A framework for learning activities of office occupants

Abstract

Energy consumption in buildings has a correlation with the activities of occupants. Buildings account for about 40% of the total energy consumption in many developed countries, making them the largest end-user consumer sector [1][2]. Non-residential buildings comprise 25% of the European building stock [2]. Average energy consumption in the non-residential sector is on average 280 Kwh/m2 which is 40 % greater than an equivalent for the residential sector [2]. Commercial buildings like offices (23 %) and wholesale and retail shops (28 %) constitute the major part of total energy consumption in the non-residential sector[2]. Hence learning and understanding occupants’ activities especially in a commercial landscape like offices is a complex process but with great benefits. The continuously changing patterns in data over time, ever generating new data types and dynamic streams of data make the task challenging as well as exciting. Understanding and analyzing the patterns in data produced as a result of human activities can lead to an increase in efficiency and higher performance resulting in lower energy consumption and more productivity of workers in the office environment [3]. The noisy raw data coming from multiple sensors installed at different locations in the office premises needs to be converted into some useful information and interpreted in an intelligent manner. This thesis investigates approaches to recognize occupant’s activities in office premises and predict whether any of these activities will occur in some time window and how long these may last. The contextual information from raw sensor data can be used for learning different types of common office activities. i.e., working or not working on a computer, reading a book using a lamp, presence, and absence from the room, preparing a meal using the microwave or making a hot coffee using a coffee machine in the kitchen. The sensors are selected keeping the privacy of the user in mind while deliberately not using cameras for video recording and capturing images. Our contributions can be grouped into a three-fold approach. Firstly, developing a framework which can address the problems of recognizing and predicting activities in real-time and off-line mode. Secondly, applying our framework on the real sensor data collected using a wireless sensor network in an office environment. Lastly, providing separate evaluating metrics for both problems of activity learning (recognition and prediction).