Fast Nonlinear Model Predictive Control Using Long Short-Term Memory Networks

Abstract

Nonlinear Model Predictive Control (NLMPC) is one of the most convenient and optimal ways to control nonlinear systems but is plagued by issues with computational efficiency. Most of the algorithm is quite fast, but solving an optimization problem online at each time step causes NLMPC response times to be unpredictable and too slow for many systems. To address this issue, this paper uses supervised learning to learn the behavior of an NLMPC controller, replacing the entire control system architecture with a single Long Short-Term Memory (LSTM) network. We discuss the dataset structure and a procedure for generating it and show that the LSTM has comparable performance to NLMPC when deployed to an inverted rotary pendulum system. Computation time for the LSTM during inference is stable and several orders of magnitude higher than that of the NLMPC controller, with typical controller response times around 13 [ms].