ST-NILM: A Wavelet Scattering-Based Architecture for Feature Extraction and Multilabel Classification in NILM Signals

Nonintrusive load monitoring (NILM) is a relevant tool for improving energy consumption habits, contributing to energy conservation and distribution system planning. In recent years, high-frequency strategies using deep learning have been presented in the literature, achieving the state-of-the-art results for detection, feature extraction, and classification of aggregated electrical loads, particularly with the architecture defined as deep neural network model for detection, feature extraction, and multilabel classification (DeepDFML). DeepDFML used a deep convolutional network (DCN) whose trained weights were shared for different output fully connected networks. The performance of DeepDFML depended on the availability of data and data augmentation (DA) strategies. Given this scenario, we propose the ST-NILM, a new integrated architecture based on the scattering transform (ST). ST-NILM has a DCN with analytical wavelet-based nontrained weights, shared with fully connected output networks that perform event detection and multilabel classification of aggregate loads. We compared ST-NILM and DeepDFML for the LIT-SYN dataset. ST-NILM achieved equivalent detection results to DeepDFML for two and three aggregated loads and performed better for single loads. The hardware implementation shows that ST-NILM consumes less memory, less GPU load, and substantially less computational effort than DeepDFML. ST-NILM presents comparable or even superior results than other state-of-the-art deep-learning-based methods.