(58g) Comparison of Finite Impulse Response Vs. Subspace Vector Identification in Model Identification of a Fractionator

Model predictive control (MPC) is oricess model based technology to optimize process performance by manipulating controlled variable setpoints of regulatory control loops. One of the most common package for MPC is dynamic matrix control (DMC) which traditionally uses finite impulse response identification (FIR) method to compute steady-state gains matrix.

While MISO (multiple input single output) structure-based FIR ID is superior for low order parametric models, subspace vector ID implements intensive optimization to become a true MIMO (multiple input multiple output) structure and results in a balance model between steady state gains and short-term dynamics. This work presents a quantitative and qualitative comparison of these two model identification techniques in computing steady state gains matrices by using Aspen DMCplus.

Given fractionator column data, the steady state gain matrices are computed with different amount of data to see the effect of data size on the quality of the resulting model. Further, different levels of noise are introduced to the data to test the effect of noisy data on the quality of steady state gains matrix. The results indicate subspace vector ID has led to better models in term of deviations from the full-blown models based on full amount of data without noises.