Diversity behind the meter – machine learning from household activities

Phil Grunewald, University of Oxford, Environmental Change Institute

As the conference theme states, consumers use energy “without knowing or wanting to know much about how energy is being delivered to them”. For consumers to genuinely move to the ‘heart of the energy system’, the reverse disconnect may also need to be addressed: generators and suppliers of energy know surprisingly little about what their customers need this energy for. This two-way gap in knowledge is becoming more problematic in low carbon systems, where the timing and flexibility of demand is critical [“Smart Power”. National Infrastructure Commission. 2016].

We present data from the first study of its kind exploring at scale what happens behind the meter, not just in terms of appliance use, but the actual activities that give rise to times of high or low consumption. A better understanding of these dynamics is important to develop more consumer focussed business models that reflect the significant diversity in usage patterns and their flexibility.

Time-use data has become widespread in attempts to better understand and model energy use [CREST Model. I. Richardson 2007, Sekar. Joule 2018], the temporality of demand [“Laundry, Energy and Time”. B. Anderson. 2016] and even intrinsic flexibility of demand [“Peak Residential Electricity Demand and Social Practices”. J. Torriti 2015]. However, time use data is collected without energy recordings and these studies have to make difficult inferences on how activities may relate to consumption at aggregate level [J. Ramirez. Energy and Buildings 161. 2018].

Our data collection method is a combination of high resolution household electricity recordings, with simultaneous app based activity reporting [www.energy-use.org]. The parallel collection of data allows new analytical tools to shine a light on the relationship between activities, socio-demographics, appliance ownership and the timing of electricity use. The finding presented here are based on a sample of over 8000 activity records. The collection of data started in 2016 and is ongoing.

Machine learning and advanced regression techniques have been applied to establish often complex relationships. Our paper gives specific examples of some of the activity patterns that most strongly relate to high peak time usage and other prevalent usage patterns (Figure ‘TUC_Watt’, others ongoing). The act of ‘having a cup of tea’ is especially insightful in this context. Use of kettles is clearly distinguishable in the load profiles and serves as a validation of the activity reporting accuracy (Figure ‘kettle_profiles’). More complex patterns are also presented. The approach allows new and refined clusters of consumers to be identified, including groups that are more or less likely to be able to engage in load shifting.

We conclude that the diversity of demand is an opportunity for better targeted business models to achieve greater demand response engagement. Without a detailed understanding of what energy is used for, policies are likely to be designed for ill defined ‘typical’ users, thereby missing commercial opportunities while at the same time failing many customers who could be engaged better. This understanding will help to ensure that consumers can indeed be at the heart of the energy system.