Recent Advances in Algorithms for Energy Disaggregation · Recent Advances in Algorithms for Energy...

Preview:

Citation preview

Recent Advances in Algorithms for Energy Disaggregation

J. Zico Kolter MIT Computer Science and Artificial Intelligence Laboratory

BECC Conference, 2011

In collaboration with: Tommi Jaakkola, Matthew J. Johnson, Andrew Y. Ng, Sidhartha Batra

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

The Goal

• Determine breakdown of power given whole-home consumption (e.g. from smart meter)

Total Power

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

The Goal

• Determine breakdown of power given whole-home consumption (e.g. from smart meter)

Total Power Disaggregated Power

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Event-Based” Detection

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Event-Based” Detection

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Event-Based” Detection

Dryer? Fridge? Microwave? Noise? …

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Event-Based” Detection

Dryer? Fridge? Microwave? Noise? …

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Event-Based” Detection

Dryer? Fridge? Microwave? Noise? …

• The standard paradigm for energy disaggregation (e.g, Hart, 1992; Leeb, 1993; Laughman et al., 2003; Patel et al., 2006, Bergés et al., 2010; Gupta et al., 2010)

• Problems: doesn’t provide actual energy breakdown, difficultly handling long-term devices, can make physically impossible predictions (e.g., device comes on twice without turning off)

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

Coffee Machine

Dryer

Computer

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

• Helped by nature of AC Power

Curr

ent Time =

Coffee Machine

Dryer

Computer

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

“Power Packing” Detection

• A relatively recent approach (Suzuki et al., 2008; Liang et al., 2010; Inagaki et al., 2011)

• Provides true energy breakdown, can handle “always on” devices

• Problems: combinatorial (hard) optimization problem, extremely sensitive to noise/transients and devices that are too similar

Our work: Algorithms for “Joint” Energy Disaggregation

[Kolter et al., 2010; Kolter and Jaakkola, 2011]

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Interconnected Prediction Tasks

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Interconnected Prediction Tasks

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Modeling Interconnected Tasks

• Powerful methods for handling such tasks: probabilistic graphical models

• Exactly solving resulting optimization problem still intractable

Factorial Hidden Markov Model [Ghahramani and Jordan, 1997]

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Modeling Interconnected Tasks

• Powerful methods for handling such tasks: probabilistic graphical models

• Exactly solving resulting optimization problem still intractable

Factorial Hidden Markov Model [Ghahramani and Jordan, 1997]

• Our method presents an approach for approximately solving such optimization problems; based on methods from convex optimization, probabilistic inference

• Incorporates both the “event-based” features and the “power packing” features, classifies all energy jointly

• Tractable for large numbers of devices, many states for each device, and long time periods

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance Total Power

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance True Disaggregation

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance Our Method

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance Over Two Weeks

Device (Circuit) Our Method Previous Approach

Microwave 97.5% 94.8%

Bath GFI 15.0% 80%

Bath GFI 95.1% 0%

Bath GFI 94.7% 37.4%

Electronics 41.6% 39.4%

Kitchen Outlets 37.5% 9.7%

Furnace 91.7% 18.5%

Kitchen Outlets 45.2% 12.5%

Washer / Dryer 98.8% 89.3%

Total 87.2% 44.7%

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Performance Over Two Weeks

Device (Circuit) Our Method Previous Approach

Microwave 97.5% 94.8%

Bath GFI 15.0% 80%

Bath GFI 95.1% 0%

Bath GFI 94.7% 37.4%

Electronics 41.6% 39.4%

Kitchen Outlets 37.5% 9.7%

Furnace 91.7% 18.5%

Kitchen Outlets 45.2% 12.5%

Washer / Dryer 98.8% 89.3%

Total 87.2% 44.7%

Publicly Available Data and the Behavior / Algorithms

Interplay

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

REDD: A Publicly Available Dataset

• Reference Energy Disaggregation Data Set • 16.6 kHz whole-home current

and voltage • 1 Hz power consumption for

each circuit (Powerhouse Dynamics eMonitor)

• 1 Hz power consumption on select outlets (Enmetric PowerPort)

Initial release (6 homes, ~2 weeks) available at: http://redd.csail.mit.edu

[Kolter and Johnson, 2011]

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

REDD: A Publicly Available Dataset

• Reference Energy Disaggregation Data Set • 16.6 kHz whole-home current

and voltage • 1 Hz power consumption for

each circuit (Powerhouse Dynamics eMonitor)

• 1 Hz power consumption on select outlets (Enmetric PowerPort)

Initial release (6 homes, ~2 weeks) available at: http://redd.csail.mit.edu

• >12 GB of data, per home, per day • Compress using greedy approximation

to total variation regularization

[Kolter and Johnson, 2011]

J. Zico Kolter | Recent Advances in Algorithms for Energy Disaggregation

Behavior and Algorithms

Behavioral Studies

Algorithmic Advances

Recommended