Identification of dynamic models for the Shell benchmark problem

Abstract

Dynamic process and disturbance models are identified for the 2×2 distillation column described in the summary paper by Cott. Process models are identified using both classical least squares and the newer AUDI identification algorithm developed at the University of Alberta. The identified models are validated and gain-adjusted in the frequency domain by comparing the magnitude spectrum of the models, | G (e −iω )|, versus the spectrum obtained by fast Fourier transformation of the original plant I/O time series. The noise/disturbance models are identified using standard time series tools and all models are verified by showing that the filtered residuals (measured process output minus model based estimate of the output) are essentially white noise. The estimated SISO models are all first-order and very close to the actual process whether compared on the basis of step-response coefficients, model parameters or frequency domain plots. The control problem is so simple that two standard PID feedback loops with one-way (lead/lag) interaction compensation produce results that are essentially identical to a 2×2 multivariable, model-based (GPC) controller.