Data‐Driven Process Simulation Using Connected Surrogate Unit Models Exemplified on a Steam Methane Reforming Process

Abstract

Process simulation based on physical models often faces computational problems with respect to convergence, especially if the underlying flowsheets are complex. The use of data-driven surrogate models connected to flowsheets promises to overcome these challenges. Using the steam methane reforming process, this paper presents the development of surrogate models -artificial neural networks -for the key units of the process that are subsequently connected to form the entire flowsheet. The accuracy of the individual surrogate models is analyzed based on the test error; the accuracy of the flowsheet is evaluated by a benchmark process simulation performed in Aspen Plus . Therefore, the predicted key variables, here outlet temperatures and compositions, are compared to the benchmark. It is shown that their maximum error is below the typical measurement error. The comparison of the accuracy of the surrogate-based flowsheet simulation with the Aspen Plus simulation proves to match very well, as long as the training ranges of the underlying surrogate models are not violated. The promising results of this paper pave the way for future work, such as the optimization of process parameters or superstructure optimization.

Critical analysis of the use of white-box versus black-box models for multi-objective optimisation of small-scale biorefineries

“…The Aspen Plus model could be used as a source for generating data in order to, subsequently, train a new, surrogate model based on the obtained dataset. The surrogate model can range from a response surface model (RSM) in the case of a limited dataset, or an artificial neural network (ANN) when a larger dataset is available (Schack et al, 2021).…”

Real‐Time Interactive Navigation on Input‐Output Data Sets in Chemical Processes

“…In [3], an overview of surrogate modeling in chemical process engineering is provided, while in [4] a survey regarding the use of machine learning is given. Note that not only a single surrogate could be adopted to describe the entire flowsheet, but it is also possible to interconnect multiple metamodels, each describing a single process unit [5,6]. This makes a surrogate (or a combination of them) a good choice not only for what-if analyses, but also for being embedded in multi-criteria optimization [7,8].…”

Real‐Time Optimization Using Machine Learning Models Applied to the 4,4′‐Diphenylmethane Diisocyanate Production Process

“…A solution to this problem is surrogate modeling. Data is generated by the simulators, and then classical regression or machine learning methods (e.g., artificial neural networks) can be used to identify black-box models that represent the process while being computationally inexpensive [4][5][6].…”