| 引用本文: | 李钊涛,罗庆全,余 涛,等.基于模型-数据混合驱动的空调监测及其域适应方法[J].电力系统保护与控制,2026,54(05):34-48. |
| LI Zhaotao,LUO Qingquan,YU Tao,et al.Air conditioner monitoring and its domain adaptation method based on a model-data hybrid-driven approach[J].Power System Protection and Control,2026,54(05):34-48 |
|
| 摘要: |
| 高耗能的空调负荷因其热储特性成为源荷互动的重要对象。作为可调潜力评估的基础，非侵入式负荷监测仅需总表数据即可获取空调运行信息。然而，现有监测模型不仅难以捕捉空调的复杂运行特性，在从源域训练数据迁移到新用户的目标域时，还高度依赖难以获取的运行功率标签。因此，提出基于模型-数据混合驱动的空调监测及其域适应方法。首先，构建融合电气与环境特征长短期关联性的空调监测模型。其次，通过神经网络辨识空调热力学模型参数，推算室内温度变化，进而在训练中交替嵌入温度估计损失以充分利用空调与温度的强相关性。最后，再以物理模型作为桥梁，在新用户中采用易获取的温度数据作为监督信号适应新的数据分布。公开及自建数据集实验结果表明，所提方法在监测精度、可迁移性和域适应性方面均优于现有方法，展现出良好的应用前景。 |
| 关键词: 空调监测 多元特征 物理模型 混合驱动 域适应 |
| DOI：10.19783/j.cnki.pspc.250606 |
| 分类号: |
| 基金项目:国家自然科学基金企业创新发展联合基金集成项目资助(U24B6010)；广东省基础与应用基础研究基金项目资助(2025A1515010118) |
|
| Air conditioner monitoring and its domain adaptation method based on a model-data hybrid-driven approach |
|
LI Zhaotao, LUO Qingquan, YU Tao, LIANG Minhang, WANG Keying, PAN Zhenning
|
|
School of Electric Power, South China University of Technology, Guangzhou 510640, China
|
| Abstract: |
| Due to their high energy consumption and inherent thermal storage capacity, air conditioning loads have become key resources in source-load interaction. As a foundation for assessing flexible potential, non-intrusive load monitoring (NILM) enables the extraction of air conditioner operating information using only aggregate meter data. However, existing monitoring models not only struggle to capture the complex operating behaviors of air conditioners, but also rely heavily on hard-to-obtain power consumption labels when transferred from a source domain with training data to a new user in the target domains. To address these issues, a model-data hybrid-driven air conditioner monitoring and domain adaptation method is proposed. First, an air conditioner monitoring model integrating long- and short-term correlations between electrical and environmental features is developed. Second, a neural network is employed to identify thermodynamic parameters of air conditioners, enabling indoor temperature estimation. A temperature estimation loss is alternately embedded during training to fully exploit the strong correlation between air conditioner operation and indoor temperature. Finally, the physical model serves as a bridge for domain adaptation in new users, where easily accessible temperature data are used as supervisory signals to adapt to new data distribution. Experiments on both public and self-collected datasets demonstrate that the proposed method outperforms existing approaches in monitoring accuracy, transferability, and domain adaptability, showing promising potential for real-world applications. |
| Key words: air conditioner monitoring multivariate features physical model hybrid-driven domain adaptation |
