A bifurcation deep neural network for electricity meter error prediction under actual conditions

Abstract

The environmental conditions have a great impact on the measuring accuracy of electricity meters, once they are installed. This paper aims to find a way to accurately evaluate the measuring errors of electricity meters under actual conditions. Specifically, a novel bifurcation deep neural network (BDNN) model was designed and tested. The BDNN consists of a subnetwork and a fully-connected network. The subnetwork is a deep autoencoder-convolutional neural network (DAE-CNN) dedicated to processing harmonic features. The fully-connected network takes the subnetwork output and the environmental conditions as its inputs, and generates the output of the entire model by softmax. Then, the BDNN was trained on a dataset generated by real experiments with electricity meters. Three hyperparameters, namely, the activation function, the number of hidden layers and the autoencoder structure, were optimized through several experiments. Through the optimization, the rectified linear unit (ReLu) was adopted as the activation function, the number of hidden layers was set to 4, and the autoencoder structure was determined as 256-128-64-32. Each numerical figure refers to the number of nodes in the corresponding hidden layer. Finally, the BDNN was compared with the least squares support vector machine (LS-SVM), the fully-connected MLP (FCP) and the original CNN, and found to outshine the contrastive methods in prediction error and computing cost. The research results shed important new light on the field calibration and error prediction of electricity meters.