Screening Indicator Selection and Clustering Method for Dynamic Current Balancing of Paralleled SiC MOSFETs

Paralleling multiple SiC MOSFETs is beneficial for enhancing current-carrying capability. However, parameter dispersion in the paralleled SiC MOSFETs will result in significant dynamic current imbalance. A common solution is to screen the threshold voltage and transconductance of SiC MOSFETs before paralleling them. However, it is found that these two static parameters are insufficient to characterize the dynamic voltage behaviors of SiC MOSFETs. Paralleling SiC MOSFETs with different dynamic voltage behaviors will lead to dynamic current redistribution, thereby exacerbating the current imbalance. To ensure the dynamic consistency of paralleled SiC MOSFETs, this paper proposes expanding the screening indicators to include gate resistance and Miller capacitance. As support for this proposal, this paper separately measures the static and dynamic parameters of fifty SiC MOSFETs. Deep neural networks and SHapley Additive exPlanations (SHAP) algorithms are employed to quantitatively determine the marginal contribution of the static parameters to the dynamic parameters. The analysis result suggests selecting the five static parameters as screening indicators including threshold voltage, transconductance, input capacitance, Miller capacitance, and internal gate resistance. To cope with the expansion of screening indicators and the increase in data dimensions, this paper proposes a clustering method that integrates autoencoders with K-means clustering to classify SiC MOSFETs. The experimental results demonstrate balanced currents without redistribution when SiC MOSFETs within the same cluster are connected in parallel.