Matrix Factorisation for Scalable Energy Breakdown
Nipun Batra, Hongning Wang, Amarjeet Singh, Kamin Whitehouse
IIIT Delhi, University of Virginia
To create an energy breakdown for million homes. Can save up to 15% on energy bills.
Goal
Approach
1I. Source separation
1. Appliance submetering
Our approach can be more accurate than alternatives, with lower cost. Thus, more scalable.
Results
Web application
Alternative approaches
Existing solutions require hardware in every home, so cost scales linearly with the number of homes.
Signal
separation
Smart
meter
$20
$80
$15
Our approach can produce an energy breakdown without installing new hardware in every home
Key insight: Much of the variation in energy data across buildings occurs along a relatively small number of dimensions
| | | |
| $ 80 | $ 10 | $ 40 |
| $ 70 | $ 60 | $ 10 |
| $ 20 | $ 30 | $ 80 |
… | … | … | … |
| $ 30 | $ 40 | $ 50 |
Appliance
sensor
Our web application can potentially provide energy breakdown to 60 million homes
Step 1: Add easy to collect monthly bills in the matrix. Historical bills add more value.
Step 1I: Submeter small #homes (train) to create matrix X
Step III: Perform non-negative matrix factorisation. X~AB. Include static information, such as area of homes, to guide factorisation.
Step IV: Predict energy breakdown from factors.
| | | | |
| $ 40 | $ 10 | $ 40 | $ 500 |
| $ 70 | $ 60 | $ 10 | $ 400 |
| ? | ? | ? | $ 300 |
… | … | … | … | … |
| ? | ? | ? | $ 500 |
Monthly bill
Train
Test