"Explainable Physically Constrained Graph Neural Control and Multi-objective Reinforcement Learning for User-Centric HVAC Managem"

"Heating, ventilation, and air conditioning (HVAC) systems account for an important portion of building energy consumption, usually representing 30\\% to 50\\% of the total building energy use; optimizing them is critical for achieving the \u201cdual-carbon\u201d targets and for improving user comfort. Existing control methods based on deep reinforcement learning have achieved some success in discrete scenarios, but they often display poor generalization when faced with different building thermodynamic characteristics and user preferences, and they lack multi-objective optimization and interpretability. This paper proposes a user\u2011centric explainable physically constrained graph neural network control framework. The framework constructs a graph neural network by utilizing the building topology, embeds thermodynamic conservation constraints into the controller, and, in combination with multi\u2011objective reinforcement learning methods, seeks a Pareto\u2011optimal balance among energy consumption, user comfort, and carbon emissions. Experiments were conducted in a multi\u2011zone building simulation environment, where data collected from real settings with synthetic noise were used to train the model, and comparisons were made with baseline methods such as physically constrained deep learning models, traditional PID control, and multi\u2011agent RL. Experimental results show that the proposed method can reduce energy consumption by about 18\\%, reduce comfort violation rates by 45\\%, and exhibit better generalization in scenarios beyond the testing conditions. By adopting explanatory algorithms, the influence of factors such as outdoor temperature and occupancy on the policy is revealed, thereby enhancing the transparency and reliability of the control process."

采暖、通风与空调（Heating, Ventilation and Air Conditioning, HVAC）系统是建筑能耗的重要组成部分，通常占建筑总能耗的30%至50%；对其进行优化对于实现“双碳”目标与提升用户舒适度至关重要。现有的基于深度强化学习的控制方法在离散场景中已取得一定成效，但在面对不同建筑热工特性与用户偏好时，泛化能力往往较差，且缺乏多目标优化能力与可解释性。本文提出一种以用户为中心的、具备物理约束的可解释图神经网络（Graph Neural Network, GNN）控制框架，该框架通过利用建筑拓扑结构构建图神经网络，将热工守恒约束嵌入控制器中，并结合多目标强化学习方法，在能耗、用户舒适度与碳排放之间寻求帕累托最优平衡。研究在多分区建筑仿真环境中开展实验：采用带有合成噪声的真实场景采集数据训练模型，并与物理约束深度学习模型、传统比例-积分-微分（Proportional-Integral-Derivative, PID）控制、多智能体强化学习等基线方法进行对比。实验结果表明，所提方法可将能耗降低约18%，将舒适度违规率降低45%，且在测试条件外的场景中展现出更优的泛化能力。通过采用可解释性算法，本研究揭示了室外温度、人员占用情况等因素对控制策略的影响，从而提升了控制过程的透明度与可靠性。