Machine Learning Junction Temperature Estimation Method Based on TSEP in MMCs

Abstract

Modular multilevel converters (MMCs) represent an attractive solution for various high- and medium-voltage grid applications. However, they consist of a series connection of several submodules (SMs), which are subject to non-uniform thermal stresses, leading to differing rates of aging among the various SMs and causing unexpected failures and malfunctions in the converter. For this reason, various active thermal control strategies are employed to equalize thermal stress by utilizing information on the junction temperature of the MMC’s power devices. There are several methods for estimating junction temperature; the most common are based on analytical models with lumped parameters or on thermo-sensitive electrical parameters (TSEPs), utilizing the measurement of the semiconductor’s collector-emitter voltage. However, these methods have several limitations, such as high sensitivity to parameter variations and an inability to track rapid thermal changes; furthermore, TSEP methods are strongly influenced by PWM noise at the switching frequency and by non-linearities present near the inflection point of the collector-emitter voltage versus junction temperature characteristic. To address these limitations, this paper proposes a machine learning (ML)-based estimation architecture that employs a multilayer perceptron (MLP) neural network (NN). The proposed data-driven model resolves non-linearities and eliminates measurement errors when combined with Bayesian regularization and selective conduction sampling. The ML-enhanced TSEP approach suppresses transient estimation spikes, significantly reducing both the mean absolute error (MAE) and the root mean square error (RMSE), as demonstrated by experimental validation using a Power-Hardware-in-the-Loop (PHiL) system.