In this paper, a convolutional neural networks (CNN) based analysis method to detect amplitudes and phases of harmonic components in the power system generated as a result of the existence of non-linear loads. The aim of the proposed method is to generate real time and accurate harmonic estimations for active filters applications. The harmonic components are obtained by parallel processing technique on a graphics processing unit (GPU) framework. In order to the CNN 8000 power system data with harmonics generated randomly in the simulation environment and test is achieved on a separated set of 2000 data. It has been showed that the proposed method estimates power system harmonics and generates satisfactorily accurate results for active filter applications.