Active Disturbance Rejection Control Based on Deep Reinforcement Learning of PMSM for More Electric Aircraft

• Published in: IEEE transactions on power electronics 2023-01, Vol.38 (1), p.1-11
• Main Authors: Wang, Yicheng, Fang, Shuhua, Hu, Jianxiong
• Format: Article
• Language: English
• Subjects: Active control; Active disturbance rejection control(ADRC); Aerospace control; Aircraft; Algorithms; Artificial intelligence; Control methods; Control systems; deep deterministic policy gradient (DDPG); Deep learning; deep reinforcement learning (DRL); flux weakening; Fly by wire control; Heuristic algorithms; Heuristic methods; Machine learning; Modules; more electric aircraft (MEA); Multiple objective analysis; Observers; Optimization; Parameter modification; parameter optimization; Parameter sensitivity; Reinforcement learning; Rejection
• ISSN: 0885-8993; 1941-0107
• DOI: 10.1109/TPEL.2022.3206089

In this article, an active disturbance rejection controller (ADRC) based on deep reinforcement learning (DRL) algorithm is proposed to be used in the flux weakening control (FWC) system of motors for more electric aircraft. Artificial intelligence algorithm is introduced into ADRC motor control system for the first time, and DRL is designed as the automatic tuning for the parameters optimization of ADRC. The interface module scheme is proposed to realize the conversion between the relevant quantities of the control system and the DRL Agent according to the characteristics of ADRC. The parameters are optimized in the form of parameter modification, and a new DRL-ADRC control framework is proposed which can avoid being trapped into local optimum. The ADRC model designed for the speed loop of FWC system are first introduced. An interface module is subsequently built to enable DRL to interact with the FWC system automatically. DRL agent is trained to optimize the internal parameters of ADRC, which have the characteristics of large quantities, weak sensitivity and strong coupling. Deep deterministic policy gradient is used as the strategy of DRL, which can quickly determine the descent gradient and converge the multiobjective optimization problem. Simulation and comparison with classical heuristic algorithms and disturbance rejection methods are carried out to show the superiority of DRL. The feasibility and effectiveness of the proposed control method are verified by experiments on an aerospace motor for MEA.