In the first part of this two part special edition we focussed on studies of the impact of occupants' presence and behaviour on building energy and environmental performance, before closing with an article by Lam et al. that tackles the challenging task ofpassively detecting occupancy levels – a valuable commodity for the calibration and validation of any predictive model of occupants' presence. Picking up this thread, we open this second part of the special edition with an article describing a new model to simulate occupants' presence. We then go on to focus on new models of occupants' behaviour.
Liao et al. propose an innovative new agent-based model for the simulation of occupants' presence. This is inspired by a previous Markovian model which was designed and validated for simulating the presence of occupants of single offices. Liao et al. adapt and extend this model to handle multiple occupants of both single and multi-zone buildings. This offers the tantalizing prospect of a further extension to simulate occupants' transport within buildings, given a network description. In its present state, the algorithm has been shown to compare well with its predecessor for the single occupied single office case. Verification tests are good but less convincing in the multiple occupant cases, in part due to a lack of suitable model calibration and validation data. This brings us to a general observation: researchers' desires to develop and validate sound models of occupants' presence and behaviour are often thwarted by a lack of reliable data. As a community, we would benefit significantly from the more generous sharing of data! Perhaps Liao et al. and Lam et al. can collaborate to their mutual advantage here by way of example.
Up until now, the majority of the effort that has been invested in the development and validation of models of occupants' presence and behaviour – in particular of interactions with the envelope and internal light fittings – has focussed on workplace environments, particularly offices. This is partly a reflection of the availability of pre-existing data recorded within office environments and partly because of a continued bias in interest towards the more rigorous modelling of office environments, in particular from practicing consultants. It may also reflect the relative simplicity of office occupants' behaviours as compared with, say, dwelling occupants. This is because the range of activities available to office occupants is somewhat constrained compared with the home in which we may also work as well as sleep, cook, clean, entertain ourselves and others, etc. Now certain of these activities have a direct bearing on the behaviours of interest to us: the use of lights and appliances (both water and electrical) as well as of windows and shading devices. For example, the activity ‘cooking’ is rather likely to involve the use of a cooker, and possibly the opening of a kitchen window to evacuate pollutants. The cooker may also, for the braver amongst us, be used concurrently whilst we bathe. We are less likely to open the bathroom window whilst bathing than we are when we have finished, this time to evacuate excess moisture. It is important that we tackle this complexity to ensure that we have coherent models of occupants' behaviour; but how? In answer to this question, Widen et al. demonstrate an extremely powerful technique in which time use survey (TUS) data is used to calibrate a stochastic model of occupants' activities. Furthermore, by associating the activities of each household member with a location, they are able to predict their occupants' movement. But of more interest in the current context is that, by linking activities with water and electrical appliance use and by making plausible assumptions regarding theownership, duration of use and power-using characteristics of these appliances, they are also able to estimate households' transient water and electricity demands. On this basis, they also demonstrate an approach for examining how certain of our occupants' activities may adapt (intensify or otherwise at particular times of day) in response to certain stimuli. For example, we may perform certain activities more intensively to make better use of cheaper off-peak electricity tariffs. This is particularly interesting as we move towards the realization of smart grids, which integrate a greater proportion of renewable energy technologies!
Another exception to the hitherto trend to focus onoffice environments has been the work of Jun Tanimoto and his team at Kyushu University in Japan, who have in the past also studied residents' activities, relating these to the use of appliances as well as their use of Heating, Ventilating and Air-Conditioning (HVAC) equipment. In this most recent study, Tanimoto et al. develop a new model for predicting occupants' activation of HVAC systems within their homes and compare the accuracy of this model with measured data and with that of a previous model. Their former model treated occupants' switching of HVAC plant as an homogenous Markov Chain – predicting the transition probability in HVAC systemstate as a 2 × 2 Markov Chain, which is independent of time but dependent on the prevailing indoor climate conditions (globe temperature is the sole explanatory variable), expressed as a logistic function. As such the possible dependence of HVAC system control on occupants' presence and on time is not encapsulated by the model. To overcome this, the authors derive a new stochastic model based on a Multilayered Artificial Neural Network (MANN) in which the input layer, which consists of nine explanatory variables (groupings of time of day, whether the current day is a weekday, a comfort indicator and density of presence), is related to the output layer (on/off probability) via a hidden layer, through a set of weights; these weights being learnt using an empirical training dataset. This model better reproduces the dynamics of occupants' control of their HVAC system, and this is reflected in an improved predictive accuracy as compared to the previous model. There is, however, scope for rationalizing the number of explanatory variables in this model and, asthe authors point out, there is a need to extend the training dataset in order to achieve a more robust and generally applicable model.
As a final foray into behavioural modelling in the residential context, and as an example of the mutual advantages of data sharing, Schweiker et al. develop models of window opening behaviour for occupants of both Swiss and Japanese residential buildings and then challenge their predictive accuracy by cross-validation and blind verification. In each case, part of the local dataset was used as the training dataset to calibrate a model to predict the remainder of that dataset, for which predictive accuracy was judged. Then the entire dataset was used as the training dataset to predict the performance of the other dataset, for which predictive accuracy was again judged. From this, it was found that the model derived from the Swiss dataset reproduces well the behaviour of the associated occupants. The model derived from Japanese data produced less-convincing predictions of its occupants' behaviour; but this model performed rather well when compared with the behaviour of Swiss occupants. Thisis thought to be a fortuitous result: behaviours attemperatures considered to be relatively extreme inSwitzerland correspond well with those same temperatures that are considered to be relatively temperate in Japan. On the other hand, the Swiss model was derived from situations in which windows are not closed and air-conditioning (AC) units switched on whilst temperatures are high, as in Japan, and this particularity was expressed in its relatively poor performance: it predicts poorly Japanese occupants' behaviour when their temperatures are considered to be relatively extreme (when they close their windows and turn the AC on). It seems then that, in line with common sense, models should be applied within a clearly defined range of applicability and notto contexts which have specific local behaviours that may not be reflected in a given model. But when amodel is used within its range of applicability, predictions may be highly robust. This is demonstrated in this paper by the comparison of a model derived from Swiss residential buildings to predict behaviours of Swiss office building occupants and vice versa: the observed deviation is within that arising from differences in behaviour between different occupants!
This special edition is testimony to the considerable advances that have been made in our understanding ofoccupants' presence and behaviour and the corresponding impacts on buildings' energy and environmental performance. But this by no means implies that our work is done. Far from it! Some examples of where further research would be particularly welcome include:
| • | further development of models of occupants' activities using a wider range of TUS data, | ||||
| • | coupling of models of occupants' (or rather agents') activities with those of their behaviour; in particular with respect to their use of electrical appliances, | ||||
| • | further development of models of occupants' behaviour for a broader range of buildings and climates, | ||||
| • | the posting of new model calibration parameters, along with a description of their basis, to a public database, | ||||
| • | development of a coherent model of agents' internal movement + activity + behaviour … the list goes on. | ||||
One thing appears to be clear though: to carry out this research programme in a reasonable timeframe, we, as a community, should better coordinate our efforts and more generously share our data!
In conclusion, we would like to sincerely thank the editors of JBPS for according us this special edition on‘Modelling Occupants’ Presence and Behaviour', a subject for which we both have a passionate interest. We would also like to warmly thank the contributors to this special edition and the many reviewers who have helped to shape it, for their wisdom and their time.