Article Preview
TopIntroduction
Residential and commercial buildings consume almost 60% of world's electrical energy. Around 54% of electrical energy in these buildings are used for Heating, Ventilation, Air Conditioning (HVAC) and lights (Johansson, 2012). Smart building initiatives aim to reduce energy consumption by optimizing appliance usage, based on occupancy and user preference. Occupancy information is crucial to enable energy optimization, especially room-level occupancy data. PreHeat and TherML, similar to the commercially available Nest thermostat, uses occupancy information to optimize heating system energy consumption (Scott, 2011; Koehler, 2013). Furthermore, it can also be used to enable contextual personalized services such as child monitoring, intrusion detection, and elderly care.
Currently, room level occupancy data can be obtained using additional infrastructure, for example, passive infrared sensor (PIR sensor) (Scott, 2011), camera, radio frequency identification (RFID) tag, or Bluetooth beacon. However, most of the existing households does not have such infrastructure in their house. Adding them will be an expensive and cumbersome to maintain. Indirect occupancy sensing aims to estimate occupancy using existing infrastructure and is easier to adopt in real life. WiFi-based localization is a possibility, but it requires extensive setup and training to map rooms and WiFi signals (Balaji, 2013). On the other hand, energy-consumption information is also available from smart meters which are already installed in buildings. In this paper, we try to answer the question, using smart-meter data, can we identify occupants’ location in a house accurately? Unlike WiFi-based localization, this approach only requires mapping of appliances and rooms, which remain constant most often.
Several challenges exist to obtain room-level occupancy information using just the smart-meter data. First, smart-meter data needs to be translated into appliance-level energy data. From this appliance-level energy information, we need to distinguish their state i.e., on or off. The translation from smart-meter data to appliance-level energy data must be accurate to obtain correct set of appliances and their states. Another challenge is to distinguish foreground and background appliances. Foreground appliances -- loads that are actively controlled by users’ actions, while background appliances -- loads that typically run without user intervention (Iyengar, 2016). For example, a refrigerator being turned on does not necessarily mean somebody is in the kitchen, or even home. By removing background loads from consideration, room-level occupancy can be predicted more accurately.
In this paper, we propose to use NILM-based energy analytics approach to obtain appliance-level energy data from smart-meter data. Furthermore, we also include house-level occupancy data into the feature set. We perform house-level occupancy prediction, using both a simple threshold mechanism and a classification technique, based on smart-meter data. This house level occupancy prediction data can help to distinguish cases where an appliance consumes power, but no one is actually in the building. We also use association rule to provide a smarter way to group appliances that are commonly used together. This way, we can get a consistent set of appliances that helps to indicate whether the room that contains such appliance set is being used or not. Finally, we use features mentioned above to train a multilabel classifier, which helps us to predict room-level occupancy. Several classification algorithms are evaluated to obtain the most suitable classifier for room-level occupancy prediction. The proposed system requires a one-time mapping of appliances present in the household along with their location.