Dynamic Process Flexibility Analysis Using Neural Networks and a Volumetric Flexibility Index

Abstract

Uncertain parameters are common in real-world chemical processes due to inherent variations, underscoring the essential need for operational flexibility. In dynamic process systems, the feasible operation region evolves over time, complicating the assessment of flexibility. Current approaches for evaluating dynamic process flexibility are largely adaptations of techniques used for steady-state flexibility analysis, including the extended active set method and the extended vertex method. These strategies aim to identify the maximum allowable deviations of uncertain parameters from their nominal values. However, such conventional indices may lack reliability when the selected nominal point significantly deviates from the central position and/or when the feasible region exhibits nonconvex characteristics. In this paper, we propose a volumetric flexibility index to the dynamic systems and combine Physics-Informed Neural Network for Control (PINNC) and Convolutional Neural Network (CNN) to determine the flexibility index value. The PINNC model acts as a surrogate for the system’s dynamic model, while the CNN classification network model identifies the feasible region for uncertain parameters. The proposed framework effectively handles nonconvex feasible regions. Its effectiveness and advantages are highlighted through comparisons with existing methods.