An SAGPSO-CNN Approach for Supervised Fault Classification in MVdc Shipboard Power Systems

All-electric ships (AESs) utilizing medium-voltage dc (MVdc) shipboard power systems (SPSs) rely on a limited number of generators to supply power to propulsion units and onboard loads. To meet the demands of dynamic maritime missions, frequent adjustments in network topology and operating power are required, which pose significant challenges for fault management by complicating fault detection and reconfiguration. This study proposes a convolutional neural network (CNN) optimized via surrogate-assisted gradient particle swarm optimization (SAGPSO) to enhance the efficiency and accuracy of short-circuit fault classification in MVdc systems. A two-zone MVdc simulation model, developed in power system computer aided design (PSCAD)/electro-magnetic transient design and control (EMTDC), is used to generate a comprehensive dataset covering high-impedance faults and various transient disturbances. These signals are converted into spectrograms using the short-time Fourier transform (STFT), enabling rich time-frequency feature representation. A supervised CNN is then employed to learn discriminative features from the spectrograms for fault classification, while SAGPSO fine-tunes the hyperparameters by incorporating gradient information and surrogate modeling to improve model performance and convergence speed. Experimental results demonstrate that the SAGPSO-CNN approach effectively distinguishes high-impedance dc interpole faults from strong transient disturbances, such as pulse load variations and sudden increases in propulsor loads, with a classification accuracy exceeding 99%. Furthermore, the method significantly refines network parameter configurations, supporting robust fault management in MVdc SPS.