Model Discrimination and Selection in Evolutionary Optimization of Batch Processes with Tendency Models

Batch processes increasingly feature products with short market windows that make the development of a detailed kinetic model unattractive in terms of both time and economy. Tendency modeling is already established as a systematic methodology for timely optimization of batch processes using a gray-box model approach. In this work, the problem of effective discrimination among alternative tendency models is addressed using rank correlation methods and pair-wise bisection approach to model concordance. The Kendall tau statistics is used to measure model correlation and concordance with regards to alternative optimum predictions. Less promising tendency models are gradually eliminated using a softmax criterion which trades off exploitation with exploration for experiment planning.