基于深度强化学习的热泵供热系统节能控制  

Energy-efficient Control of Heat Pump Heating System Based on Deep Reinforcement Learning

作　　者：秦浩森 于震 李太禄 李立 QIN Haosen;YU Zhen;LI Tailu;LI Li(Hebei University of Technology,Tianjin 300131,China;China Academy of Building Research,Beijing 100013,China)

机构地区：[1]河北工业大学,天津300131 [2]中国建筑科学研究院有限公司,北京100013

出　　处：《建筑科学》2022年第12期1-6,共6页Building Science

基　　金：国家重点研发计划项目“净零能耗建筑适宜技术研究与集成示范”(2019YFE0100300)。

摘　　要：热泵供热系统广泛应用于住宅建筑,其优化控制对于提高需求侧的能源效率至关重要。基于模型的控制方法需要精确的建筑模型,而无模型控制方法前期效果较差,收敛速度较慢。针对这些问题,提出了1种基于Deep Q-Learning及其改进算法的强化学习方法。该方法具有较快收敛速度,能够根据不同建筑环境自适应学习建模,在热舒适收益和能耗成本之间寻找平衡。北京市某近零能耗住宅建筑实际验证结果表明,该算法与基准策略相比综合收益提高15.3%。Heat pump heating systems are widely used in residential buildings,and their optimal control is essential to improve energy efficiency on the demand side.An accurate building model is required for model-based control,whereas it will be ineffective upfront and slow to converge via a model-free control.To this end,a reinforced learning method based on Deep Q-Learning and its improved algorithm was proposed.This method features a fast convergence rate and is able to find a balance between thermal comfort benefits and energy costs by self-adaptive learning modeling based on different building environments.An increase of 15.3%in comprehensive benefit of the algorithm compared with the benchmark strategy is achieved in a near-zero energy residential building in Beijing,according to a practical validation.

关 键 词：强化学习 Deep Q-Learning 控制策略 暖通空调 近零能耗建筑 

分 类 号：TU831[建筑科学—供热、供燃气、通风及空调工程]