithm :
Simulated Annealing
• Cost function of Simulated Annealing
–
Topology 2
x1 x2 x3 x4 x5
y0  y2
y2
Unsynchronized
Synchronized
Evaluated against the initial solution,
x1 x2 x5
x3
• Z0=(1,1…,1) : Placing meters on all available electrical outlets.
Z 
 MSE( P | Z0 ) 
c ( Z)    
  (1   ) 

MSE
(
P
|
Z
)
Z


 0
Weight Coefficient:
#
of meters vs performance Estimation Quality
Nov 4, 2010
Node Efficiency
Deokwoo Jung
x4
Evaluation - Case Study 2:
A large scale electricity network with meter deployment
•
Performance evaluation by
increasing the number of Apt
units from 1 to 12
With a single power meter for a
large electricity network
Relative Error, (%)
•
150
100
50
0
0
20
40
60
80
100
120
The number of appliances
Meter Deployment by Algorithm
•
–
–
–
–
Compared by random deployment
For λ= 0.5, x10 in performance
Or reduce x 2~3 in # of meters
λ = 0  Single power meter
λ = 1 Full deployment
10
Random Deployment
B-SEND Algorihtm
6
Max
=0
RSS(P)
•
10
Mean
5
Min
=0.3
10
=1
4
=0.5
=0.7
10
Nov 4, 2010
3
0
2
=0.9
4
6
8
10
The Number of Electricity Meters
12
14
Deokwoo Jung
Conclusions and Future Work
Developed an energy breakdown estimation algorithm for a single
power meter and the knowledge of ON/OFF states
•
10% of relative error for 12 home appliances and a single power
meter
Developed an algorithm for optimally placing additional power meters
to improve estimation accuracy in large networks
•
Deployment algorithm can reduce 3-4 times of the number of
power meter for the simulation of 12 households
Future work:
- Experimental deployment on a Yale building in January 2011
- Handle incomplete binary state sensing
- Leverage history information and user inputs
Nov 4, 2010
Deokwoo Jung
Discussion & Comparison with
Related Work
•
•
•
The question on high frequency systems makes some sense. Assuming
that you can detect signatures, if the frequency of measurement is high
enough you may have enough information to computer itemized
consumption.
The key argument to make is that this approach could work today with
existing low-frequency meters. The central meter in a home only has to
same using 1Hz. Also, in the home, we may be able to do this without any
additional hardware by just completing forms on a GUI.
While we work out details for a journal version it is important to identify and
propose the next problem to solve on load disaggregation
Nov 4, 2010
Deokwoo Jung