Modelling for Nonlinear Predictive Control of Synchronous Machines: First Principles Vs. Data-Driven Approaches

In this work, a data-driven modelling approach for synchronous machines is proposed based on the use of long-short term memory (LSTM) neural networks (NNs). Moreover, a comparison between the conventional first-principles and the proposed data-driven modelling approaches is made for the use in nonlinear model predictive controllers. The first-principles modelling is preceded by an illustration of the current and voltage measurements synchronization on a real test bench, an inverter nonlinearity compensation of a 2-level voltage source inverter (VSI), and an angle delay correction to compensate for the unavoidable delay that occurs due to the digital implementation of the control algorithms. The obtained LSTM prediction model is implemented and validated online on a 500 W synchronous motor controlled by a deadbeat controller based on the first-principles nonlinear model of the machine. The presented results yield a good prediction accuracy, and motivate further research on the use of data-driven modelling methods with predictive controllers in the field of power electronics and electrical drives.