Data-Driven Reduced Order Modeling of Induction Motors Using Dynamic Mode Decomposition and Deep Koopman Theory

The presence of unbalances modifies the nonlinear dynamics of an induction motor drive. Data-driven models are becoming increasingly relevant for model predictive control as they can infer dynamics directly from data, thereby automatically calibrating to the system at hand. In this work, dynamic mode decomposition with control is improved for induction motors by learning a time-variant system. Hereafter, Koopman theory is applied to linearise the state nonlinearity. The algorithms are trained on simulation data using Direct Torque Control. Time-invariant dynamic mode decomposition with control and a symmetrical multiple coupled circuit model are used as reference models. Two orders of magnitude improvement of the stator current prediction is achieved, which is maintained for extensive prediction distances. Moreover, Koopman theory increases longterm stability. The proposed models are validated for torque control in a simulated environment. They are found to be useful for situations where high accuracy is required or where drive dynamics are particularly unknown. In future work, neural architecture optimisation and experimental validation will be explored.