A Lightweight Neural Network for Energy Disaggregation Employing Depthwise Separable Convolution

Abstract

Non-Intrusive Load Monitoring (NILM) is a blind source separation problem which aims to estimate the electricity usage of individual appliances by decomposing a household’s aggregate electricity consumption. Recent state-of-the-art neural network models have delivered a good performance on load disaggregation, however these models have tremendous model size with huge numbers of parameters to tune and require large amounts of data for training, thus are computation and memory demanding, which makes it difficult to be practically implemented. To address this problem, we study the mechanism of depthwise separable convolution and propose a modified convolutional neural network model, which results in a lightweight NILM algorithm with the aim to be deployed on edge devices with limited computation power as mobile phones or smart meters. The lightweight NILM algorithm is evaluated with public available dataset: UKDALE. Assessment results prove that the proposed model delivers acceptable disaggregation performance with dramatically reduced model size and computation requirements.