Open-circuit Fault Detection and Location in MMCs with Multivariate Gaussian Distribution

Abstract

Open-circuit fault of sub-module (SM) attracts more attention with the increasing applications of modular multilevel converter (MMC). To diagnose the SM open-circuit fault in MMC, this paper proposes a fault detection and location (FDL) method based on machine learning (ML). According to the open-circuit fault characteristics, the SM capacitor voltages are selected as the key indicator for anomaly detection model. A method of time-domain feature extraction from voltages is then introduced to construct the dataset for the model. After trained with the samples, the model based on multivariate Gaussian distribution can realize FDL by making predictions for anomaly detection and tracing back the predicted faulty SM. The proposed method can locate the faulty SM accurately without extra sensors or mathematical model of circuit. The results on a simulation of 21-level three-phase MMC present that the anomaly detection model achieves high FDL accuracy as well as low false alarm rate.