A novel series arc fault detection method for photovoltaic system based on multi-input neural network. (September 2022)

• Record Type: Journal Article
• Authors: Chen, Xiaoqi; Gao, Wei; Hong, Cui; Tu, Yanzhao
• Is Part Of: International journal of electrical power & energy systems. Volume 140(2022)
• Publication Date: 2022-09
• Subjects: Photovoltaic (PV) system -- Series arc fault (SAF) -- Hankel-singular value decomposition (Hankel-SVD) -- Multi-input convolutional neural network (MICNN) -- Squeeze-and-excitation-inception (SE-Inception); Electrical engineering -- Periodicals; Electric power systems -- Periodicals; Électrotechnique -- Périodiques; Réseaux électriques (Énergie) -- Périodiques; Electric power systems; Electrical engineering; Periodicals; 621.3
• DOI: 10.1016/j.ijepes.2022.108018
• Languages: English
• ISSNs: 0142-0615

Abstract

Highlights: Based on the traditional light-weight model of CNN, a multi-input convolutional neural network (MICNN) is established, which can input the features of time domain and frequency domain at the same time and train them respectively. The inception structure is introduced into the CNN model to realize the fusion of different scale features, which can avoid the over fitting and complex computation caused by the increase of the depth and width of the network. The channel attention mechanism is introduced into the CNN, and the squeeze-and-excitation (SE) module is combined with the perception structure to realize the feature recalibration, which further improves the accuracy of the model. Abstract: There is a risk of fire caused by series arc failure in the operation of photovoltaic (PV) system. Therefore, it is required to discuss a solution for rapid arc fault detection. To address the series arc fault (SAF) detection under different working conditions, a method based on squeeze-and-excitation (SE)-inception multi-input convolutional neural network (MICNN) is proposed. Firstly, normalization and Hankel-singular value decomposition algorithm are used to denoise the current, which effectively avoid the influence of switching frequency on the subsequent diagnostic accuracy. Subsequently, the filtered time-domain signal and the frequency-domain signal after Fourier transform are input into a variant one-dimensional convolutional neural network (1D-CNN) model for training and testing. The proposed model is characterized by transforming the traditional CNN into MICNN, and introducing the inception network with spatial scaling function and the SE network structure with channel attention mechanism. Extensive simulations are performed to evaluate the efficacy with a desirable result of 97.48%, which is superior to traditional methods such as CNN, wavelet decomposition, and mathematical statistics. The proposed method can not only detect arc faults occurring in different locations, but also resist the disturbance of dynamic shading, maximum power point tracking (MPPT), strong wind, etc. In addition, this model achieved satisfactory results in three cases of long line fault, single series and array ageing.