The ability to sense and predict occupancy - i.e. to establish when the 
residents are and will be in a building - represents a basic requirement for 
the energy-efficient operation of many building automation systems. In 
residential households, in particular, the absence of all residents allows a 
heating controller to automatically lower the temperature of the home, thereby 
saving energy that would have been otherwise wasted on heating an empty 
building. However, if the home has been thus allowed to cool, a boiler and 
heat distribution system need a non-negligible time to reheat the home to a 
comfortable temperature. Therefore, to avoid a loss of comfort, a heating 
control system also requires a sufficiently accurate prediction of when the 
occupants are going to return in order to trigger the heating at the right 
time. Since space heating accounts for a large fraction of residential energy 
use (e.g. 68% in the European Union member states), heating control systems 
based on occupancy sensing and prediction - often referred to as smart 
thermostats – play an important role in reducing energy consumption and 
carbon dioxide emissions, while at the same time ensuring occupant comfort. 

The objective of this thesis is thus to investigate how the two main 
computational components of a smart thermostat - occupancy sensing, based on 
sensors that typically exist in a residential environment, as well as 
occupancy prediction from historical occupancy patterns - can be used to 
automatically reduce the energy consumption of a heating system while trying 
to maximise thermal comfort. 

Current smart thermostats require the installation of dedicated hardware to 
sense whether the occupants are at home or away. This increases installation 
and maintenance costs and thus prevents widespread adoption of such 
potentially energy-saving solutions. To overcome this hurdle, we investigate 
the suitability of opportunistically using devices already existing in 
households to sense occupancy. This opportunistic sensing approach seeks to 
utilise available devices to replace or augment dedicated infrastructures. An 
example are smart electricity meters, which are mandated to be installed in 
many households worldwide. We hypothesise that the information contained in 
the electrical load of the household, as measured by the smart electricity 
meter, can be used to infer its occupancy. To verify this hypothesis, we have 
performed an extensive data collection campaign over seven months in six Swiss 
households to collect occupancy ground truth data as well as the aggregated 
and device-level electrical consumption of the households. Using this data, we 
employ supervised machine learning algorithms to infer occupancy solely from 
the households’ aggregated electricity consumption. We show that such an 
approach yields a classification accuracy of up to 94%. 

As soon as the occupancy sensing infrastructure detects that residents left 
the house, the temperature can be allowed to drop resulting in energy savings 
during this setback period. However, a reactive strategy cannot be employed 
upon the arrival of the occupants as it may take a considerable amount of time 
to bring the house back to a comfortable temperature. To avoid loss of 
comfort, occupancy prediction algorithms are used to predict the time of 
arrival of the occupants to determine the right time to start pre-heating the 
house. To analyse the performance of such prediction approaches we have 
derived occupancy schedules from a large, publicly available mobile phone 
location dataset. Using the schedules from 45 participants we show that 
current state-of-the art occupancy prediction algorithms achieve an accuracy 
around 85%, which is close to the theoretical optimum given by the 
predictability of the schedules (which in practice always feature some level 
of irregular behaviour). 

The accuracy of the occupancy prediction alone does not necessarily reflect 
the energy savings and comfort loss that can be achieved or caused by a smart 
thermostat. The actual savings depend upon the occupancy schedule of the 
household, the prediction accuracy, the weather conditions and the physical 
properties of the building. The final part of this thesis thus deals with the 
simulation of various heating scenarios to investigate the effect of a smart 
thermostat on the overall energy savings under different environmental 
conditions. To this end, we assess the overall energy expenditure in several 
building scenarios. Furthermore, we develop a new methodology to accurately 
assess the impact of the weather conditions on the energy savings. We show 
that building parameters result in a range of savings from 6% to 17%, while 
the savings in the 25% of households with the lowest occupancy are 4-5 times 
higher than in the quarter with the highest occupancy. 

The unifying theme of this thesis is to show how current technology, which 
already exists in many homes, can help to save energy without sacrificing 
comfort. For this purpose, we draw upon recent work in the distributed systems 
domain to access smart electricity meters and machine learning algorithms to 
derive occupancy data. We show how predictable occupancy schedules are and, by 
providing a simulation framework to evaluate different occupancy prediction 
algorithms, we seek to answer the question how much energy a smart thermostat 
can save.