Inferential control using high-order process models with application to a continuous pulp digester

Philip Alexander Wisnewski, Purdue University

Abstract

In the paper industry, the continuous pulp digester poses a challenging control problem. This is in part due to the complex physics that take place in converting wood chips to pulp, the non-uniformity of the feedstock, and the incomplete measurement information of the quality variable, the Kappa #. Given an accurate process model and secondary measurements, an inferential control structure is employed to estimate the Kappa # and compute control action which minimizes its variation around a setpoint. This thesis examines modeling, control structure selection, and application of model predictive control to the continuous pulp digester. The continuous digester model is developed as an extension of the well known Purdue model. A lumped parameter approximation is used to describe the flow transport mechanism of the continuous digester, and a model based on fundamental principles is derived. Mass and volume fractions are re-defined so as to allow fewer simplifying assumptions. The model compares closely with the original Purdue model as well as the Weyerhaeuser Digester Problem model. A systematic approach to selecting the manipulated inputs, secondary measurements, and measurement locations, based on robust control analysis tools, is applied to the Weyerhaeuser Digester Problem. A linear, multi-rate data-sampling model predictive controller using the selected input/measurement pairings provides robust closed-loop performance for the rejection of both deterministic and stochastic unmeasured disturbances. For white liquor composition disturbances, it is determined that the white liquor flowrates should be used as the manipulated inputs rather than liquor temperatures. Given that many industries are relying more heavily on complex, fundamental process models, there exists the need to systematically incorporate these models into the control design. Three model predictive controllers, each using a model of differing degree of complexity, is applied to the continuous digester "plant". The first two are linear models, one obtained utilizing subspace identification and the other obtained from the linearization of the fundamental model. The third MPC controller uses the complex, nonlinear digester model with extended linearization to update a model for future predictions and control computations. A novel model reduction technique is employed to allow for on-line application of the large scale process model.

Degree

Ph.D.

Advisors

Doyle, Purdue University.

Subject Area

Chemical engineering

COinS