FPGA Implementation of Low Complexity Nonlinear Model Predictive Control Using Deep Learning Approach

Abstract

The main bottleneck in the embedded real-time implementation of nonlinear model predictive control (NMPC) is to solve a complex online optimization problem within a sample time on resource-limited hardware. This computational complexity limits its applicability in real-time control applications running on embedded hardware. Motivated by the recent developments in machine learning methods, our idea is to approximate standard NMPC control law with deep neural networks (DNNs) for a nonlinear system. The developed DNN-based NMPC is implemented on a field-programmable gate array (FPGA) using low-level C/C++ code, where the activation function is evaluated at each sample time. The performance of the proposed DNN-NMPC is demonstrated with a flying robot control application. The proposed DNN-NMPC is compared with the standard NMPC implemented on the same FPGA board (Xilinx's ZYNQ-7000 SoC ZC706). The hardware-in-the-loop (HIL) co-simulation results of both controllers are presented with a detailed analysis of computational complexity, memory utilization, clock speed, and power utilization. Results show that the FPGA-based DNN-NMPC is fast, resource-efficient, and delivers comparable closed-loop performance compared to standard NMPC. The proposed framework allows one to use an NMPC for systems with high computational complexity, significant resource demand and on low-cost embedded hardware like microcontrollers, digital signal processors (DSPs), programmable logic controllers (PLCs), etc.