Reinforcement Learning-Based Deep Q Direct Torque Control with Adaptable Switching Frequency Towards Six-Step Operation of Permanent Magnet Synchronous Motors

Abstract:
Deep Q direct torque control (DQ-DTC) for permanent magnet synchronous motors is a reinforcement learning-based finite-control-set optimal control approach, which enables a model-free control requiring minimal drive system knowledge. The reinforcement learning agent chooses an action for the controlled system in consideration of a policy, which is learned in a data-driven training process to maximize the outcome of a reward function. This paper extends the originally introduced reward function of the DQ-DTC framework by a term that allows to determine the optimal zero-voltage vector and to define the resulting average switching frequency. Further, a term is introduced to learn the six-step operation mode more easily. The investigation is carried out in simulation with two drive systems, one fractional horsepower motor and one highly utilized motor for electric traction applications.