Thermal Estimation of Modular Multilevel Converter Submodule Using Deep Regression on GRU and LSTM Network

Abstract

This paper proposed a GRU/LSTM-based deep regression model for thermal estimation of modular multilevel converter submodule. The MMC is composed of many submodules with the power semiconductors such as IGBTs and MOSFETs. The switches are the main components determining the reliability of the MMCs, and the swing of junction temperature causes most switch failures in the power semiconductors. So, thermal estimation is essential to improve the reliability of the MMC systems. Thermal modeling is a regression problem of time-series data, considering various environmental conditions. The conventional models cannot reflect the complex environmental conditions due to their fixed mathematic formulas. Therefore, this paper proposes the deep regression model that can estimate the junction temperature by using the arm current of the MMC submodule. The proposed model improved the accuracy of thermal estimation by more than 7.2 times compared to the existing method. Moreover, it does not require pre-processing and takes about 4.5ms on average to process 100ms data.

Keywords

• Thermal estimation
• modular multilevel converter (MMC)
• deep regression
• gated recurrent unit (GRU)
• long short term memory (LSTM)
• recurrent neural network (RNN)