Evaluation of low-complexity supervised and unsupervised NILM methods and pre-processing for detection of multistate white goods

According to recent studies by the BBC and the Scottish Fire and Rescue Service, malfunctioning appliances, especially white goods, were responsible for almost 12,000 fires in Great Britain in just over 3 years, and almost everyday in 2019. The top three “offenders” are washing machines, tumble dryers and dishwashers, hence we will focus on these, generally challenging to disaggregate, appliances in this paper. The first step towards remotely assessing safety in the house, e.g., due to appliances not being switched off or appliance malfunction, is by detecting appliance state and consumption from the NILM result generated from smart meter data. While supervised NILM methods are expected to perform best on the house they were trained on, this is not necessarily the case with transfer learning on unseen houses; unsupervised NILM may be a better option. However, unsupervised methods in general tend to be affected by the noise in the form of unknown appliances, varying power levels and signatures. We evaluate the robustness of three well-performing (based on prior studies) low-complexity NILM algorithms in order to determine appliance state and consumption: Decision Tree and KNN (supervised) and DBSCAN (unsupervised), as well as different algorithms for preprocessing to mitigate the effect of noisy data. These are tested on two datasets with different levels of noise, namely REFIT and REDD datasets, resampled to 1 min resolution.