Verena Marie Barthelmes

Office workers respond to a variety of environmental and contextual stimuli, which can influence their comfort and behaviour. Although recent models address the unpredictability of human–building interactions, they often neglect multi-domain factors and user motivations. A holistic understanding entails environmental, personal and contextual influences. The use of direct occupant feedback can reveal perceptions and drivers of behaviour that physical data alone cannot capture. This study introduces a multi-domain and mixed-method data collection framework for open-plan offices entitled eCOMBINE. This framework integrates both subjective and objective data across domains to better capture the factors influencing occupant behaviour and perception, particularly in buildings with operable windows and blinds. Central to this framework is the ‘OBdrive’ mobile application, which records self-reported motivations for window, blind and light control actions. The framework was implemented in two Swiss case studies, involving 44 participants. Results demonstrate the framework’s feasibility and participant acceptance. Benefits arise from embedding comprehensive data collection within research workflows. The eCOMBINE framework can support future research and inform more effective and efficient building design and operation by investigating user motivations, adaptive actions, health impacts, behaviour-driven simulations and optimal sensor set-ups. PRACTICE RELEVANCE Open-plan offices merit specific attention in energy-related behavioural research as their occupants typically have diverse personal needs and preferences that need to be met in a shared environment. The eCOMBINE framework adopts an integrated approach to study the relationships among combined indoor environmental quality, occupant behaviour, global environmental comfort and building energy use. By capturing multi-domain environmental conditions alongside subjective perceptions, motivations and contextual variables, eCOMBINE enables research into human–building interactions and facilitates integrated analysis into post-occupancy evaluation. These insights can support the development of more accurate and human-centred behavioural models, which in turn can help to improve the predictive reliability of building performance simulations. Such models can be used to optimise office building design and enhance both indoor environmental quality and energy performance by considering the synergies and trade-offs across environmental domains, adaptive opportunities, individual preferences and workplace context.

Hierarchical models are introduced first as a natural extension of the generalized linear model introduced in Chapter 4, where the so-called first stage model is a member of the exponential dispersion family and the second stage model is Gaussian. The examples included window opening, ceiling fan speed and clothing insulation level. The estimation is done through the Laplace approximation, which opens up for much more general structures, this being non-linear or non-Gaussian and such models are discussed along with appropriate tools that include automatic differentiation, again illustrated with examples. The chapter concludes with models that can be estimated without the Laplace approximation using conjugated priors, again all presented models are supported by examples.

The chapter presents survival analysis as a way of modelling time to event in data, the method presented in the chapter is illustrated using vacancy of rooms. The chapter introduces the special features of survival data in particular the concept of censoring which is common in survival data. Censoring also implies that the usual statistics for summarising data are not so useful, and the basic object in survival modelling is the survival function, which is discussed throughout the chapter, starting with the non-parametric Kaplan-Meier estimator and properties of that, including uncertainty estimation and tests, is discussed using the example of vacancy of rooms. Emphasis is put on parametric survival models, i.e. Weibull, exponential and log-logistic models, that are introduced through the hazard function. The chapter concludes by introducing a class of mixed-effect survival models.

State space models are introduced as a natural extension of linear time series models where the phenomena of interest are not observed directly, e.g. the phenomenon of interest is CO₂ production, but the actual observed time series is CO₂ concentration. The important concept of filtering is introduced and in particular, the Linear Kalman Filter is introduced and the resulting likelihood is given such that parameters can be estimated. The ARIMAX models presented in a previous chapter are given in state space form. The method is introduced using time series of CO₂ production. The chapter is finalized using modelling of window position (open or closed) using a generalized state-space model based on the exponential family of distributions.

In occupant modelling the interest is often more on the population of occupants, rather than the individual occupants, and this calls for the so-called mixed effect or random effect models. The chapter introduces linear mixed effects models as a natural extension of the general linear model. The general approach throughout the chapter is to describe the correlation structure induced by the model and hence formulate the likelihood. The chapter discusses the REML estimator and nested random effects. The theory is illustrated using an example of clothing insulation level. The estimation of random effects is discussed using the projection theorem and again illustrated using the example.

Generalized linear models are introduced through the exponential family of distributions. The exponential family includes many distribution assumptions and the most common ones are treated through examples ranging from occupants' window operation to their clothing insulation level and choice of ceiling fan speed, and the appropriate distribution assumptions are discussed through these examples. Statistical tests and confidence intervals are discussed using the results from Chapter 2. A wide range of modelling possibilities are available in the generalized linear models, e.g. choosing the proper link function and including over-dispersion is discussed through relevant examples. Further model diagnostics are discussed throughout the chapter.

Do you have data on occupant behaviour, indoor environment or energy use in buildings? Are you interested in statistical analysis and modelling? Do you have a specific (research) question and dataset and would like to know how to answer the question with the data available? Statistical Modelling of Occupant Behaviour covers a range of statistical methods and models used for modelling energy- and comfort-related occupant behaviour in buildings. It is a classical textbook on statistics, including many practical examples related to occupant behaviour that are either taken from real research problems or adapted from such. The main focus is traditional statistical techniques based on the likelihood principle that can be applied to occupant behaviour modelling, including: General, generalised linear and survival models Mixed effect and hierarchical models Linear time series and Markov models Linear state space and hidden Markov models Illustration of all methods using occupant behaviour examples implemented in R The built environment affects occupants who live and work in it, and occupants affect the built environment by adapting it to their needs – for example, by adapting their indoor environments by interacting with building components and systems. These adaptive behaviours account for great uncertainty in the prediction of building energy use and indoor environmental conditions. Occupant behaviour is complex and multi-disciplinary but can be successfully modelled using statistical approaches. Statistical Modelling of Occupant Behaviour is written for researchers and advanced practitioners who work with real-world applications and modelling of occupant data. It describes the kinds of statistical models that may be used in various occupant behaviour modelling research. It gives a theoretical overview of these methods and then applies them to the study of occupant behaviour using readily replaceable examples in the R environment that are based on actual and experimental data.

The chapter introduces the most common time series models, namely the so-called ARIMAX models, with an emphasis on identification and estimation. The models are introduced as a generalization of the general linear model, but with emphasis on the identification tools needed, this being the autocorrelation function (ACF) either the theoretical ACF based on model structures, or the empirical ACF based on the available data. Seasonal structures (e.g. diurnal) are discussed and further model diagnostics are presented and finally predictions using ARIMAX models are discussed. All methods are illustrated using a data series of CO₂ levels in a room taking into account occupant actions (window openings).

A historical account for occupant behavior models and their applications is given, the chapter includes a brief overview of different types of data and data structures. This includes the way occupants interact with the built environment, uncertainty related to human perception and the resulting actions taken by the occupants. Different types of data and the necessary considerations of characteristics in order to build suitable statistical models are presented, with emphasis on suitable models for a given data structure. Guides for choosing the appropriate model structure is given from the viewpoint of the data at hand. All chapters are introduced including small hands-on examples of applications to specific data sets.

The chapter presents models for time series data with discrete outcomes (e.g. window open or closed), such data can be modelled with Markov chain models. The chapter presents some general properties of Markov chain models but focuses on estimation of the transition probability matrix, and statistical properties of the estimator using likelihood techniques. Some generalizations of the simple Markov chain models are presented, in particular, higher-order Markov models and time-varying Markov models are introduced along with guidelines on how to estimate parameters. The time-varying Markov models are introduced with an emphasis on how to include covariates such as time of day. All methods are illustrated using a dataset of window opening actions.