Transfer Learning for Multiappliance-Task Nonintrusive Load Monitoring

An accurate identification of the electricity consumption status of users is crucial in the development of smart grids. The nonintrusive load monitoring (NILM) technology plays a pivotal role in effectively recognizing the users’ energy consumption behavior. Among various NILM methods, deep learning has shown outstanding performance. However, when deep learning is applied to different data domains, it will face challenges, such as limited labeled data and extensive training times. To address these issues, transfer learning has been employed in NILM. However, existing methods have shown limited accuracy and efficiency in disaggregating multiple appliances. This article proposes a novel transfer learning approach for NILM, which incorporates dual objectives: energy disaggregation and appliance state detection. By employing a regression-classification framework within a subtask-gated network (SGN), the approach enhances the model’s generalization capabilities and significantly improves posttransfer performance. In addition, the model adapts from single appliance to multiappliance settings under the transfer learning framework. Furthermore, attention mechanisms are utilized to refine the extraction of generalized features, enabling the multiappliance models to outperform their single-appliance counterparts. Experimental results show that the proposed method improves mean absolute error (MAE) by 60% and increases the $F1$ score by 200% compared with other transfer learning methods, highlighting its effectiveness in multiappliance-task NILM.