Self-Adaptive Deep Learning Framework for Non-Intrusive Load Monitoring: Addressing Aging Appliance Challenges with Transfer Learning and Pseudo Labeling

Abstract

Efficient energy management practices are recognized as crucial for optimizing energy utilization. Non-intrusive load monitoring (NILM) has been identified as a promising solution, particularly with the use of deep learning techniques. Conventional NILM models often face difficulties in adapting to changes in power consumption patterns, especially as appliances age. To address this challenge, a self-adaptive NILM model is proposed, which integrates deep learning techniques with transfer learning and pseudo labeling. Unlike traditional NILM models, this approach incorporates a unique self-adaptive feature that enables the model to automatically adapt to changing power patterns resulting from aging appliances. Synthetic data generation and advanced neural network architectures are used for training and validating the model, achieving exceptional accuracy rates in disaggregating power consumption. Electrical appliances used for this experiment are categorized into two groups: on-time fixed devices and on-time variable devices. Experimental results demonstrate the effectiveness of the Self-Adaptive NILM approach with on-time variable devices, such as three-phase refrigerators. The model was tested over a six-year period, focusing on a three-phase refrigerator, and an accuracy rate exceeding 97% in disaggregating power consumption was achieved. It was found that for on-time fixed devices, the conventional NILM model gives better predictions. This high level of accuracy and the findings underscore the potential of this approach for energy management systems. By addressing a significant gap in existing NILM literature, this research introduces the way for the development of more robust, resilient, and adaptive energy management solutions.