Design a neural network for features selection in non-intrusive monitoring of industrial electrical loads

This paper proposes to compare the performance of neural network classifiers between back propagation (BP) and learning vector quantization (LVQ) for pattern analyses of features selection in a non-intrusive load monitoring (NILM) system. Load recognition for identifying loads being connected and disconnected is applied to a NILM by using a neural network, especially for industrial electrical loads, even though some loads are activated at the nearly same time. In order to accurately decompose the aggregate load into its components, a feature-based model for describing the signatures of individual appliances and load combinations is used. The model will suggest the certain signatures which can be detected for all loads in order to indicate the activities of the separate components. To verify the performance of the model for the features selection, the data sets of the electrical loads and the load recognition techniques apply an electromagnetic transient program (EMTP) and a neural network, respectively. The effectiveness and computation equipment of load recognition are analyzed and compared by using the back propagation classifier and the learning vector quantization classifier. To obtain a maximum recognition accuracy rate, the calculation of the turn-on transient energy signature employs a window of samples, Δt, to adaptively segment a transient representative of a class of loads. Experiments performed with a variety of model data sets which reveal the back propagation classifier is superior to the learning quantization classifier in the effectiveness and computation equipment of load recognition.