A hierarchical HVAC optimal control method for reducing energy consumption and improving indoor air quality incorporating soft Actor-Critic and hybrid search optimization.

• It offers a hierarchical scheme solution for the optimal control of multi-zone HVAC systems. • A deep reinforcement learning-based method is proposed to optimize airflow demands. • A genetic algorithm-based method is proposed to accurately track multi-zone airflows. • It achieves energy saving and good IAQ under dynamic environmental changes. • It exhibits good generalization under different occupancy schemes and system topologies. Heating, ventilation, and air conditioning (HVAC) systems are designed to maintain a healthy indoor environment, where indoor air quality (IAQ) and energy use issues are of top concerns. This paper proposes a hierarchical optimal control method for multi-zone HVAC systems to improve IAQ while reducing fan energy consumption. The proposed hierarchical optimal control method consists of two levels. At the upper level, a virtual multi-zone HVAC environment is established and a soft actor-critic-based agent is trained under reinforcement learning framework to optimize the fan energy consumption while maintaining satisfactory IAQ in each zone. At the lower level, a "proportional balance + proportional recovery" strategy is devised to accurately track the terminal airflow via a hybrid search optimization method incorporating genetic algorithm and fmincon function. Compared with existing ventilation control methods, the proposed hierarchical optimal control method offers the following advantages: a) It can achieve good IAQ in multiple zones and low fan energy consumption by optimizing the demand airflow in response to the changes in real-time environment. b) It can accurately control the airflow with further energy saving, by optimizing the duct static pressure implicitly. c) Simulations demonstrate that the proposed method can achieve a maximum energy saving of 38.0 % and 43.6 % compared with two traditional methods, respectively. d) The proposed method exhibits good generalization ability under different occupancy scenarios and ventilation system topologies. [ABSTRACT FROM AUTHOR]