Load Recognition With Few-Shot Transfer Learning Based on Meta-Learning and Relational Network in Non-Intrusive Load Monitoring
Dong Ding, Junhuai Li, Huaijun Wang, Kan Wang
Abstract
Non-intrusive load monitoring (NILM) monitors the operating status and power consumption of residential appliances with only one main meter, providing a new measure for energy management. Deep learning (DL) shows outperformance in NILM. However, the lack of appliance data caused by rapid growth appliance types and high-cost data sampling reduces the DL-based load recognition accuracy. Transfer learning (TL) enables DL-based model generalization. However, the generalization would be limited when data in the target domain are insufficient. Therefore, a non-intrusive load recognizing (NILR) few-shot TL based on meta-learning and relational network is proposed to improve the load recognition generalization performance, named MRNILR-TL. First, the method constructs an episode task dataset by task sampling to provide diverse learnable tasks for few-shot load recognition training. Afterward, a multi-classification load recognition model based on meta-learning and relational network is constructed, and a meta-learning based relational network enhances the ability to learn the laws of similarity among appliance features from few-shot data. Finally, achieving the few-shot multi-classification load recognition generalization by directly transferring the source task knowledge and strategies to the target task. Experimental results in four transfer scenes demonstrate the proposed method achieves generalization and outperforms most existing NILM TL methods.