Steady-State Error Reduction of Reinforcement Learning based Indirect Current Control of Permanent Magnet Synchronous Machines

Abstract:
Deep reinforcement learning (DRL) can achieve favorable dynamic performance compared to conventional control methods. However, steady-state errors are often present. This paper investigates the reduction of steady-state error in DRL-based current control of permanent magnet synchronous machines (PMSMs) by augmenting the integrated tracking error to the observation vector. More specifically, this paper assesses the performance of a DRL-based method under nominal and adverse operating conditions by considering PMSMs with linear and nonlinear magnetic circuits, which exhibit saturation, cross-coupling, and spatial harmonics. The latter include parameter mismatches between the training model and the physical system and misalignment of the dq-frame with respect to the identified position of the d-axis. As shown with the presented experimental results, the DRL-based control method can successfully operate the drive system under all operating conditions, with the steady-state and dynamic performance being similar to that of field-oriented control.