Machine Learning-Based Prediction of Power Conversion Efficiency in a DC–DC Boost Converter under Variable Operating Conditions

Abstract

Accurate prediction of power conversion efficiency is essential for selecting operating points, comparing component designs, and implementing condition-aware energy management in DC–DC boost converters. This study develops and evaluates a machine-learning framework for predicting steady-state efficiency under simultaneous variation of input voltage, duty ratio, load resistance, switching frequency, inductance, ambient temperature, and output power. A transparent averaged electrical-thermal loss model was used to generate 3,600 feasible continuous-conduction operating points. The data were divided into independent training, validation, and test subsets, and six regression approaches were compared: polynomial ridge regression, support vector regression, random forest, gradient boosting, histogram gradient boosting, and a multilayer perceptron. Support vector regression produced the most accurate held-out predictions, with an R-squared value of 0.979, a root-mean-square error of 0.093 percentage points, and a mean absolute error of 0.073 percentage points. Ninety-five percent of test errors were below 0.190 percentage points. Permutation analysis identified input voltage and output power as the dominant predictors, followed by switching frequency, inductance, and duty ratio. Error remained below 0.110 percentage points across low-, medium-, and high-power regimes. The findings demonstrate that a compact nonlinear surrogate can reproduce converter efficiency maps with high precision, while also showing that the model must be validated experimentally before use outside the simulated component and operating domain.