Rapid control prototyping of sensorless control system for PMSM based on multi-agent reinforcement learning and fractional order sliding mode control

Abstract

Based on the overall Field Oriented Control (FOC) control strategy of a Permanent Magnet Synchronous Motor (PMSM), a flexible and efficient control system architecture is developed in this work to achieve superior control performance. Sliding Mode Control (SMC) laws are utilized for both the inner and outer loop, but the typical cascade control characteristic of the system is maintained. Thus, the inner loop (IL) control laws are designed to provide increased flexibility by using fractional order (FO) computation and a response speed that is an order of magnitude higher than that of the outer loop (OL). The optimization of the tuning parameters of these controllers is performed by a computational intelligence (CI) algorithm, more specifically the Improved Grey Wolf Optimizer-Cuckoo Search Optimization (IGWO-CSO). The minimization of the computation time in the implementation of control algorithms is achieved by using a neural network (NN) that estimates the derivative value of the sliding surface in the structure of the SMC type speed controller. A term is added to the control law to cancel global perturbations of the system model, estimated with a Disturbance Observer (DO). Mitigation of the numerical stability problems of the derivative computation is achieved by using a Levant observer tracking differentiator. The use of Multi-Agent Reinforcement Learning (MARL) based on three properly trained Twin-Delayed Deep Deterministic (TD3) RL agents, which provide correction signals overlapping the control signals, contributes to the superior performance of the sensorless control system of the PMSM (SCS-PMSM). These include both parametric robustness to parameter and load torque variations, but also the use of an adaptation law to estimate the stator resistance, which can vary significantly. The superiority of the proposed SCS-PMSM over a benchmark control system based on Proportional Integrator (PI) controllers is demonstrated by following both the Software-in-the-Loop (SIL) and Hardware-in-the-Loop Simulated-Rapid Control Prototyping (HILS-RCP) phases. The realization of an RCP for the proposed RCP SCS-PMSM at different sampling periods corresponding to the implementation in both high performance and low/medium performance Digital Signal Processors (DSPs) is achieved using a SpeedGoat Performance Real-Time Target Machine platform.