Open-Circuit Fault Detection and Localization in Modular Multilevel Converters Based on Transformer Neural Networks

Abstract:
A Modular Multilevel Converter (MMC) with Half- Bridge Submodules (HBSMs) consists of multiple HBSM units connected in series, each containing two switches. The high number of switches necessitates rapid fault detection and localization (FDL) to maintain system reliability and efficiency. This paper presents a method for the detection and localization of opencircuit faults in MMCs. The MMC under study incorporates a phase voltage controller, an output current controller, a sorting algorithm for capacitor voltage balancing, and a circulating current controller. The proposed approach utilizes circulating currents from the three phases and the maximum capacitor voltage index from each arm, obtained via the sorting algorithm, to detect and localize open-circuit faults, significantly reducing the dataset size. Data is scaled using the RobustScaler method and fed into a Transformer network with a Positional Encoding Layer. The model is optimized using the Rectified Adam optimizer and categorical cross-entropy loss function, with the ReduceLROn- Plateau method for adaptive learning rate adjustment, enhancing convergence and stability. Once trained, the model utilizes the obtained weights and biases to improve FDL speed and accuracy without requiring complex and repetitive computations. Simulation results confirm the efficiency and performance of this method.