Optimal Design Methodology for A High-Frequency
Transformer Using Finite Element Analysis and
Machine Learning
by
Eunchong Noh
A thesis submitted in partial fulfillment of
the requirements for the degree of
Master of Science
(Electrical and Computer Engineering)
December 2021
Thesis committee:
Gyu-Sik Kim, Professor, ECE, University of Seoul
Seung-Hwan Lee, Associate Professor, ECE. University of Seoul
Moon-Que Lee, Professor, ECE, University of Seoul
Abstract
The demand for isolated DC-DC converters is increasing due to the spread of
electric vehicles (EV) and the expansion of renewable energy use. Accordingly, the
need for a high-frequency transformer, a key component of an isolated DC-DC
converter, is also increasing. This trend is also taking place in the field of railway
locomotive systems. Solid state transformer (SST) technology to improve the
performance and efficiency of railway locomotive propulsion systems is being
actively researched, and high-frequency transformer is the core of SST. Highfrequency
transformer design for railway locomotive systems has more complex
design elements that must be considered for volume-loss optimization and
insulation and thermal design.
This thesis investigates an optimization design methodology using machine
learning and NSGA-II for optimized high-frequency transformer design. For
machine learning, Finite-element analysis (FEA) simulation was used to obtain
high-frequency transformer parameter data. Conventional high-frequency
transformer optimization design methods used analytical models for parameter
calculation. However, this analytical model has a significant error when the shape
of the high-frequency transformer becomes complicated. In particular, the leakage
inductance of the high-frequency transformer is difficult to calculate with an
analytical model. So, it is difficult and cumbersome to apply it in the design. This
thesis obtained magnetizing inductance, leakage inductance, and copper loss of
shell-type transformer models in various shapes using FEA simulation. Then, using
the data obtained from the FEA simulation, a machine learning regression model
was created to minimize the parameter calculation error in complex shapes. In
addition, the NSGA-II algorithm, which is widely used in multi-variable
optimization design, is used to find the optimal transformer shape to perform
optimization that can satisfy multiple design elements at the same time.
Each parameter inferred by the machine learning regression model showed a
high correlation and sufficiently low inference error rate, used for parameter
calculation in the NSGA-II algorithm. The inferred parameters are used to calculate
transformer loss for optimization design or check whether constraints are satisfied.
Through the optimization design using NSGA-II, a Pareto front curve for volume
and loss that satisfies all design conditions was obtained. The designer can select
and use the designs according to the situation.
The methodology can be designed for more complex shapes to achieve higherperformance
high-frequency transformer design. In addition, the complexity of the
design is reduced because numerous consideration factors can be easily considered
through constraint setting in the NSGA-II. Finally, unlike the conventional design
methodology, which has a significant influence on the skill and intuition of the
designer, once the environment is set up, the design proceeds only by inputting
target parameters and executing the code so that the design time can be reduced.
Therefore, it is possible to design a high-frequency transformer with constantly
high performance regardless of the designer's skill level.
View full text:
