Reward functions in reinforcement learning applied to field-oriented control of permanent magnet synchronous motors

Abstract

Permanent magnet synchronous motors (PMSM) are used in various industrial applications due to their high efficiency and ability for their current to be precisely controlled. This thesis explores the application of reinforcement learning as an alternative method for current control as opposed to a traditional proportional-integral controller in such motor systems. Using field-oriented control a reward based learning model was trained to act as the current controller with the aim of improving performance and adaptability. However, the results indicate that the trained models performed significantly worse than a tuned proportional-integral controller under the evaluated conditions. This shortcoming is attributed to limitations in the training dataset and the testing environment, which were unable to capture scenarios where reinforcement learning is advantageous. The thesis further explores the main principles of permanent magnet synchronous motor control, reinforcement learning and the use of MATLAB tools for training the models and its implementation in this context.