Study on Model Transferability for Non-Intrusive Load Monitoring of Households Appliances to Other Seasons and Locations

Abstract

Non-Intrusive Load Monitoring is a process that uses machine learning to disaggregate the mains power into appliance powers to reduce the number of sensors. In this study, we used 10 months of newly measured load data of Japanese households to evaluate the transferability of the disaggregation model to other seasons and locations. In the case of transfer to another season, we found that Seq2Point model can minimize the error by constantly adding new data to the training data, but the factorial hidden Markov model should be trained only for similar months. Migration of the model to another location resulted in large aggregation errors due to the different use and combination of appliances. In particular, appliances with multiple modes tend to be used in different ways in different locations, i.e., their power loads tend to be different, which reduces model transferability.