Refinery-wide optimization using neural network surrogate models

Process optimization helps operate chemical processes more efficiently. Currently, the state-of-the-art in petroleum refineries is to use nonlinear unit-wide optimizers in conjuction with a linear optimizer to schedule the fiow between different process units. This approach, however, does not capture all of the potential benefits of optimization. In this work, a structure for performing a nonlinear refinery-wide optimization in an implementable manner is investigated. This structure uses neural networks to model the process in place of detailed, first-principle models. Neural networks are used because they can be solved much faster than detailed, first-principle, differential/algebraic equation (DAE) models.

The detailed models and neural network models necessary to implement this method have been developed. Some discussion of neural network structure is presented, and methods of unconstrained optimization are developed for use in neural network training. Since the refinery optimization problem is subject to a number of nonlinear process constraints, the topic of constrained optimization is discussed as well. The neural network models are optimized and the resulting optimum was compared with the optimum found by optimization of the detailed models. The resulting program was profiled to determine the most computationally-intensive components.

Since neural network training is a time-intensive problem, the use of parallel computers is examined. The platforms tested were a SGI Origin 2000 supercomputer and a cluster of Compaq Alpha computers. Finally, the time required to perform the optimization on a full-scale refinery is estimated.