Study of Direct Neural Control of an Induction Motor Trained in a Digital Twin

Objective: to assess the feasibility of direct neural control of an induction motor drive without a fully specified mathematical model of the motor and to determine whether this approach can reduce active power consumption while maintaining speed-control quality comparable to classical field-oriented control. Methods: the study was carried out on a digital-twin model of the inverter-induction motor-controller system. A reinforcement learning agent was trained to generate control actions in the d-q reference frame, and its performance was compared with FOC using the stator RMS current, active power, and speed-control error as evaluation criteria. Practical significance: the trained policy can potentially be transferred to an embedded drive-control platform. This reduces the dependence of control quality on accurate parameter identification and makes the approach relevant for energy-efficient electric drives.