Modeling and control of power electronics converters using machine learning

Abstract

In this thesis, a novel black-box modeling technique for power electronics converters using machine learning is proposed. The conventional modeling techniques may be impractical for over-the-counter converters that lack accurate data sheets. A purely data-driven model is developed using Long Short-Term Memory (LSTM) networks, and then its performance is compared against Gated Recurrent Units (GRUs). The comparison demonstrates that both LSTM and GRU are viable options depending on the desired balance between performance and accuracy. Additionally, to improve computational efficiency, transfer learning is employed to speed up the training process for converters with similar topologies. MATLAB was selected for developing and training the machine learning models due to its robust toolboxes for control systems and deep learning, offering seamless integration with Simulink for dynamic system simulations. Its flexibility and computational efficiency made it a suitable choice over other platforms for implementing LSTM networks and DRL algorithms, while Hypersim facilitated the transition to real-time testing. Furthermore, Hypersim, a popular real-time simulation software for electrical systems and power electronics, is mainly designed for mathematically based models, which can be limiting for commercial converters that lack detailed data sheets. Since Hypersim does not natively support the proposed black-box modeling method, an algorithm for real-time simulation in Hypersim is developed. This algorithm bridges the gap between the machine-learning-based converter models and real-time simulation environments and it provides a significant step forward for hardware-in-the-loop as well as practical implementation. The study also presents a model-free control strategy for power electronics converters using Deep Reinforcement Learning (DRL). Through broad simulation and experimental results, the DRL-based control method is compared against conventional Model Predictive Control (MPC), a conventional non-linear control method. The DRL approach is shown to not only eliminate the need for an accurate mathematical model but also to be more resilient to parameter mismatch, variations in operation points, uncertainty, and noise. Ultimately, the DRL method is applied to a more complex converter, the Hybrid Pakced UCell (HPUC) converter, which generates 23 voltage levels using a single DC source. Control and voltage balancing in this converter are challenging, however, the DRL method demonstrated advantages over traditional MPC in terms of robustness under parameter mismatch, noise, and other disturbances. This research presents new opportunities for utilizing the power of emerging machine learning techniques to solve challenging problems in the domain of modeling and control of power electronic converters.

Date	18 Jun 2025
Original language	American English
Awarding Institution	École de technologie supérieure
Supervisor	Kamal Al-Haddad (Supervisor)