Unsupervised Learning Techniques for HVAC Terminal Unit Behaviour Analysis

In the pursuit of improved energy ef ciency, older and new buildings are being tted with Building Energy Management System (BEMS). BEMS can be used to extract valuable building data that can be further analysed to discover problems related to user comfort, building maintenance and energy wastage in buildings. The main focus of this paper is to demonstrate and effective method to remotely analyse and categorise the different Heating, Ventilation and Air-Conditioning (HVAC) Terminal Unit (TU) behaviours using BEMS data. Using a data-driven, unsupervised learning strategy to identify anomalous behaviours enabling noti cations to the building manager regarding faulty TUs can go a long way in providing energy savings and improving building performance. A novel feature extraction method based on event discovery from TU data is proposed and applied to multidimensional data streams retrieved from a building based in the city of London. Further, X-means clustering has been performed over the extracted features to group the different TU behaviours. The clustering results, validated through established statistical methods, successfully yield several distinct TU behaviour patterns in addition to the outliers. The clustering behaviour has been further veri ed across daily and weekly TUs.