Transformative enhancement of predictive models via fourier transformer-based denoising for non-intrusive load monitoring

Energy disaggregation is a pivotal technique employed in Non-Intrusive Load Monitoring (NILM) to accurately estimate the power consumption of individual appliances in households. Recent advancements in trasformer models, particularly in domains like Natural Language Processing (NLP), have underscored their potential for enhancing classification performance. In this paper, we propose a novel algorithm for energy decomposition using an adapted transformer-based architecture. Our approach integrates the rapid Fourier transform with a multi-head attention mechanism to efficiently transform time-series feature data into frequency domain features, followed by noise reduction achieved by multiplying the data with a trainable matrix. The DFT-NILM model, also known as the denoising Fourier transformer, has been evaluated on publicly available electrical data. The evaluation shows that the DFT-NILM model outperforms existing state-of-the-art methods in terms of performance.