Computationally Efficient Nonlinear Model Predictive Control

For nonlinear systems, Nonlinear Model Predictive Control (NMPC) is preferred to linear Model Predictive Control(MPC) since the nonlinear dynamics of the plant and the control performance index can be incorporated directly. In certain applications the computational resources available for calculating the control solution are severely restricted or the solution is required at high frequency. To overcome these computational challenges this paper presents a computationally efficient update scheme for NMPC using the Forward Dif-ference Generalized Minimum RESidual (FDGMRES) method with a neuro-fuzzy nonlinear dynamic model to describe the plant. Following a description of the FDGMRES approach and a simple case study, an evaluation of the algorithms computational performance is presented using the example of a reference tracking controller for control of a nonlinear Continuously Stirred Tank Reactor (CSTR) system. The online execution time of the FDGMRES algorithm based controller is compared in real time with the more conventional approach of the Sequential Quadratic Programming (SQP) algorithm using Rapid Controls Prototyping hardware.