Authors:

Zachary Rasmussen

Date Created:

Not specified

Course Title:

Professor:

Not specified

About Paper:

Distillation column sequences are an important component in industrial plants, however, their energy usage can account for up to 60-70% of the plants' annual capital. Therefore, synthesizing an optimal distillation sequence provides grounds for reducing energy consumption and increasing profit. Synthesizing distillation sequences through a superstructure approach proves to be difficult due to the nonlinear behavior of the columns, so a simplifying surrogate model is required for the problem to be solvable. This research investigates the use of a novel physics-informed neural network that can guarantee linear physical constraints, KKT-hPINN, to capture the behavior of the distillation columns and simplify the problem. The KKT-hPINN model was created by obtaining training and test data through a commercial process simulator, and the model was trained with linear mass flow constraints embedded into the architecture. A distillation sequence optimization problem with non-sharp splits is then solved once with the KKT-hPINN model embedded in the optimization formulation, and again with a standard feed forward neural network. The accuracy of each model is then evaluated using root mean square error to determine the efficacy of using the KKT-hPINN architecture over a neural network. This research presents an alternative to standard surrogate models for applications where it is essential for linear constraints such as mass balances to be conserved. Keywords: [no keywords provided]

Source:

Purdue University / 2024

Topics:

No topics listed

Co-authors:

Zachary Rasmussen