Identification and predictive control of a multistage evaporator

A recurrent neural network-based nonlinear model predictive control (NMPC) scheme in parallel with PI control loops is developed for a simulation model of an industrial-scale five-stage evaporator. Input–output data from system identification experiments are used in training the network using the Levenberg–Marquardt algorithm with automatic differentiation. The same optimization algorithm is used in predictive control of the plant. The scheme is tested with set-point tracking and disturbance rejection problems on the plant while control performance is compared with that of PI controllers, a simplified mechanistic model-based NMPC developed in previous work and a linear model predictive controller (LMPC). Results show significant improvements in control performance by the new parallel NMPC–PI control scheme.

Research highlights
► A differential recurrent neural network model is developed for an industrial 5-stage evaporator.
► An efficient training algorithm based on the Levenberg-Marquardt method and automatic differentiation is applied to the network.
► The developed model is used as the internal model for a nonlinear model predictive controller implemented in a parallel NMPC-PI scheme.
► The performance of the new NMPC-PI scheme is compared with that of a PI controller, a linear MPC and a simplified mechanistic model based NMPC to show superiority.