Study on Adaptation of Model-Agnostic Meta-Learning Integrated Deep Q-Learning for HVAC Control

With the growing concern about climate change and the carbon neutrality caused by the energy sector, heating, ventilation, and air conditioning (HVAC) systems are significant contributor in energy usage, especially in space cooling, which leads to an annual increase in demand. This motivates research focused on designing high-performance HVAC control system using machine learning approaches because they have shown their proven capability and performance. However, the control methods frequently depend on a specific data source, such as locations, building types, and environmental conditions. This reliance on specific data can lead to inadequate performance when dealing with unforeseen situations, thus limiting the realistic use of these methods in real-world implementation, especially when the environment dynamically changes. Therefore, this study focused on developing a control model using model-agnostic meta-learning (MAML) and integrated reinforcement learning (RL) algorithms to address adaptation challenges. The result found that model-agnostic meta-learning with deep Q-learning (MAML-DQL) demonstrated inadequate control performance across environmental diversity and required improvement. The MAML-DQL was ineffective owning to its involvement with various buildings, while the algorithm's parameters were inadequately configured to cover all buildings throughout the training process. Therefore, to achieve a higher ability to cope with the versatile weather conditions and building types, the algorithm's parameters, including layer size, learning rate, and episode number, should be calibrated and optimized. This resulted in future directions aiming to evaluate and improve the performance of the MAML by performing hyperparameter tuning and further focusing on real-world applications.