January 19, 2020

349 words 2 mins read

Predicting residential occupancy and hot water usage from high-frequency, multivector utilities data

Predicting residential occupancy and hot water usage from high-frequency, multivector utilities data

In EU households, heating and hot water alone account for 80% of energy usage. Cristobal Lowery and Marc Warner explain how future home energy management systems could improve their energy efficiency by predicting resident needs through utilities data, with a particular focus on the key data features, the need for data compression, and the data quality challenges.

Talk Title Predicting residential occupancy and hot water usage from high-frequency, multivector utilities data
Speakers Cris Lowery (Baringa Partners), Marc Warner (ASI)
Conference Strata Data Conference
Conf Tag Make Data Work
Location New York, New York
Date September 11-13, 2018
URL Talk Page
Slides Talk Slides
Video

In EU households, heating and hot water alone account for 80% of energy usage. If we could predict when a resident requires heating or hot water, we could then optimize energy use to meet these needs without the need for direct interaction by the householder, taking account of real-time price signals. This could deliver significant financial, societal, and environmental benefits. Cristobal Lowery and Marc Warner share a nonintrusive machine learning approach to predict the residents’ needs. The project, led by the UK’s Energy Technologies Institute, uses data from electricity and water utility meters together with internal humidity and temperature measurements from five residential properties. The data was recorded at very high frequency, up to 200,000 readings a second for electricity. This generates 3 TBs of data a month for a property. Hence, a fully exploratory approach to developing the data features is unrealistic. As such, the algorithm was structured using a more traditional modeling approach. A key consideration in the work was how to achieve data compression while minimizing information loss, which forced the team to extrapolate domain knowledge to unexplored territories. Another key consideration of the work was how such a system could be productionized given the significant data volumes, which pose both storage and memory challenges for a reasonably priced device. Cristobal and Marc outline a potential solution and approach, which uses cheap hardware to achieve the necessary compression, starting with a Fourier transform, which can be achieved through a relatively cheap chip.

comments powered by Disqus