non-intrusive appliance load monitoring: promise and practice · non-intrusive appliance load...
TRANSCRIPT
© Fraunhofer USA
Non‐intrusive Appliance Load Monitoring (NIALM): Promise and PracticeMichael Zeifman, Ph.D. and Kurt Roth, Ph.D.
March 1st, 2012
Building America Stakeholder Meeting
© Fraunhofer USA
What Is NIALM?
Non‐Intrusive Appliance Load Monitoring A.k.a. Non‐Intrusive Load Monitoring
Main breaker/circuit level Data acquisition (hardware) and disaggregation algorithms (software)
2
07:00 07:30 08:00 08:30 09:00 09:300
200
400
600
800
1000
1200
1400
Time
Pow
er, W
07:00 07:30 08:00 08:30 09:00 09:300
200400
Pow
er, W
Coffee Maker
07:00 07:30 08:00 08:30 09:00 09:300
5001000
Pow
er, W
Refrigerator
07:00 07:30 08:00 08:30 09:00 09:300
200400
TimeP
ower
, W
TV
© Fraunhofer USA
NIALM: Interest
Significant growth in U.S. granted patents
Both large (Belkin, GE, IBM, Intel,) and small (4home, PlotWatt, Enetics, Navetas) companies
3
1989-1995 1996-2002 2003-2006 2007-20110
0.5
1
1.5
2
2.5
3
3.5
Year
Pat
ents
/yea
r
© Fraunhofer USA
NIALM: Motivation
Energy saving potential Electricity constitutes >70% of residential primary energy consumption Whole‐house electricity consumption reduction of 5‐15%
More specific, actionable feedback has greater savings!
Potential for further savings for fault detection and diagnostics
Smart Grid – unique business opportunity Large smart meter deployment underway Data acquisition specs for NIALM can be incorporated into smart meters Non‐intrusive “manual” alternative to home automation
Time‐of‐use electricity pricing – valuing Demand Response Potential transition to time‐dependent pricing in Europe and North America
4
© Fraunhofer USA
The HEM Barrier: Complexity of HEM use and deployment
Complexity of HEM systems impedes their acceptance and use by consumers due to: Deployment complexity Complexity of use Complexity of information presented
NIALM if fully developed can potentially close this gap Deployment: May not require professional installation Complexity of use: Detect appliances and track their presence in time Complexity of information presented: Provide specific information for detected appliances (e.g., state, current
power draw, historical data and comparisons) Tailor energy‐saving tips to each household Increase credibility (accuracy) of information and tips
© Fraunhofer USA
How Does NIALM Work?
6
Data Acquisition Appliance Signal Features
Disaggregation Algorithm
NIALM
© Fraunhofer USA
Data for NIALM: Sampling Rate
7
Typical daily number of switching events in a household: up to ~ 103‐104 (Baranski 2004)
0.1 Hz – minimum sampling rate
22:35 22:40 22:45 22:50 22:55 23:00 23:05500
1000
1500
Pow
er, W
0 0.05 0.1 0.15 0.2-0.1
-0.05
0
0.05
0.1
Time, s
Cur
rent
, A
Inexpensive hardware: 1 Hz
More expensive hardware: > 5 kHz
© Fraunhofer USA
Variability of signatures
-540 -520 -500 -480 -460 -440 -420 -400 -380 -3600
50
100
150
200
250
300
350
-P, W
Freq
uenc
y
Coffee maker CFL 0 2000 4000 6000 8000 10000
-0.2
-0.15
-0.1
-0.05
0
0.05
0.1
0.15
“Signatures”
Current waveformPower
Conventional NIALM: Challenges
No distinct signatures (overlap, missing/ corrupted data, multi‐state‐, permanent‐, and variable power devices)
Matching signatures with appliances
Low
frequency
High
frequency
© Fraunhofer USA
Basic Method (MIT)
Commercial Product:
SPEED (Single Point End‐use Energy Disaggregation) by Enetics
Requires professional hardware installation
Capable of NIALM of large appliances
Requires training
$1,300
Hart (1992)
© Fraunhofer USA
Advanced method (Patel) Commercial Products:
In development by Belkin(since 2010)
In development by Intel (since 2009)
Major challenges to overcome:
Distinctive signatures
No‐SMPS appliances
Power‐line communication
Power draw estimation
Cost
Basic idea 1: appliance connected to a socket induces noise (electromagnetic interference) in another socket
