Non-Intrusive Appliance Identification with Appliance-Specific Networks

Abstract

Non-Intrusive Load Monitoring (NILM) is a technique for load identification and energy disaggregation. The problem is usually formulated as a single-channel blind source separation. NILM algorithms aim to identify the operating characteristics of individual appliances from aggregate power measurement. Recent advances in deep learning gave rise to many methods that mostly focus on learning a direct mapping from aggregate measurement to individual appliance power, but these methods often suffer from overfitting and don't generalize well. In this paper, we propose a novel NILM method that leverages advances in both supervised and unsupervised learning techniques. The proposed method consists of three stages: a) a Bayesian non-parametric learning-based approach is used to extract appliance states; b) synthetic minority oversampling technique (SMOTE) is employed to mitigate the heavy imbalance in switching events present in the NILM problem; and c) lightweight long short-term memory (LSTM) networks are employed for status classification for each appliance. We argue that making the differences before and after the switching event as the input to the networks can reduce complexity of network training and makes the proposed method robust to multi-appliance scenarios. Experiments are conducted to demonstrate the effectiveness of the proposed method, achieving better performance when compared to recent methods. Furthermore, an ablation study is conducted to demonstrate the effectiveness of each module of our method.