Photovoltaic (PV) systems installed in sensory systems networks require advanced fault handling. Existing methods of diagnosis find it hard when faced with the non-linear nature of PV installations, especially in the circumstances of high-impedance and transients. Four deep-learning models, including Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), Transformer, and one using Autoencoder as a baseline, are considered in the given work with the purpose of identifying anomalies in the system, including line faults, transient, and high-impedance faults. The training is performed on a simulated dataset which was generated based on Discrete Wavelet Transform (DWT) of microgrid EA, EB, and EC features extracted in MATLAB-Simulink. The findings suggest that the Transformer and the CNN are the best at finding the highest accuracy (89) and at F1-score (0.88). The two models exhibit significant ability to learn temporal and spatial characteristics, hence the ability to carry out good classification under noisy conditions or weak signals, which is precisely the kind of a scenario that is typical of high-impedance faults. These results contribute to the use of Transformer based models in the real-time PV system diagnosis.