A Novel Control Scheme for Variable Load Drive Systems with Reinforcement Learning

In this paper, the authors proposed a novel Reinforcement Learning based controller for the optimal control of a variable frequency drive (VFD) system for an induction motor. Traditional research in this area concentrates on a different level of analysis, that of operational control – e.g. different PWM schemes to create an idealize output – while this research looked on operational optimization – to select the action given the state of the device. The agent is based on a class of algorithm known as Policy Gradient, and the entirety of the study was done through computer simulation. The agent was trained and tested on a series of complex spatial-temporal load sequences to demonstrate its robustness and generalizability. In addition, the agent does not need any explicit knowledge of the device model itself, it could learn just by observing a set of user defined observations of the environment and receiving a reward signal as feedback for its action. The novel control scheme outperforms an uncontrolled benchmark significantly in multiple areas such as power factor, slip performance, among others by a significant 25% margin.