Nonlinear model predictive distillation control using an extended neural Hammerstein model

Model Predictive Control has been successfully applied in the chemical and petrochemical industries due to its intuitiveness and constraint handling capabilities. However most currently applied techniques use linear models that are valid only in the neighborhood of the operating point. Model predictive control using nonlinear models does have significant potential for efficient control over a wide operating range. This is particularly important for distillation control which is characterized by highly nonlinear, interactive and nonstationary behavior. The main challenge of nonlinear model predictive control is to develop accurate dynamic models.

Phenomenological modeling is difficult, and computationally intensive. Hybrid models, that combine conventional identification techniques with alternative modeling approaches like neural networks, are favored because of their flexibility, computational efficiency, and ability to learn complex nonlinear mappings in a reasonable time. The Hammerstein modeling strategy simplifies the identification by separating the steady-state and transient components.

In this project, an extended Hammerstein model was developed for use in a nonlinear model predictive control framework. The static nonlinear element of the Hammerstein model was modeled as a feed-forward neural network model, and the nonlinear dynamic element was identified as transfer function models with input-dependent adaptive dynamic parameters. Two distillation columns were modeled: a propylene-propane (C3) splitter operating at base case and at high purity and a toluene-xylene column. Steady-state and dynamic data were obtained from rigorous simulators developed previously. A dynamic model of the C3 splitter at base case using internally recurrent neural networks was also developed.

Nonlinear model predictive control using the extended Hammerstein model was tested on dynamic simulations of each column. Nonlinear model predictive control using the recurrent dynamic model was tested on the C3 splitter at base case. The control performance was compared with that of PI controllers for each column for setpoint and disturbance rejection.