Comparative Research between Several Lightweight CNN Models for Non-intrusive Load Monitoring Using Transfer Learning Approach

Abstract

V-I trajectory is widely used as load signatures in non-intrusive load monitoring (NILM) problems. This paper proposes a simple but efficient method to obtain RGB V-I trajectory, using the time derivatives of instantaneous reactive current and voltage to map gray V-I trajectory to RGB space. With two kinds of V-I trajectory as load signatures, i.e., the greyscale and the RGB respectively, comparative research between 7 Lightweight CNN models is carried out for NILM problems using transfer learning approach. The results show that an averaged improvement of 3.1% of F1-score is obtained using the RGB V-I trajectory as load signatures than the greyscale one. Larger input resolution of V-I trajectory benefits the CNN models’ performance, and those models trained on the RGB V-I trajectory of various resolutions have a smaller variance, indicating the RGB V-I trajectory could alleviate the influence of resolutions. According to comparative research results, EfficientNet-BO achieves the highest F1-score, no matter using the greyscale or the RGB V-I trajectory. The performance of MobileNet V2-1.0 is better than MobileNet V3-small-1.0 while slightly worse than MobileNet V3-large-1.0 when input resolution larger than 96*96. In the end, inference speed test is carried out, and a new comprehensive metric is proposed to evaluate the adaptability of different CNN models for NILM problem, which trades off accuracy and reference speed. According to this metric, ShuffleNetV2-1.0 and MobileNetV2-1.0 are much more efficient for NILM problem than other CNN models.