Abstract
Air Conditioning (AC) has become an integral part of the modern lifestyle, with energy consumption of ACs contributing up to 40% of the total electricity consumption in buildings in Singapore. Not only is AC an energy guzzler, AC usage also contributes to global warming, spewing 100 million tons of carbon dioxide to the atmosphere annually. With AC energy demand expected to triple by 2050, there is a strong need to improve the energy efficiency of ACs. Traditional Building Energy Management Systems (BEMS) seldom integrate and process information from various components within and external to the building, leading to performance gaps between the designed and operational performance of ACs. With growing popularity and lower cost of Internet of Things (IOT) devices in buildings, it is now possible to collect vast amount of contextual data for AC management from various sensors placed in and around buildings, to augment AC control. Artificial Intelligence (AI) techniques are primed to assist in data integration and analysis in order to come up with appropriate control decisions. However, current usage of AI techniques lacks standardization, contextual understanding, and holistic consideration of factors. This thesis reports on the research in the field of Air Conditioning Load Management using AI techniques in order to make them holistic, self-adaptive, and generally applicable. The work consists of three topics, arranged from macro systems within buildings to micro systems in individual households. The first topic attempt to standardize AC load management by describing an ontology-based AC load management framework with an adaptive benchmarking module. The second topic considers a holistic set of factors for the modelling and control of zonal AC loads within commercial buildings, by modelling the zone of interest using a grey box method, and making set point decisions based on a fuzzy logic system. The third topic discusses the modelling and control of single unit AC loads within residential buildings despite their inherent uncertainty, by building a simulation environment using uncertainty-aware deep neural networks and a reinforcement learning algorithm for optimizing AC set point.