LSTM and GRU type recurrent neural networks in model predictive control: A Review

Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) neural networks are known for their capability of modeling numerous dynamical phenomena. Model Predictive Control (MPC) refers to a family of advanced control methods in which a dynamical model predicts online the future behavior of the controlled process, and an optimization procedure finds the best control policy. From the point of view of the control quality and the computational complexity of MPC, two issues are crucial: the model structure and the way the model is used in MPC. Both factors determine the resulting control quality and computational complexity of MPC. This article reviews possible methods of using LSTM and GRU type networks in MPC. First, we characterize several model variants that are utilized in MPC. Next, we review possible approaches to MPC based on LSTMs and GRUs, particularly the MPC methods leading to low computational complexity. Stability and robustness issues are also discussed. For a chemical pH reactor, the efficiency of LSTM and GRU models and a few neural network-based MPC algorithms are compared. Finally, we review numerous applications, including applications to real processes, hardware-in-the-loop solutions and example simulation studies.