Appliance Identification in NILM Applications by means of a Convolutional Auto-Encoder

In energy efficiency applications, Non-Intrusive Load Monitoring techniques (NILM) are typically used to deduce which electrical loads are being used in a building at a given time. The identification of household appliances, in particular manually operated ones, is relevant information that can also be applied to infer the routines of tenants in Active and Assisted Living environments (AAL). These tools and applications are becoming increasingly interesting, especially in Western countries, where the ageing population is putting a strain on public social and health services. In this context, this work aims to classify the on/off events of the devices considered in the BLUED database. For this purpose, an architecture is presented, consisting of a Convolutional Auto-Encoder (CAE) followed by a classifier neural network. The CAE is used to implement a dimensionality reduction process after the encoder. Input data are formatted as images, created with extracted sections of the high-frequency electric current signal captured around the switching events. It is noteworthy that this dimensionality reduction also allows a decrease in the computational load of the classifier. Regarding the CAE functionality, the reconstruction error reaches a value of 1.579 · 10
⁻³
, whereas in the validation stage a weighted average classification F1-score of 87 % is obtained for the whole architecture.