Nonintrusive load identification using extreme learning machine and TT-transform

This paper presents a new nonintrusive load identification method to disaggregate the target load in a typical commercial building. For this purpose, experiments were carried out in the laboratory implementing real load switching activity and power measurements were made with a smart meter. Nonintrusive load monitoring is performed by analysing the power signals obtained from the smart meter and detecting the operation of load appliances. A new feature extraction technique based on the time-time (TT)-transform is applied to improve the load identification. For classifying and predicting the various load operations, a new intelligent technique called as extreme learning machine with single hidden layer feedforward neural network is developed. ELM has high efficiency and simple to be implemented. The inputs to the ELM are the extracted TT-transform features together with other signals like real and reactive powers while the ELM outputs are the switching states of the load appliances. The load appliances considered in the study are the lightings, personal computers and air conditioners. The ELM accuracy is validated by testing with unknown dataset recorded by the smart meter at 1 minute sampling rate. The ELM testing results showed that the desired level of load identification can be achieved by using additional TT-transform features.