The implementation of advanced methodologies for optimizing energy efficiency requires comprehensive analysis of energy consumption patterns. Non-Intrusive Load Monitoring (NILM), which enables device- or process-level energy disaggregation, has emerged as a pivotal strategy for this purpose. With the rapid development of machine learning (ML) in time-series data analytics, cutting-edge ML algorithms are increasingly being applied to energy consumption disaggregation, reducing reliance on extensive smart meter deployment and thereby lowering monitoring costs. However, in practical usage, ML models trained on public datasets often exhibit substantial performance degradation when transferred to specific industrial processes, due to the inherent diversity of industrial systems. Although fine-tuning methods can help mitigate this gap, they typically require large volumes of on-site energy data, which imposes an intensive commitment to industry users. In practice, only a limited number of samples -often few-shot data- are available. In order to address the issue, this paper proposes a few-shot learning-based, Sequence-to-Point Convolution Neural Network (FSL-S2P-CNN) that minimizes the need for extensive on-site fine-tuning. Using aggregated meter readings as input, our proposed FSL-S2P-CNN generates time-varying energy consumption split ratios at device or process levels. The few-shot-learning feature is built on top of well-cited S2P-CNN architecture, addressing industrial process diversity. Particularly, FSL-S2P-CNN incorporates a Siamese-network-based dataset reconstructor, which enriches fine-tuning datasets by identifying analogous patterns from large base datasets using similarity assessments within a learned embedding space. Experiments were conducted using few-shot energy consumption time-series data collected at a workshop over two days at one-minute resolution. Results demonstrate that FSL-S2P-CNN achieves approximately 2.5-times improvement in training-to-validation loss ratio compared to baseline models with conventional fine-tuning. These findings highlight the potential of the proposed approach to significantly enhance the generalizability and scalability of ML techniques for industrial NILM applications.
Du et al. (Thu,) studied this question.