Learning-based model predictive current control for synchronous machines: An LSTM approach

Abstract

In this work, a data-driven model predictive control (MPC) approach for the current control of synchronous machines is presented. The model of the motor is represented via a long-short term memory (LSTM) neural network (NN). The model is obtained purely from collected data and doesn’t include any physical knowledge. As an online optimization using the obtained data-driven model is not easily implementable in the available sampling time, the neural model is used to solve an MPC problem offline. Finally, the control policy is learned via another computationally implementable NN that runs in real-time as a current controller. The proposed data-driven MPC controller is tested experimentally, and is bench-marked against MPC schemes that incorporate the well-known physically-based first-principles linear and nonlinear model11The term first-principles model is used throughout the paper to call the machine model derived from the physical laws governing the functionality of the synchronous motor, whether it is the mostly used LTI model or the nonlinear model with the inductance/flux current-dependant maps.of the machine.