PMSM MTPA-Control with Reinforcement Learning and CO2 Burden

Paper i proceeding, 2023

Electric machinery often needs to be accurately controlled for efficiency. Maximum Torque per Ampere (MTPA) approach gives rapid & precise control, while assuring the efficiency. To incorporate the MTPA approach, look-up tables (LuT) are populated based on physical principles, synthesizing current combinations in the D-Q plane. Especially for embedded systems like encountered in vehicles, parsing data from densely populated tensors like the LuT is rather resource intensive, which reflects itself in the CO 2 emissions. In this approach supervised learning will be used to devise a regression based surrogate model for LuT by means of Neural Networks (NN). Subsequently, principles of Reinforcement Learning (SARSA) based MTPA control will also be developped in an effort to reduce the training data needs. It will be shown that, the RL based approach provides a valid method for motor control. Finally, a CO 2e emission comparison for chosen approaches will also be reported.