Towards smart control and energy efficiency for multi-zone ventilation systems via an imitation-interaction learning method in energy-aware buildings

In this paper, an imitation-interaction learning (Imit-IR) method is proposed for multi-zone ventilation systems, to enhance the ventilation control performance and optimizes system energy efficiency. This method synergizes online reinforcement learning (RL) with offline imitation learning to expedite the convergence speed of RL and bolster the validity of exploration in the training process. In the imitation learning phase, the RL policy network is pre-trained via behavioral cloning using suboptimal demonstration sampled from expert trajectories. In the interaction learning phase, the control performance of the RL agent is significantly boosted by migrating network parameters from the offline phase and through environmental interaction. The proposed Imit-IR method offers the following advantages: a) The convergence speed of RL online learning is effectively accelerated, leveraging on biased sampling expert demonstration in offline learning. b) The proposed method exhibits good dynamic characteristics under time-varying demands, with the maximum airflow error controlled within 6 %. c) Test results demonstrate that the proposed method can achieve a maximum energy savings of 18.9 % compared to the three benchmark methods. d) The proposed method has good generalization ability under various dynamic scenarios and ventilation system topologies.