Basic idea 2: switch mode power supplies (SMPS) generate distinctive noises
Patel et al. (2007)
© Fraunhofer USA
Fraunhofer Approach to Low‐Frequency Data
Principles: Appropriate to major appliances Detection and use of transient events to obtain signatures Power signature statistics (not unique) Time on/off statistics (self‐learned) in addition to power signatures Probabilistic algorithm that optimally decodes appliance states Currently applied to on‐off appliances, but can be extended
Advantages: Can work with a 3rd‐party data, potentially including smart meter data High accuracy for overlapping appliances Robust to missing and/or corrupted data Easier appliance matching – can be based on both power and time of use
© Fraunhofer USA
Our Approach to Low‐Frequency Data: Scheme
Data are collected for ~ two weeks Negative changes of power are clustered and matched to positive changes Each cluster can correspond to appliance Cluster signature (power, time) are statistically characterized Obtained statistical distributions can be used to disaggregate new data Clusters need to be matched to real appliances
DataClustering & Distribution Estimations
Distributions of power & time features
Disaggregation algorithm
© Fraunhofer USA
Our Approach to Low‐Frequency Data: Some Results
Real household publicly‐available data (Kolterand Johnson, 2011)
9 appliances, 3 weeks of monitoring
Overall accuracy
Our method: 0.89
Other methods : 0.71
9.95 9.952 9.954 9.956
x 105
0
100
200
300
400
Time,s
Pow
er, W
SubmeteredEstimated startEstimated end
4 4.5 5 5.5 6 6.5
x 104
0
500
1000
1500
2000
2500
Time,s
Pow
er, W
SubmeteredEstimated startEstimated end
DishwasherRefrigerator
Accuracy in Kim et al. (2011)
© Fraunhofer USA
Our Approach to High‐Frequency Data*
Principles: Appropriate to both MELs and major appliances Detection and use of transient events Two or more “orthogonal” methods for waveform characterization Probabilistic fusion of orthogonal methods to enhance accuracy Requires ~ 50‐100 kHz sampling rate
Advantages: Enhanced classification accuracy (high true positive rate, at low false positive
rate) Can incorporate additional waveform characterization methods to further
enhance accuracy
*Supported in part by DOE through a subcontract from Lawrence Berkeley National Laboratory (LBNL).
© Fraunhofer USA
Our Approach to High‐Frequency Data: Scheme
Method currently requires training (waveform templates for each MEL) After transient is detected, waveform of the change is obtained Two or more “orthogonal” methods implemented to match waveform to template
We use image processing method (Fourier descriptors ≈ harmonics) and multivariate statistical technique (Partial least squares, used in spectroscopy)
Results are fused together to classify the waveform
DataTransient detection
Waveform matching 1
Waveform matching 2
Waveform
Templates
Classification
© Fraunhofer USA
Our Approach to High‐Frequency Data: Some Results
Experiment 1: CFL lamp, CRT TV, halogen lamp, fan (no overlap in waveforms, power draw ~ 20 – 200 W)
Experiment 2: printer, CRT monitor, CFL lamp, incandescent lamp, fan (noticeable overlap in waveforms, power draw ~ 20 – 200 W)
Experiment 3: LCD TV, Pioneer receiver, netbook, vacuum cleaner, CFL light, pencil sharpener, hair dryer, night lamp, DVD, incandescent lamp (noticeable overlap in waveforms, power draw ~ 2 – 2,000 W)
© Fraunhofer USA
Conclusions
NIALM can enable significant energy savings
NIALM deployment low‐cost IF embedded in smart meter
NIALM is challenging!
Appreciable research, one product
Technical Challenges
Variability and overlap of “signatures”
Automatic appliance matching
Fraunhofer NIALM
Developed two new methods that outperform most other methods
Current status: proof of concept
Further applied research is required