Sequential dynamic optimization of complex nonlinear processes based on Kriging surrogate models

This paper presents a sequential dynamic optimization methodology applicable to solve the optimal control problem of complex highly nonlinear processes. The methodology is based on the use of kriging metamodels to obtain simpler, accurate, robust and computationally inexpensive predictive dynamic models, derived from input/output (training) data eventually generated using the original complex first principles process model (mathematical or analytical model) or from the real system. Then these metamodels can easily take the place of the complex first principles process model in any of the well-tailored computational schemes of sequential dynamic optimization. The results of applying this approach to three well known problems from the process systems engineering area are compared with the ones obtained using the corresponding first principles models, showing how the proposed approach significantly reduces the computational effort required to get very accurate solutions, and so enables the use of dynamic optimization procedures in applications where robustness and immediacy are essential practical constraints.