【摘要】 建筑暖通空调系统在动态热扰动下存在高维非线性耦合,导致节能控制工作在能效寻优与热舒适协同方面难以实现自适应滚动权衡。为此,提出一种基于实时测量与数据驱动的暖通空调系统自适应节能控制技术。通过部署多源异构传感网络实时测量建筑围护结构表面温度、室内外温湿度、太阳辐射照度及空气渗透率等热工参量,形成原始高维特征流形。然后,采用核主成分分析与本征正交分解相结合的降维重构算法,剔除原始测量数据中的冗余分量,获得低维紧致表征的热工状态空间。在此基础上,基于降维重构后的时序特征序列,构建具备动态时滞补偿能力的长短时记忆网络(Long Short-Term Memory,LSTM)模型,实现对建筑热负荷在预测时域内的滚动动态预测。将LSTM预测输出嵌入模型预测控制框架,以系统耗电量最小化与室内热舒适偏差范数为双目标函数,在满足执行机构物理约束的前提下,通过自适应滚动时域优化生成控制指令,实现制冷剂流量与风机转速的协同调节。实验结果表明,所提技术使暖通空调进、出水温差稳定在2_℃以内,控制指令平均变化率仅为0.081 Hz/s,较基于多层Re LU网络和强化学习的技术分别降低64.09%和71.67%,且24 h总耗电量最低。该方法在制冷效果、控制平滑性及节能性等方面均表现出色,有效解决了能效寻优与热舒适协同之间的固有矛盾,为暖通空调系统在复杂热环境下的鲁棒节能运行提供可量化的技术路径。更多还原

【Abstract】 The building heating,ventilation,and air conditioning(HVAC) system exhibits high-dimensional nonlinear coupling under dynamic thermal disturbances,making it difficult to achieve adaptive rolling trade-off in energy efficiency optimization and thermal comfort coordination during energy-saving control.Therefore,this study proposes an adaptive energy-saving control technology for HVAC systems based on real-time measurement and data-driven methods.By deploying a multi-source heterogeneous sensor network to measure thermal parameters such as the surface temperature of the building envelope,indoor and outdoor temperature and humidity,solar radiation intensity,and air permeability,a raw high-dimensional feature manifold is formed.Then,a dimensionality reduction and reconstruction algorithm combining kernel principal component analysis and intrinsic orthogonal decomposition is adopted to eliminate redundant components in the original measurement data,obtaining a lowdimensional compact representation of the thermal state space.On this basis,an long short-term memory(LSTM) model with dynamic time-delay compensation capability is constructed based on the time-series feature sequence after dimensionality reduction and reconstruction,enabling rolling dynamic prediction of the building s heat load within the prediction time domain.The LSTM prediction output is embedded into the model predictive control framework,minimizing the system power consumption and the norm of indoor thermal comfort deviation as the dual objective functions.Under the premise of satisfying the physical constraints of the actuator,control instructions are generated through adaptive rolling time-domain optimization to achieve coordinated regulation of refrigerant flow and fan speed.Experimental results show that the proposed technology keeps the inlet and outlet water temperature differences of the HVAC system within 2℃,with an average change rate of control instructions of only0.081 Hz/s.Compared with technologies based on multi-layer ReLU networks and reinforcement learning,it reduces by 64.09%and 71.67% respectively.The total power consumption over 24 h is the lowest.It also demonstrates excellent performance in terms of cooling effect,control smoothness,and energy efficiency,effectively resolving the inherent contradiction between energy efficiency optimization and thermal comfort coordination,and providing a quantifiable technical path for the robust energy-saving operation of HVAC systems in complex thermal environments.更多还原