Modelling Building Occupant Behaviour Using Hidden Markov Models

Abstract

This thesis deals with the development of building occupant behaviour models, using hidden state methods. Reducing greenhouse gas emissions and the use of fossil fuels is a global challenge. A significant part of the global energy demand can be ascribed to buildings. Consequently, major efforts have been undertaken to reduce emissions and the energy use in buildings. As a requirement, this assumes a proper understanding of how the energy is used. Building energy simulation tools attempt to estimate the energy use, in many cases even before the building is constructed. These tools utilise computer models of a building to simulate the energy demand, taking the influence of its physical properties, the weather conditions, but also the occupant behaviour on the energy balance of the building into account. While modelling the physical properties of a building is well understood, it has been acknowledged that models of the occupant behaviour are over-simplified. This is due to the lack of proper understanding of the variation of occupant behaviour and its drivers. However, occupants have a significant influence on the energy demand of a building. Hence, they are a key part in the estimation of energy use and greenhouse gas emission of buildings. Moreover, a thorough understanding of occupant behaviour can contribute to improved control strategies of heating, cooling and air-conditioning. Using occupant presence forecasts, it is possible to heat, cool and ventilate a building only when needed. This way, the indoor climate can be controlled in a way that minimises energy usage, yet satisfying occupant comfort and health requirements. This thesis addresses several aspects of occupant behaviour modelling, with a focus on occupant presence and window opening behaviour models. In particular, the use of hidden state methods is explored. This means, that information about occupant behaviour is inferred without directly observing it. Instead, other quantities such as environmental variables and any other available information sources are used. These methods are important, since occupants’ actions are seldom measured directly. Traditionally, occupant behaviour models are derived from data of field studies or controlled experiments in which rooms are equipped with monitoring devices that record the behaviour in question. These monitoring campaigns have been an important factor, especially for model validation purposes. However, the use of indirect methods, such as hidden Markov models, provides this research discipline with more possibilities, since datafrom a wide range of existing buildings without dedicated occupant behaviour measurements can be used for model development.