I attended the second non-intrusive load monitoring (NILM) workshop, hosted at the University of Texas at Austin from June 3-4, 2014.
The first day was an academic session on NILM, run by Mario Berges and Zico Kolter. The second day was hosted by Pecan Street Inc., and highlighted the new WikiEnergy site providing the residential data they collect. From these sessions I took home a number of important messages:
- Interest in NILM algorithms and applications is growing. Conference attendance was up by 50% from 2012 to 90 participants, and is reflective of the growth of published work in this field.
- It remains difficult to create algorithms with easily reproducible results. Researchers use their favorite metrics, sampling rates, and data sets to the detriment of comparability. Tools like NILMTK, an open source toolkit under development by Jack Kelly, Oliver Parson, and Nipun Batra will help this effort.
- Most research efforts remain focused in the residential problem. The vast majority of papers and posters presented focused on single-family residences for their target application, and discussion always came back to these buildings.
- NILM provides a general framework for broader issue of “disaggregated single-point sensing”. Similar work is starting with water sensors in buildings, but this requires more expensive equipment and greater intrusion in test systems. There was discussion as to what other “flows” this could be applied to in order to make it more generically applicable and how active learning could improve performance. See paper by Eric Larson (paper)
- New NILM dataset from the UK is available. Jack Kelly and William Knottenbelt have authored (paper) a new disaggregated data set from a single family residence that has 6-second appliance-level data and 16kHz aggregate power data. The study consists of data from four homes. The paper describes their data collection system as well.
- Interesting algorithmic developments. (a) A prominent team in the Belkin disaggregation competition wrote a paper on their algorithm, which was claimed high accuracy in steady-state and transient event detection using only macroscopic real and reactive power features in an unsupervised setting, then afterwards performing data labeling. (paper by Barsim, Streubel and Yang). (b) A method on detecting and measuring the energy usage of a single appliance type in a home by leveraging a priori information was presented by Oliver Parson (read his excellent blog here). The example he used was a refrigerator, and the fact that you know at certain times in the day (namely at night), there is a high probability that the only operating appliance will be the refrigerator. (paper by Oliver Parson et al.)
- “Jamming” of energy disaggregation systems. If data privacy is a concern, energy storage systems can be optimized to both confuse disaggregation algorithms in an attempt to increase the privacy of building occupants, while also gleaning benefits from time of use pricing. (paper by Jinyun Yao)
- WikiEnergy Data Respository. Pecan Street Inc. has created this online repository in the hopes of becoming the “world’s largest database of customer energy research”. They are well on their way with, among other data sets, 1-minute disaggregated residential data from over 200 homes across multiple U.S. states.
- NILM conference papers: http://nilmworkshop.org/schedule/
- WikiEnergy: http://wiki-energy.org/
- Pecan Street: http://www.pecanstreet.org/
- Pike Powers Lab (affiliated with Pecan Street Inc.): http://www.pikepowerslab.com/
- UK-DALE data set: http://www.doc.ic.ac.uk/~dk3810/data/
- NILMTK: http://nilmtk.github.io/