Long short-term memory and gated recurrent unit networks in model predictive control for multivariable processes

• A computationally uncomplicated Model Predictive Control scheme for multivariable processes is described • LSTM and GRU neural networks are used for prediction • Modeling efficiency of LSTM and GRU networks is compared for a chemical reactor benchmark • Control accuracy and computational load of the algorithm are demonstrated This work describes a computationally efficient Model Predictive Control (MPC) algorithm for Multiple Input Multiple Output (MIMO) processes. Two types of neural networks can be used as dynamic models for prediction: the Long Short-Term Memory (LSTM) network and the Gated Recurrent Unit (GRU) network. To obtain a computationally uncomplicated scheme, the algorithm finds online a linear approximation of the predicted trajectory of controlled variables. As a result, we formulate a quadratic optimization MPC task, which can be solved online efficiently and fast. For a chemical reactor, we compare features of LSTM and GRU models in terms of accuracy and complexity. We demonstrate the efficiency of the presented MPC algorithm, taking into account the control quality and computational load. The proposed MPC algorithm achieves control quality comparable to full nonlinear MPC, but with lower computational complexity.