A semi-supervised sparse representation neural network for error estimation of electricity meters with insufficient tagged samples

Abstract

Once installed, the electricity meters will face a huge variation in operating conditions, which exerts a major influence on the measuring accuracy. Thus, it is imperative to develop an effective way to estimate the errors of electricity meters in varied operating conditions. Considering the limited sample size, this paper develops a semi-supervised sparse representation (SSR) algorithm for error estimation of electricity meters with insufficient tagged samples. Each sample was considered a combination of two sub-signals, namely, the prototype dictionary P and the variation dictionary V. The prototype errors of electricity meters were taken as the P, using the Gaussian mixture model (GMM). The linear noises were sparsely characterized by the V. On this basis, the mixed tagged and untagged samples were processed in a semi-supervised manner, to obtain the nonlinear variations between the two types of samples. To verify its effectiveness, the proposed SSR algorithm was compared with other error prediction methods through four experiments on actual datasets of different sizes. The results show that our algorithm greatly outperformed the other methods in the accurate estimation of the errors of electricity meters in operating conditions. The research results provide an innovative way to onsite calibration of electricity meters.