Disaggregated power system signal recognition using capsule network

Non-intrusive-load-monitoring (NILM) is normally based on power series analysis. In the load classification stage we use an image based deep convolutional neural network (DCNN) which is modelled on the biological visual cortex thereby achieving extremely high levels of object recognition and classification. However, the downsize to the DCNN is the requirement of a large image training dataset, translational invariance and loss during max pooling of information captured in small signal perturbations. In this paper to reduce the training dataset, provide appliance signature equivariance recognition and replace max pooling with routing by agreement for improved NILM recognition and classification we use Hinton’s capsule network (CapsNet). Disaggregated appliance current, real power and power factor signals are converted to two-dimensional (2D) images and then complementary fused together for increased recognition accuracy before final input into the CapsNet. We implement the Discrete Wavelet Transform (DWT) since it is able to transform within a large frequency band portfolio and fuses well low pixel images. By using image fusion technique we show that for only fifteen images per appliance and with no data augmentation we are able to achieve average prediction accuracies of up to 93.75% and hence consolidate the validity of the CapsNet in the NILM recognition and classification scheme for limited data memory and improved recognition.