Transcript Document

Demand Response
Verification (DRV)
Outline
• DRV Use Case Overview
• Non-Intrusive Load Monitoring (NILM)
• Compression Algorithm
customer
web portal
AMI
headend
Internet
load shed request
verification
smart
meter
HVAC
customer
network
DR gateway
fridge
plugin
hybrid
pool
pump
solar
panels
AMI
headend
verification
smart
meter
Link from meter to AMI headend
•
Frequency: twice year ~ once per 15 min
•
Data: Total wattage
Limited resources
•
Bandwidth
•
Processing power of meters
•
Memory/Storage of meters
AMI
headend
Assumption
Power company has sufficient processing power to
analyze large amount of load data
verification
smart
meter
Tradeoff: Accuracy vs Transmission Speed
•
Accuracy - Amount of useful info received at
AMI headend
•
Transmission Speed (local preprocessing)
•
Extracting info from data
•
Compressing data
AMI
headend
smart
meter
Proposed Method
Amount of info
at AMI headend
Transmission
Speed
Local
Processing
Time
Raw data
Complete
Very slow
Fast
Compression (loss rate)
Large
Slow
Medium
Transmission Frequency =
Func(load)
Medium
Fast
Medium
Fit load data with standard
distribution function
Small
Fast
Medium
Make local decision,
transmit “yes/no”
Small
Very fast
Slow
1. Smart meters are not that trusted yet - lack proper authentication
2. Power company has well-established ways to analyze large amount of load data
Non-Intrusive Load
Monitoring
• Developed in 1982 at MIT by George W.
Hart
• While looking at load data for a
photovoltaic study, the research team
noticed that on/off events for major
appliances in the home could be read
visually
The Idea
• Individual On/Off events of high power
appliances are easy to detect
Improvements
• Wanted the system to be able to
recognize individual loads based on the
aggregate data
• Determined that real power alone
would not give enough information
about which appliance was turning
on/off
• But we are looking at AC power….
Real vs Reactive Power
• AC Power is made up of AC current and
AC voltage
• Each is a sinusoidal wave that
oscillates at some frequency
• Recall that power is calculated by the
simple equation P = I*V
Real Power
• In phase I and V yields real power:
Reactive Power
• Out of phase I and V yield reactive power:
Improvements
• Using the real and reactive power gives a
two dimensional plane to identify appliances
The Algorithm
• This observation led to a simple
algorithm to ascertain that load in
the system with the following steps:
1.
2.
3.
4.
5.
Edge Detection
Cluster Analysis
Cluster Matching
Anomaly Resolution
Appliance Recognition
Edge Detection
• Analyze the incoming data for transitions
Cluster
Analysis/Matching
• Group like transitions together
• Match the on/off transitions that appear
similar
Anomaly Resolution
• For cases that do not match known patterns,
analyze the waveform for the possibilities of
multiple on/off transitions for the net change
Results
• Simple algorithm
has a high
probability of
identifying major
appliances in
residential settings
Results
• Pros:
• Measurement data was simple
• Small amount of data
• Cons:
• Algorithm had difficulty identifying low power
appliances uniquely
• Algorithm could not identify clustered systems
• Systems that needed to turn on slowly sometimes
passed edge detection
• The power signatures of appliances needed to be known
ahead of time for the appliance recognition
Recent Work: Improving
Granularity
• In order to increase the ability to discern
between individual appliances, more
detailed appliance fingerprints are required
• The strategy in more detailed system is to
look at harmonic properties of load data
• This gives more granularity at the expense of
needing more detailed measurement and
more processing power
Recent Work: Clustered
Systems
• Clustered systems presents a problem
• Examples
• Light bulb in a refrigerator
• Automatically Defrosting Refrigerator
• Multistage Light bulb
• Deal with this by introducing state based
recognition using finite state machine
models
• Modern methods can find clustered loads,
but introduces even more computational
complexity
Recent Work: SelfLearning
• All these methods depend on pre-determined
list of appliances and load patterns
• These have been initialized ahead of time at
installation of the system
• Would be nice if this was self-learning
• Solutions are being researched that use neural learning
algorithms to create the appliance load data
• This introduces even more computational
complexity
Current Status
• There is a lot of research in making NILM
extremely accurate
• Papers report results accurate down to
individual 10 watt light bulbs
• These algorithms are able to deduce power
drawn from clustered sources and systems
that are slow to ramp up to power
• The algorithms are also self learning so no
initial setup is required
What we really need….
• If the load data is sent to a processing
station, computational constraints are
not as severe to complete the NILM
• The big difference is that we have the
following limitations due to meter
constraints:
• We may only have real power
• There is a limitation on the amount of data we
can take due to BW issues
This is Ideal
• The fact that we may* only have real power lowers
granularity
• We are interested in turning off large appliances
• There is an inverse relationship between the size of
the appliance and the difficulty of detecting an
on/off transition
• Thus, the loss in granularity due to not having
reactive power may not be a problem
• Additionally, we may not care as much if we cannot
distinguish between two different 1000 Watt
appliances
* There is an initiative to include reactive power measurements in smart meters
If Only…..
• Classic NILM can be used to analyze
major appliance use
• In order to do this every major change
in power level would need to be
reported
• This could be problematic over low BW
links
So much data!
40 steps
Let’s Floor Some Values
(20)
12 steps
Let’s Try Again (40)
8 steps
Time To Compare
Instead of Flooring…
• Rounding
• Create a point based on input and
change values only if above some
threshold
• Take a rolling average (has to be used
in conjunction with other ideas)
Threshold (±20) Vs Full
10 steps
Time To Compress
• Deflate (e.g., gzip, zip, png)
– LZ77
• Blah blah blah blah blah! => Blah b[D=5,L=5]lah blah blah! =>
Blah b[D=5,L=18]!
– Huffman (Prefix) Encoding
A 16
B 32
C 32
D8
E8
(source: gzip.com)
Other Compression
Algorithms?
• Lossless
– LZMA/LZO (hash chains, binary trees and
Patricia tries)
– Bzip2 (effective, but slow because it has 9
steps)
• Lossy
– Discrete cosine transform (audio/video)
– Vector quantization (finds centroids)
(source: wikipedia.com)
Discussion