A deep reinforcement learning-based method for dynamic quality of service aware energy and occupant comfort management in intelligent buildings

Buildings are responsible for 30 % oof the world's energy consumption, about half of which is consumed by Heating, Ventilation, and Air Conditioning (HVAC) systems. Intelligent control of these can significantly reduce global energy consumption. However, these systems are also one the most important means of providing comfort to occupants within buildings. These two control objectives may conflict with each other, which means that reducing the energy consumption of the HVAC systems may lead to dissatisfaction among buildings' occupants. Thus, in most cases, to handle the trade-off, weights are assigned to these objectives which indicate their importance. The priority of these goals may change throughout the operation, meaning that the value of assigned weights must change. To solve this problem, existing methods needed to be retrained every time that the weights changed, which is time-consuming and costly. Thus, an algorithm that can adapt to weight changes in operation time is needed. In this research, a dynamic multi-objective deep reinforcement learning algorithm is introduced to control the power of the HVAC system. Since it is needed to control the continuous power of the HVAC system, a Deep Deterministic Policy Gradient (DDPG) algorithm empowered by Tunable Training, is proposed which can adapt to changes of objectives weights. The controller can adapt to changes in the importance of objectives during operation and does not need to be retrained. The proposed method is compared with a dynamic multi-objective Deep Q-Network (DQN) Algorithm, which is used for discrete control, in a smart home. The results show that the cost of energy in the proposed method is up to 14 % lower than the DML-DQN algorithm.