Non-intrusive load monitoring based on lightweight single sequence to multiple sequences model

Abstract

Non-intrusive load monitoring (NILM) allows electricity users to obtain the power consumption of various household appliances in a low-cost way, which is conducive to promoting carbon neutrality and improving demand-side management capabilities. Aiming at the contradiction between load disaggregation errors and model calculation cost faced by general NILM algorithm, a lightweight NILM model based on single sequence to multiple sequences is proposed. The model adopts a fully convolutional structure based on depthwise separable convolution, and uses the feature extraction capabilities of different convolution kernels to achieve multiple outputs, which greatly reduces the amount of model parameters and calculation time, channel attention mechanism is then introduced to assign weights to different channels of feature maps, which reduces the load disaggregation errors of models. In terms of data processing, fuzzy C-means clustering is used to classify electrical appliances into two types, including single-operating-state appliances and multiple-operating-states appliances, and two methods of power estimation and state estimation are adopted to achieve power disaggregation errors separately. The model is verified on the REFIT dataset, and experiments show that the model can greatly reduce the computational cost while maintaining a low disaggregation error.