Data-Driven Control of Electrical Drives: A Deep Reinforcement Learning with Feature Embedding

Abstract

Classical model-based control solutions dominated the research field of numerous electrical drives applications in the past forming a strong basis, since they usually result in control approaches with excellent performance. However, the design of these controllers strongly depends on the available knowledge of the controlled plant, which often leads to the lack of robustness owing to model-dependent nature. To take account of the defect, this work aims to provide a control framework that combines intelligent data-driven-based control protocol with the deep rein-forcement learning technique for electrical drives. Specifically, the two key features of this developed control framework that, first, a data-driven control architecture along with deep rein-forcement learning technique that embedding the features of the agents' inputs is developed to enhance the performance, second, the environment for the current agent is reformulated so as to avoid mutual interference between the two agents (controllers) in training for both speed and current in a dual-loop system. Finally, we demonstrate our solution and highlight its superiority on a case study, and the results presented are promising and motivate further research in this field.