Advanced design and control techniques for DAB power converters with Machine Learning

This thesis investigates the application of advanced design and control techniques for a Dual Active Bridge (DAB) power converter using machine learning. The goal is to enhance computational efficiency in design modeling and improve the controller performance of the converter. The DAB power converter has significant attributes, such as galvanic isolation, a wide input-output voltage range, bidirectional power flow capability, and high power density. These qualities make it suitable for various applications, including electric vehicle charging infrastructure, aerospace, railway, and more. Therefore, advancements in the converter’s design and control can positively impact multiple sectors. This research integrates theoretical analysis, simulation, and experimental results to identify enhancements in design, control algorithm performance, and their implementation. The findings of this study are compiled in conference and journal articles and are organized into five chapters that summarize the results.

To achieve higher power density with a smaller and lighter design, the switching frequency of power electronic converters can be increased, made feasible by the use of wide bandgap semiconductors such as SiC or GaN. However, this higher switching frequency can result in unexpected converter behavior, where stray components change high-frequency characteristics, potentially causing electromagnetic interference with nearby devices or increasing component stress that may damage the equipment. In this context, modeling at high frequencies can be particularly challenging due to the multitude of factors involved and the extensive calculations needed to develop a model capable of assessing various design and operational conditions. Furthermore, incorporating additional objectives such as efficiency, reliability, cost, or weight will complicate the overall design process. In this thesis, we analyze the generation of current ringing by identifying its root cause and propose a machine learning-based methodology for its analysis. Experimental results are employed to validate the model, and the ensuing discussions and conclusions underscore the key findings.

Achieving high-performance, multi-objective control in power electronic converters can pose a significant challenge. Model Predictive Control (MPC) is a powerful algorithm recognized for its simplicity and flexibility. One of its primary strengths lies in its capability to decouple and enable the control of different variables, as well as incorporate constraints within its cost function. Achieving fast output voltage control and stabilizing input voltage, while dynamically limiting current, can be challenging for the DAB power converter. We proposed a dynamic reference approach to enable current limitation and enhance the performance of the MPC algorithm. This facilitates the implementation of high-performance control on both the input and output sides of the converter. To reduce the computational burden of MPC, we introduce a surrogate controller model that acts as an MPC imitator for controlling the DAB. The methodology for creating the surrogate model is outlined through defined steps, which include preliminary testing of the target controller, modeling, data collection, training, and testing. Experimental results are presented to evaluate its performance under various conditions, comparing it with results obtained using the original MPC algorithm and a PI-based controller structure to achieve the desired objectives. The discussion and conclusions highlight the main observations and identify potential areas for future improvement.