Inferential Model Predictive Control Using Statistical Tools

Abstract

With an ever increasing emphasis on reducing costs and improving quality control, the application of advanced process control in the bulk chemical and petrochemical industry is steadily rising. Two major areas of development are model-based control strategies and process sensors. This study deals with the application of multivariate statistical techniques for developing soft-sensors in an inferential model predictive control framework. McAvoy (2003) has proposed model predictive statistical process control (MP-SPC), a principal component (PC) score control methodology. MP-SPC was found to be very effective in reducing the variability in the quality variables without using any real-time, on-line quality or disturbance measurements. This work extends McAvoy's formulation to incorporate multiple manipulated variables and demonstrates the controller's performance under different disturbance scenarios and for an additional case study. Moreover, implementation issues critical to the success of the formulations considered such as controller tuning, measurement selection and model identification are also studied. A key feature is the emphasis on confirming the consistency of the cross-correlation between the selected measurements and the quality variable before on-line implementation and that between the scores and the quality variables after on-line implementation. An analysis of the controller's performance in dealing with disturbances of different frequencies, sizes and directions, as well as non-stationarities in the disturbance, reveals the robustness of the approach. The penalty on manipulated variable moves is the most effective tuning parameter. A unique scheme, developed in this study, takes advantage of the information contained in historical databases combined with plant testing to generate collinear PC score models. The proposed measurement selection algorithm ranks measurements that have a consistent cross-correlation with the quality variable according to their cross-correlation coefficient and lead time. Higher ranked variables are chosen as long as they make sufficiently large contributions to the PC score model. Several approaches for identifying dynamic score models are proposed. All approaches put greater emphasis on short term predictions. Two approaches utilize the statistics associated with the PC score models. The Hotelling's statistic and the Q-residual information may be used to remove outliers during pre-processing or may be incorporated as sample weights.

The process dynamics and controller performance results presented in this study are simulations based on well-known, industrially benchmarked, test-bed models: the Tennessee Eastman challenge process and the azeotropic distillation tower of the Vinyl Acetate monomer process.