A learning-based generalized dynamic predictive control for PMSM position servo system with uncertain dynamics

This paper presents a learning-based generalized dynamic predictive control (GDPC) approach for a permanent magnet synchronous motor (PMSM) position servo system, aiming to achieve intelligent and fine-grained operation under uncertain environments. Considering the limitations of the human-configured generalized predictive control (GPC) scheme with the fixed horizon to cope with diverse operating conditions, the investigated strategy further integrates a deep reinforcement learning (DRL)-based horizon online-updating mechanism. In particular, an extended state observer (ESO) is first constructed for the estimation of lumped disturbance to modify the deviation of the prediction model. The analytical solution of the benchmark GPC algorithm is then obtained by solving an optimization problem of the designed performance index. To optimize the prediction horizon of GPC, a DRL agent is then trained offline. Real-time horizon adjustment is finally implemented on an experimental setup that combines a digital signal processor (DSP) and a Beckhoff controller. A series of simulations and experiments validate the efficacy of the proposed control approach. The proposed GDPC method utilizes a deep deterministic policy gradient (DDPG) algorithm to optimize the prediction horizon in real time, achieving improved control performance under varying operating conditions.