Towards Fast Nonlinear Model Predictive Control for Embedded Applications

Real-time implementation of NMPC has its limitations for the embedded applications such as solving a non-convex optimization problem within a sample time, high computational complexity, and limited onboard resources (such as processor speed, memory, power, etc.). To mitigate these limitations, we have proposed application of Successive Online Linearization-based NMPC (SOL-NMPC) to cater the fast sample time requirements of realtime embedded applications. The idea of SOL-NMPC is to derive a convex optimization (Quadratic Programming (QP)) problem at each sample time through the linearization of the nonlinear model which eventually results in a linear MPC formulation. Subsequently, the resulting LMPC is executed online by using state-of-the-art QP solvers. The developed SOLNMPC is implemented on a low-cost STM32 Nucleo development board. The 2-DOF- Helicopter model is used to demonstrate the effectiveness of the proposed SOL-NMPC. Hardware-Inthe-Loop (HIL) co-simulation results demonstrate the implementation feasibility for real-time control applications. The performance is compared with the classical NMPC implemented using the GRAMPC toolbox in C/C++ environment. The comparison results show that: (i) the proposed SOL-NMPC is comparatively faster and takes less computational resources as compared to classical NMPC, (ii) successfully satisfies the constraints, and (iii) however, at the cost of reduced complexity, one needs to scarify with the accuracy as compared to NMPC. Nevertheless, the accuracy obtained by the presented SOL-NMPC is acceptable in embedded applications over the high computational burden of the classical NMPC.