Impact of Normalization on BiLSTM Based Models for Energy Disaggregation

Non-Intrusive Load Monitoring is the decomposition of each appliance energy from the total energy recorded at a central smart meter. The success of machine learning based disaggregation models immensely depends on the quality of the input data. This paper evaluates different normalization techniques and their impact on the performance of a Bi-directional Long Short-Term Memory (BiLSTM) model for energy disaggregation. The group of appliances consists of a non-linear device (television), three multi-sate devices (washing machine, rice cooker and microwave) and two continuously operating devices (refrigerator and Kimchi refrigerator). The comparison was conducted on two publicly available datasets and the results showed that applying normalization to the input samples rather than feature values can improve the performance significantly. The best results have been obtained for normalization with L
₂
-norm.