Performance modeling of power converters is indispensable for the design, control, and maintenance of power electronics systems. Conventional knowledge-based modeling approaches suffer from low accuracy and heavy computation burden induced by heavy human involvements. Data-driven modeling methods have emerged as potential solutions, leveraging empirical data to automatically extract functional relationships between input variables and output objectives. However, existing data-driven methods face challenges in terms of generalization capability, computational cost, and model interpretability. To address these challenges, this paper proposes the use of the light gradient boosting machine (LightGBM), a high-performance ensemble learning algorithm for performance modeling. LightGBM is a gradient boosting decision tree that utilizes gradient-based sampling and feature bundling techniques for efficient training and pruning. The paper applies LightGBM to model the efficiency and current stress of a multi-level DAB converter under a hybrid modulation, which is compared with several prevalent benchmark algorithms. 2-kW hardware experiments comprehensively validate the superiority and feasibility of the LightGBM for performance modeling.