Data-Enabled Predictive Control for Nonlinear Load Compensation in DC Motor Drives

Abstract

This paper compares linearized model predictive control (MPC), data-enabled predictive control (DeePC), and Koopman-based data-enabled predictive control (K-DeePC) for controlling the speed of a DC motor driving a centrifugal fan. The quadratic fan load introduces a nonlinearity that a linear MPC prediction model cannot represent exactly. On the other hand, a Koopman lifting dictionary that includes system nonlinearities makes the motor dynamics perfectly linear in the lifted state space, giving K-DeePC a structural advantage: the fan load is exactly in the span of the Koopman Hankel matrix. A multi-step reference profile spanning three distinct operating points demonstrates that K-DeePC achieves the lowest integral absolute error and nearly uniform steady-state tracking error across all reference levels, while linearized MPC accumulates a speed-dependent bias and standard DeePC shows operating-point-dependent performance asymmetry.