Recurrent Neural Network based MPC for Process Industries

Abstract

Autonomous operation of industrial plants requires a cheap and efficient way of creating dynamic process models, which can then be used to either be part of the autonomous systems or to serve as simulators for reinforcement learning. The trends of digitalization, cheap storage, and industry 4.0 enable the access to more and more historical data that can be used in data driven methods to perform system identification. Model predictive control (MPC) is a promising advanced control framework, which might be part of autonomous plants or contribute to some extent to autonomy. In this article, we combine data-driven modeling with MPC and investigate how to train, validate, and incorporate a special recurrent neural network (RNN) architecture into an MPC framework. The proposed structure is designed for being scalable and applicable to a wide range of multiple-input multiple-output (MIMO) systems encountered in industrial applications. The training, validation, and closed-loop control using RNNs are demonstrated in an industrial simulation case study. The results show that the proposed framework performs well dealing with challenging practical conditions such as MIMO control, nonlinearities, noise, and time delays.