Chillers consume considerable energy in building HVAC systems, and the optimal operation of chillers is essential for energy conservation in buildings. This article proposes a model-free optimal chiller loading (OCL) method for optimizing chiller operation. Unlike model-based OCL methods, the proposed method does not require accurate chiller performance models as a priori knowledge. The proposed method is based on the Q-learning method, a classical reinforcement learning method. With the comprehensive coefficient of performance (COP) of chillers as the environmental feedback, the model-free loading controller can learn autonomously and optimize the chiller loading by adjusting the set points of the chilled water outlet temperature. A central chiller plant in an office building located in Shanghai is selected as a case system to investigate the energy conservation performance of the proposed method through simulations. The simulation results suggest that the proposed method can save 4.36% of chiller energy during the first cooling season compared to the baseline control, which is slightly inferior to the value for the model-based loading method (4.95%). Owing to its acceptable energy-saving capability, the proposed method can be applied to central chiller plants that lack a system model and historical data.

冷水机组在建筑HVAC系统中消耗大量能源，而冷水机组的最佳运行对于建筑物的节能至关重要。本文提出了一种用于优化冷却器运行的无模型最佳冷却器装载（OCL）方法。与基于模型的OCL方法不同，该方法不需要先验知识即可获得准确的冷却器性能模型。所提出的方法基于经典的强化学习方法Q学习方法。通过将冷水机组的综合性能系数（COP）作为环境反馈，无模型负载控制器可以通过调节冷水出口温度的设定值来自主学习并优化冷水机组的负载。以上海某办公楼的中央冷水机组为例，通过仿真研究了该方法的节能性能。仿真结果表明，与基线控制相比，该方法在第一个降温季节可节省4.36％的冷水机能量，略低于基于模型的加载方法的数值（4.95％）。由于其可接受的节能能力，因此该方法可用于缺乏系统模型和历史数据的中央冷水机组。略逊于基于模型的加载方法（4.95％）的值。由于其可接受的节能能力，因此该方法可用于缺乏系统模型和历史数据的中央冷水机组。略逊于基于模型的加载方法（4.95％）的值。由于其可接受的节能能力，因此该方法可用于缺乏系统模型和历史数据的中央冷水机组。