【摘要】 针对公共建筑现有空调负荷分解方法误差较大的问题，在总负荷季节性作差分解空调负荷的传统方法的基础上，提出了基于动态基准线预测模型的分解方法。首先，提出了基于改进K-Medoids模型的基准日划分方法，通过疏离系数修正K-Medoids聚类结果，准确划分季节交替期的空调日和基准日，为动态基准线预测提供准确的标签数据。其次，构建了基于两阶段注意力机制-长短期记忆（dual-stage attention mechanismlong short-term memory,DAM-LSTM）的动态基准线预测模型，引入两阶段注意力机制（dual-stage attention mechanism,DAM）优化配置长短期记忆（long short-term memory,LSTM）网络，通过特征注意力机制提升对关键特征的侧重、时序注意力机制强化对时序信息的筛选，为空调负荷分解提供准确的基准负荷曲线。最后，以镇江某办公建筑和商业建筑为例进行实验，结果表明该方法显著提升了空调负荷分解精度和鲁棒性，尤其对小样本算例和尖峰负荷预测等难点优势明显。更多还原

【Abstract】 To address the issue of large errors in existing air-conditioning load decomposition methods for public buildings, this paper proposes a decomposition method based on a dynamic baseline prediction model, building upon the traditional seasonal difference approach using total load. Firstly, an improved K-Medoids-based baseline day classification method is introduced. By incorporating a separation coefficient to refine clustering results, the method accurately distinguishes air-conditioning days and baseline days during seasonal transition periods, providing reliable labeled data for dynamic baseline prediction. Secondly, a dynamic baseline prediction model based on DAM-LSTM is constructed, where a dual-stage attention mechanism(DAM) is integrated to optimize the long short-term memory(LSTM) network. The feature attention mechanism enhances the model’s focus on key features, while the temporal attention mechanism strengthens its ability to filter temporal information, thereby improving the accuracy of baseline load curve prediction for air-conditioning load decomposition. Finally, experiments conducted on a government office building and a commercial building in Zhenjiang demonstrate that the proposed method significantly improves both the accuracy and robustness of air-conditioning load decomposition, showing particular advantages in small-sample scenarios and peak load prediction tasks.更多还原