Fast Nonlinear Model Predictive Control Using LSTM Networks: A Model Linearisation Approach

Abstract

This work describes a fast Model Predictive Control (MPC) algorithm in which Long Short-Term Memory (LSTM) networks are used to model dynamical processes. To obtain a computationally simple quadratic optimisation MPC task, a linear approximation of the model is repeatedly determined on-line using an original linearisation method that is specially tailored for the LSTM model. For a benchmark polymerisation process, it is shown that the described approach results in more precise prediction and better control quality than the classical model linearisation method. It is also shown that the described algorithm gives very similar control quality to that observed in MPC with nonlinear optimisation.