ZEMCH 2015 - International Conference Proceedings | Page 244
1 Introduction
Different studies and organisations claim that buildings are one of the major sources of carbon
emissions. In fact, it is been estimated that the energy consumptions from buildings in the UK
reaches values over 40%(de Wilde et al., 2013). In order to comply with the national and international CO2 reduction targets set, various solutions have been adopted for energy saving in buildings in recent years, such as improvements in the thermal insulation or renovation of windows,
boilers etc. In addition to this, recent studies (Haldi and Robinson, 2009)(Azar and Menassa, 2012)
show that occupants have a large impact on energy consumption in building. Occupants not only
generate heat and moisture through their natural metabolism, but also their behaviour patterns
will determine the use of the different building systems and consequently the final energy (Teixeira et al., 2010).
Due to all these considerations, we must try to find the best ways to automate BMS such as lighting or heating efficiently, so as to maintain the real user needs, but using only those resources that
are strictly essential for the actual occupation of the building.
It is evident that the indoor parameters such as ventilation rates, indoor temperature or lighting
will be different depending on the occupant activity performed (e.g. sleeping will require little
lighting while shower will require high rates of ventilation to eliminate moisture). Therefore, by
knowing at all times the current activities being performed, we will be able to use that information to regulate systems adequately.
To carry out these tasks we will need to collect information from users and indoor parameters and
also we will need to learn from that data to identify the patterns that occur for each activity. For
the former, we can collect information through ambient sensor readings (sensors such as motion,
temperature or contact sensors are ubiquitous, inexpensive and easy to deploy and handle). For
the latter, machine-learning (ML) based modelling approaches have shown a great potential to
learn patterns from sensor data and recognise daily activities.
In this work, we propose a new activity pattern recognition model using machine-learning approaches to accurately identify occupant activities. We train the model based on two public available datasets (Kasteren, 2011)(Cook et al., 2013). We constructed our model using a multi-class support vector (Burges, 1998) and compared its accuracy against other existing approaches namely
HMM (Rabiner, 1989) and kNN (Indyk, 1998).
The rest of the paper is organised as follows: Section 2 reviews the previous related work in the
field. Section 3 explains the methodologies used in our experiments by firstly introducing the
modelling techniques, the datasets used for model training and validations. Section 4 presents
and discusses the results of the experimental evaluation. Finally, Section 5 concludes the work
and highlights the future works.
2 Previous related works
Occupant behaviour and activity modelling has been developed lately with the introduction of
multiple models to capture human patterns in order to create novel adaptive systems to regulate
BMSs accordingly. Different machine learning approaches have been used as in Page’s model (Page
et al., 2008) where an HMM based algorithm simulated presence and absence of people in an office building using motion sensors. Other models attempted to simulate occupancy by modelling
242
ZEMCH 2015 | International Conference | Bari - Lecce, Italy