Worldwide, Nuclear Power Plants (NPPs) must have higher security protection and precise fault detection systems, especially underground power cable faults, to avoid causing national disasters and keep on safe national ratios of electricity production. Hence, this paper proposes an automatic, effective, and accurate Deep Learning (DL)-based fault classification and location technique for these cables via a One-dimensional Convolutional Neural Network (1D-CNN) and a Binary Support Vector Machine (BSVM). The proposed approach includes four main steps: data collection, feature extraction and reduction, fault detection, and fault classification and location. Signal collection from the underground cable’s sending end is performed via the Alternating Transient Program/Electromagnetic Transient Program (ATP/EMTP). Feature extraction and reduction are performed via Fractional Discrete Cosine Transform (FrDCT) and Singular Value Decomposition (SVD) methods. Fault detection is performed through leveraging BSVM with the linear Kernel method in the third step. Finally, this permits 1D-CNN to classify the fault type and locate it. Simulation results confirmed the efficiency of our proposed method, especially for 11kV underground cable faults, including different fault resistances and inception angles. Moreover, the proposed technique is applicable in real-time scenarios with a 99.6% accuracy rate, 0.15sec lowest execution time, and 0.095% maximum error rate for fault location at fractional factor ( $\alpha $ ) equals to 0.8.

Indirect Lightning Stroke (ILS) is considered an urgent issue on overall power systems due to its sudden dangerous occurrence. A grid-connected solar Photovoltaic (PV) power plant of 1MW was considered and analyzed using PSCAD/EMTDC software. The effect of grounding grid resistance ( $R_{g}$ ) on the induced voltages caused by the indirect strokes was discussed. The Transient Grounding Potential Rise (TGPR) variation with $R_{g}$ was presented and discussed. Four different models were proposed and installed for the system under study, which includes a combination of the Externally Gapped Line Arrester (EGLA) with the Non-Gapped Line Arrester (NGLA). The results show that when the $R_{g}$ was reduced from 5 to 1 ohm, TGPR decreased by about 79.63%, whereas the peak value was reduced by about 91.3% nearby the striking position. Four models of EGLAs were proposed to reduce the induced transient overvoltage effectively. The four models showed a remarkable ability to reduce the backflow current (BFC) and, consequently, the induced overvoltage. The EGLA’s type with the composite air gap reduced the TGPR by about 77.04 % and reduced the induced overvoltage nearby the striking position by about 51.3%.

Fault diagnosis of Medium Voltage power Cables (MVCs) research nuclear reactor, incredibly inaccessible/remote ones, has to be carefully identified, located, and fixed within a short time. Therefore, this paper proposes a perfect simultaneous fault diagnosis scheme based on Multiclass Support-Vector Machine (MCSVM) in the fractional Fourier domain. First, the three-phase sending currents are simulated under different conditions then their features are extracted using Discrete Fractional Fourier Transform (DFRFT). Afterward, the features reduction process occurs via the Singular Value Decomposition (SVD) approach. MCSVM scheme is used to diagnose faults (i.e., discover, categorize, and trace) using reduced features outcome from DFRFT and SVD stages. Alternating Transient Program/Electromagnetic Transient Program (ATP/EMTP) simulations have been carried out for 22 kV unreachable MVC. Different kernels of SVM, i.e., linear, quadratic, or polynomial, and diverse factors of DFRFT, i.e., α, are investigated in simulations to obtain the optimum performance (i.e., best α and kernel pair). Hence, performance analysis of the proposed diagnosis method under different conditions (i.e., various fault resistances, locations, and inception angles) concluded two highest accuracy and lowest time settings, which were found at α = 0.5 (for both) quadratic kernel, and linear kernel, respectively. Moreover, the linear kernel achieves 99.8% accuracy rate, the lowest execution time (10 ms), and fault tracing error rate of 0.525789%, which is proper for real-time applications. Besides, our proposed method is more reliable and accurate against variable operating conditions (fault resistances, distances, and inception angles), leading to more reliable power production systems.