Channel oriented approach for multivariable model updating using historical data

摘要

The model-plant mismatch (MPM) can be responsible for poor control performance. This can be solved by locating which channels (i.e., pars of MVs-CVs) are suffering the highest MPM, and then, proceed with the model updating using the same historical data used for the assessment and diagnosis steps. This paper proposes a method for updating models based on the nominal error: a benchmark that quantifies the model discrepancy by comparing the measured output of a system with its corresponding nominal output, i.e., the output of the closed-loop with no MPM or unmeasured disturbances. The main advantages of the method are to avoid usual multivariable model identification and use simpler SISO structures, thus reducing the workload of the model maintenance. The effectiveness of the method is illustrated using the quadruple-tank process with a nonminimum phase operating point to explore multivariable characteristics of the final updated model. All the required methodologies for assessment, diagnosis, and model maintenance are also presented in the paper and can be applied in the same historical data. No additional plant perturbations are required to improve the model. Although it is not limited to Model Predictive Control (MPC), the proposed methodology can be successfully applied to MPC assessment, diagnosis, and model maintenance.