Data-Driven Fault Detection of Multiple Open-Circuit Faults for MMC Systems Based on Long Short-Term Memory Networks

Abstract

This paper presents a long short-term memory (LSTM)-based fault detection method to detect the multiple open-circuit switch faults of modular multilevel converter (MMC) systems with full-bridge sub-modules (FB-SMs). Eighteen sensor signals of grid voltages, grid currents and capacitance voltages of MMC for single and multi-switch faults are collected as sampling data. The output signal characteristics of four types of single switch faults of FB-SM, as well as double switch faults in the same and different phases of MMC, are analyzed under the conditions of load variations and control command changes. A multi-layer LSTM network is devised to deeply extract the fault characteristics of MMC under different faults and operation conditions, and a Softmax layer detects the fault types. Simulation results have confirmed that the proposed LSTM-based method has better detection performance compared with three other methods: K-nearest neighbor (KNN), naive bayes (NB) and recurrent neural network (RNN). In addition, it is highly robust to model uncertainties and Gaussian noise. The validity of the proposed method is further demonstrated by experiment studies conducted on a hardware-in-the-loop (HIL) testing platform.