A fast wide-range air balancing control method based on deep reinforcement learning

Abstract

Air balancing is a crucial technology for reducing energy consumption in ventilation ducts and improving indoor environmental quality. Existing air balancing methods face challenges due to the complex and diverse internal duct topologies of ventilation duct systems, as well as the strong coupling of airflow between branches, making air balancing tuning difficult. To address the slow convergence speed and difficulty in precise control of air balancing in multi-branch ventilation systems, this paper proposes a fast wide-range air balancing control method based on deep reinforcement learning (DRL-AB). A system airflow control process with Markovian properties is designed, with state, action, and reward functions oriented towards reducing airflow errors. A dynamic single-branch target training mechanism is designed to reduce training costs while improving the agent’s training efficiency and its ability to generalize to different target airflow levels. The proposed method’s performance under various target airflow conditions is verified through experimental platforms. The results indicate that, in all test cases, the maximum percentage error is controlled within 3.33%, ensuring rapid and accurate convergence of a wide range of target airflows. This demonstrates strong universality, prevents waste of airflow energy, and enhances energy utilization efficiency.