Embedded Implementation of Deep Learning-based Linear Model Predictive Control

Abstract

Model predictive control (MPC) has emerged as an excellent control strategy owing to its ability to include constraints in the control optimization and robustness to linear as well as highly non-linear systems. There are many challenges in real-time implementation of MPC on embedded devices, including computational complexity, numerical instability, and memory constraints. Advances in machine learning-based approaches have widened the scope to replace the traditional and intractable optimization algorithms with advanced algorithms. In this paper, a novel deep learning-based model predictive control (DNN-MPC) is presented. The proposed MPC uses recurrent neural network (RNN) to accurately predict the future output states based on the previous training data. Using deep neural networks for the real-time embedded implementation of MPC, on-line optimization is completely eliminated leaving only the evaluation of some linear equations. Closed-loop performance evaluation of the DNN-MPC is verified through hardware-in-loop (HIL) co-simulation on ARM microcontroller and a 4x speed-up in computational time for a single iteration is achieved over the conventional MPC. Detailed analysis of DNNMPC complexity (speed and memory requirement) is presented and compared with traditional MPC. Results show that the proposed DNN-MPC performs faster and with less memory footprints while retaining the controller performance.