Improving multi-state appliance classification by SE-DenseNet based on color encoding in non-intrusive load monitoring

Abstract

Non-intrusive load monitoring (NILM) is a technique that efficiently monitors appliances' operational status and energy consumption by utilizing voltage and current data, without intrusive measurements. In NILM, designing efficient classification models and building distinctive load features are crucial. However, due to its continuously variable load characteristics, multi-state load identification remains the most challenging problem in NILM. In this paper, we improve the encoding of the color V–I trajectory by incorporating instantaneous power, thereby enhancing the uniqueness of V–I trajectory features. Furthermore, we investigate a NILM method based on deep learning methods and propose a densely connected convolutional network with squeeze-and-excitation network (SE-DenseNet) architecture to solve the multi-state load identification problem. Initially, the architecture leverages DenseNet's dense connectivity property to generate a multitude of feature maps from the V–I trajectory. Then, SENet's channel attention mechanism is employed to enhance the utilization of effective features, which is more effective for multi-state load identification. Experimental results on the NILM public datasets PLAID and WHITED show that the recognition accuracy of the proposed method reaches 98.60% and 98.88%, respectively, which outperforms most existing methods.