Reinforcement learning based control approach for PMSM drives — Theory, concept, design and realizations

With the recent advancements made in Artificial Intelligence, it is possible that Reinforcement Learning based data driven methods can become a next generation technology to control electrical drives instead of the classical model-based techniques. However, providing the correct setup and hyperparameters for training the agent is challenging and usually not evident. The manuscript aims to present the main steps of the design workflow to apply Reinforcement Learning for controlling the current of a PMSM drive. Along these steps, every major design consideration is summarized in addition to providing the necessary theoretical background. Furthermore, a thorough examination of the hyperparameter search stage is provided with several experiments in Python on the most influential parameters of the environment. The most optimal setup based on the compared results is selected and evaluated using a testing environment in Matlab that resembles a real-life application scenario. The paper summarize the findings in the form of practical guidelines that are crucial for achieving a high level of performance, and they also minimize the time needed for the implementation and training of the RL agent. • Data-driven reinforcement learning based control of electric drives. • Detailed description of the steps for building a custom environment for controlling electrical drives. • Practical guidelines for the correct setup and hyperparameters for training of the agent. • Comparative results about 5 hyperparameters in a Phyton based environment based on 400 different trainings. • Demonstrative results in a MATLAB based environment simulating a real-life application scenario.