Artificiële intelligentie in vermogenselektronica-ontwerp – Machine Learning voor het modelleren en optimaliseren van DC-DC converters

Power electronics design automation, incorporating artificial intelligence (AI) to optimize the design of power converters, has emerged as a novel research topic given the complexities and challenges of power converter design. This thesis explores various AI techniques that enhance DC-DC converter design assisting in semiconductor component evaluation and selection, magnetic design, converter modeling, and parameter optimization. To streamline the evaluation of power losses across the enormous number of semiconductor components, an AI-based tool is developed to extract nonlinear dynamic properties from datasheets. This tool automates the data collection process for efficiently analyzing the datasheets and collecting data for further comparison. However, due to the high dimensionality of variables and operating conditions, calculating power losses remains computationally intensive. To address this, an embedding-encoded artificial neural network (ANN) is employed to model power MOSFET losses. This technique significantly accelerates MOSFET preselection, advancing the automation of power electronics design. Magnetic design is critical to power converter design, as it significantly influences power losses and overall converter volume. This thesis introduces power loss modeling and the design process for inductors and transformers. Subsequently, the reinforcement learning (RL) approach is applied to optimize the design of DC inductors with air gaps as a case study. Given a triangular excitation current and maximum volume constraints, RL effectively improves the inductor design process, accelerating optimization and enhancing efficiency. Considering both the power switch losses and magnetic losses, parameter optimization is conducted. By conducting the simulation program with integrated circuit emphasis (SPICE) simulations under varying parameter combinations, an ANN model is trained to predict semiconductor losses. Combined with an optimized inductor design, RL is employed to refine parameter selection, maximizing the overall efficiency while maintaining volume and ripple constraints. Additionally, this simulation data-based AI methodology extends to circuit behavior modeling, which is vital for parameter design and control in complex converter topologies. For full-bridge LLC converters, a knowledge-aware neural network (KANN) is used to develop a precise voltage gain model, enhancing design accuracy. Similarly, for single-input multi-output converters, a physics-informed neural network (PINN) works as a substitute model for modulation, focusing on balanced load conditions. In summary, AI-driven methodologies significantly improve the modeling and optimization of power converters at the component, circuit, and system levels. These advancements enhance design accuracy, efficiency, and automation, positioning AI as a key enabler in modern power electronics design.