A Method for Detecting and Localizing Open-Circuit Switch Faults in MMCs Using Separable Conv2D Neural Networks

Open-circuit switch faults in submodules (SMs) of modular multilevel converters (MMCs) pose a significant challenge to system reliability and continuous operation. This article proposes a fault detection and localization (FDL) method that utilizes a neural network based on SeparableConv2D with a dataset generated from three-phase circulating currents and SM capacitor voltages. Considering the crucial role of controllers in MMCs and their influence on signal behavior during normal and fault conditions, a comprehensive control framework is employed in two operational modes: grid-connected and RL load-connected. This also ensures that the FDL method remains independent of the modulation scheme. This article considers faults even when the faulty switch is outside the arm’s current path. Preprocessing techniques, such as K-nearest neighbors (KNNs) imputation, quantile transformation, and sliding time window (STW), are used to prepare the data for optimal neural network input. Model optimization uses the AdamW algorithm, with learning rates dynamically adjusted via the ReduceLROnPlateau method. This approach minimizes computational complexity by requiring only one training session while leveraging optimized network architecture and preprocessing techniques to enhance FDL speed and accuracy. Finally, simulation and experimental results confirm the effectiveness of the proposed method.