Process identification and model predictive control using hybrid polynomial models

More stringent environmental regulations and economic competition have forced the chemical industry to use advanced process control techniques (i.e. model based control) to allow operation in regions with significant instabilities and nonlinearities. An important example currently used in industry is model predictive control (MPC) which makes use of a model of the open-loop process. Success of an MPC algorithm depends a great deal on the model used to represent the process and the model structure is a critical issue. In this research we developed a hybrid modeling approach which reproduces the open-loop dynamic behavior of the system and yields good MPC performance. We show that an existing hybrid approach can be easily modified such that linear vs. nonlinear parameter estimation techniques are used in process identification. Hybrid models were identified for a two-dimensional continuous stirred tank reactor system exhibiting saddle-node bifurcation, Hopf bifurcation, and hysteresis behavior. We required that the hybrid model capture the complete bifurcation behavior of the open-loop system as a function of control input - new model validation criterion. MPC using a hybrid model outperformed both the linear and empirical models. The hybrid modeling approach was also applied to a five-dimensional polymerization system exhibiting a Hopf bifurcation. This is the first example of a hybrid model for a high-order system (i.e. greater than three). We show that the five-dimensional hybrid model took less computational time during the MPC optimization step than the exact nonlinear first principles model. This improvement indicates the potential benefit of our hybrid modeling approach for MPC.