An Online Harmonic Estimation Technique Based on Deep Learning in Distribution Networks

Abstract

The escalating use of modern power electronics converters in distribution networks gives rise to power quality challenges, particularly harmonics. Harmonics pose a significant threat to the network, necessitating their minimization to prevent detrimental impacts such as thermal overheating, trips, and faults. Traditional analytical techniques for determining and accurately estimating the harmonic emissions in the network prove ineffective; furthermore, the online estimation of harmonics, which can assist network operators in employing diverse methods to mitigate harmonics and enhance the reliability and efficiency of networks, is currently inadequate. This article, therefore, proposes a deep learning (DL) technique based on deep neural network (DNN) for harmonic estimation in distribution networks. This technique is characterized by its online, quick, and real-time nature. The proposed DNN method can estimate complex and nonlinear relationships observed within distribution networks. A novel temporal convolution neural network (TCNN) architecture associated with two regressor heads has been proposed for estimating magnitudes and phase angles of voltage and current harmonics. It employs an online training paradigm, ensuring rapid compliance to nonstationary conditions. According to the model training results, the proposed TCNN model training brings more flexibility and adaptability to the network in various operational conditions. The proposed technique is validated using a practical experiment involving an adjustable speed drive (ASD) associated with three-phase diode-rectifiers connected to a distribution network.