Natdanai Kiewwath. Development of an artificial intelligence agent to optimize energy efficiency in air conditioning system. Doctoral Degree(Knowledge and Innovation Management). Chiang Mai University. Library. : Chiang Mai University, 2026.
Development of an artificial intelligence agent to optimize energy efficiency in air conditioning system
Abstract:
Electricity consumption from residential air-conditioning systems is an important issue in tropical countries, particularly Thailand, where air-conditioner use continues to increase due to hot climatic conditions, urban expansion, and the need for indoor comfort. However, conventional residential air-conditioner operation still commonly relies on fixed temperature settings and time-based maintenance practices, which may not respond effectively to dynamic environmental conditions such as temperature, humidity, and PM10 concentration. In addition, many intelligent control studies focus mainly on energy efficiency or thermal comfort, while the maintenance-aware dimension and the systematic transformation of structured explicit knowledge into intelligent agent design remain limited. This dissertation therefore addresses the problem of developing an intelligent agent for long-horizon decision-making that balances energy efficiency, user thermal comfort, and maintenance awareness in residential split-type air-conditioning systems.The research methodology began with the synthesis of structured explicit knowledge from academic literature, user manuals, maintenance manuals, and related technical sources. This knowledge was used as the foundation for a knowledge-to-agent transformation framework, in which relevant knowledge was converted into the agents state representation, action space, reward logic, environmental assumptions, and evaluation criteria. The air-conditioning control problem was then formulated as a long-horizon decision-making problem under a Markov decision process framework. The proposed agent was developed using an Enhanced Deep Q-Network that integrates several stability-oriented learning mechanisms, including Double Q-learning, dueling network architecture, prioritized experience replay, multi-step returns, Bayesian-style regularization, and entropy-aware exploration. The evaluation was conducted in two stages: a diagnostic benchmark using LunarLander-v3 and an applied 365-day simulation of a 15,000 BTU residential split-type air conditioner driven by Thai weather and PM10 data.The results show that the proposed agent under the RW4 reward-weight configuration reduced annual electricity consumption from 5116.22 kWh to 4440.03 ± 37.50 kWh, corresponding to an energy saving of 13.22% compared with a fixed 25 °C baseline setting. The learned policy also demonstrated adaptive control behavior under changing environmental conditions and reflected maintenance-aware tendencies through the consideration of PM10 and filter health. The main academic contribution of this dissertation is the proposed knowledge-to-agent transformation framework for designing an intelligent agent for long-horizon multi-objective decision-making. Its practical contribution lies in demonstrating the potential of reinforcement learning for sustainable and energy-efficient residential air-conditioning control. Nevertheless, the study is limited by its simulation-based evaluation, the use of a specific residential air-conditioning configuration and environmental profile, and the simplified representation of equipment degradation. Future research should extend the framework through real IoT-based validation, explicit equipment health modeling, and deployment in real residential environments.