Information and communication technology plays an important role in addressing
the world’s energy problem. Networked digital electricity meters (so-called
smart meters), for instance, can provide households with real-time information
on their electricity consumption and thus help them to conserve energy.
Initial expectations on the saving potential of this technology were too
optimistic, however. In fact, recent pilot studies conducted under realistic
assumptions have shown that savings induced by plain electricity consumption
feedback are often significantly lower than many have originally expected.

In this dissertation, we take smart metering to a new level as we explore a
data analysis-driven approach to personalize energy efficiency services that
may be offered at large scale. An example for such a service is automated
energy consulting, which consists in automatically providing energy saving
recommendations to households by taking into account their appliance stock and
usage profiles. In addition, we provide the foundation for an electricity bill
that is tailored to the household as it shows the contribution of individual
appliances to the overall bill or compares a household’s consumption with other
households that have similar characteristics. Behavioral trials indicate that
such consumption feedback is potentially more successful in motivating
households to reduce their electricity consumption than plain consumption
feedback or generic energy saving recommendations.

One contribution of this thesis is the design, development, and evaluation of a
system that automatically estimates characteristics of a household (like its
socio-economic status, dwelling properties, and appliance stock) from the
household’s electricity consumption data. We evaluate our approach on real
world smart meter data collected from more than 4000 households over a period
of 1.5 years. Our analysis shows that inferring household characteristics is
feasible, as our method achieves an accuracy of more than 70% over all
households for many of the characteristics and even exceeds 80% for some of the
characteristics. For utilities, the system creates valuable customer insights
that - without having to perform costly and cumbersome surveys - help to run energy
efficiency campaigns more efficiently by targeting each household with the
adequate service (e.g., offering energy consulting for retired people and a
smart heating system, which automatically controls the thermostat based on
occupancy, for employed people). Furthermore, these insights can be used to
realize automated peer group comparisons on the electricity bill or in an
online portal.

Providing automated, household-specific energy saving recommendations requires
more detailed information about a household than its high-level
characteristics. In particular, it is important to know when individual
appliances are running and how much they consume. To avoid measuring each
appliance individually through a complex sensing infrastructure, we investigate
inferring this information from the overall electricity consumption measured by
a smart meter. To explore this concept (non-intrusive load monitoring, NILM),
we developed an evaluation framework and analyzed the performance of several
state-of- the-art NILM algorithms. To this end, we collected electricity
consumption data in six Swiss households over a period of eight months and made
it publicly available. Along with fine-grained smart meter data (collected at 1
Hz), our data set contains ground truth measurements of 47 selected appliances
and each of the household’s occupancy state. Our analysis shows that - through
the enhancement of an existing NILM algorithm - it is possible to achieve
recognition rates of more than 90% for some typical appliances. This is
sufficient for energy consulting scenarios; its practical use is limited,
however, since a training period is required.

Ultimately, deploying smart meters comes with a cost that - for some of the
households -  can be higher than the achievable savings given today’s energy
prices. Maximizing societal benefits thus requires a well-managed interplay
between (1) regulators, which define rules for smart meter deployments and set
penalties if saving targets are not reached, (2) utilities, which develop and
run energy efficiency campaigns, and (3) households, which should invest in
energy saving solutions or adapt their lifestyle in order to use energy more
efficiently. This thesis copes with this challenge as it shows how to utilize
Internet of Things technologies and machine learning methods to enable
personalized energy efficiency services that scale to thousands or even
millions of households. We develop methods, build open source evaluation
frameworks, and collect and analyze real world consumption data in order to
better understand residential electricity consumption and improve the effect
(and thus the value) of smart meter deployments and feedback mechanisms.