一种基于云边协同的非侵入式负荷辨识框架

Non-intrusive load identification framework based on cloud-edge collaboration

中文关键词: 非侵入式负荷辨识  云边协同  马尔可夫转移场  轻量级深度学习模型  自适应合成采样

英文关键词: non-intrusive load identification  cloud-edge collaboration  Markov transition field  lightweight deep learning model  adaptive synthetic sampling

中文摘要:

      为解决海量非侵入式负荷监测（non-intrusive load monitoring，NILM）数据上传云端带来的分析时延性与云端资源大量消耗问题，提出了一种基于云边协同的非侵入式负荷辨识框架。首先，采用马尔可夫转移场（Markov transition field，MTF）编码方式对功率数据进行颜色编码，构建特征明晰的负荷标识；接着，在云服务层和边缘服务层分别部署相同结构的轻量级深度学习模型以完成训练与负荷辨识任务，在降低云边资源压力的同时，通过迁移学习方式实现负荷辨识的云边协同；最后，基于自适应合成采样（adaptive synthetic sampling，ADASYN）对REDD数据集进行扩充以解决数据集不平衡引起的模型学习偏见，并基于该数据集对所提框架辨识性能进行有效性验证，结果表明该框架不仅能够满足负荷辨识的高精度与实时性需求，同时能够显著降低云端和边缘端的存储与计算资源压力。

英文摘要:

      In order to solve the problem of analysis ductility and large consumption of cloud resources caused by massive non-intrusive load monitoring（NILM）data uploaded to the cloud, a non-intrusive load identification framework based on cloud-edge collaboration is proposed. Firstly, the Markov transition field（MTF）coding method is used to color code the power data, and the load identification with clear characteristics is constructed. Then, a lightweight deep learning model with the same structure is deployed in the cloud service layer and the edge service layer respectively to complete the training and load identification tasks. While reducing the pressure of cloud-edge resources, the cloud-edge coordination of load identification is realized through transfer learning. Finally, based on adaptive synthetic sampling（ADASYN）, the REDD dataset is extended to solve the model learning bias caused by dataset imbalance, and the identification performance of the framework proposed is validated based on the dataset. The results show that the framework can not only meet the requirements of high precision and real-time load identification, but also significantly reduce the pressure of cloud and edge storage and computing resources.