Performance Evaluation of FastICA as a Blind Source Separation Approach for Current-Based Load Disaggregation

Abstract

Deep learning has established itself as a prevalent method in non-intrusive load monitoring (NILM) due to its capacity to recognize patterns. However, the load disaggregation process, which aims to identify electricity consumption patterns, continues to face several challenges. These challenges arise from non-stationary and overlapping aggregate signals that result from highly dynamic electricity usage. Furthermore, machine learning approaches typically require extensive training data, regular updates, and significant computational resources, particularly when employing deep learning techniques. This paper assesses the performance of fast independent component analysis (Fast-ICA) as a promising blind source separation (BSS) method for load disaggregation. FastICA was applied to separate aggregate current signals using both deflation and symmetric separation schemes. Its performance was evaluated in both the time and frequency domains, incorporating wavelet transforms and principal component analysis (PCA) as preprocessing methods. In addition, three non-linear activation functions, tanh, cubic, and logcosh, were explored alongside various separation schemes. Employing FastICA combined with a tailored activation function and appropriate preprocessing techniques, proves effective as a BSS approach for load disaggregation in NILM.