Gray‐Box Modeling for the Optimization of Chemical Processes

Abstract

The availability of predictive models for chemical processes is the basic prerequisite for offline process optimization. In cases where a predictive model is missing for a process unit within a larger process flowsheet, measured operating data of the process can be used to set up such models combining physical knowledge and process data. In this contribution, the creation and integration of such gray‐box models within the framework of a flowsheet simulator is presented. Results of optimization using different gray‐box models are shown for a virtual cumene process.

Modeling, Simulation, and Optimization 4.0 for a Distillation Column

“…This data is essential for a sound model selection, validation and adjustment. Furthermore, tools to setup data driven models, e.g., using machine learning methods could be useful to either complement models (graybox modeling, e.g., ) or to replace complex and time‐consuming models by simpler surrogate models (e.g., , ). All these steps require an efficient data management for a multitude of solutions.…”

“…In these cases, a 2D projection is no longer intuitive. Navigators to investigate different simulation results, generated for example in model validation or adjustment of different operating points or by sensitivity analysis and allowing to setup empiric or machine learning based data driven models are useful for graybox and surrogate modeling as will be discussed. Furthermore, versioning allows to track the development of the model and documentation helps to recall the rationale behind made decisions.…”

“…This concept will be called graybox modeling in the following. An example of such an application is given in . Modeling of real behavior leads sometimes to complex rigorous models, which are time‐costly to evaluate or hard to solve.…”

The Potential of Hybrid Mechanistic/Data‐Driven Approaches for Reduced Dynamic Modeling: Application to Distillation Columns

“…The idea to combine mechanistic equations with efficiently evaluable data-driven model parts has gained substantial attraction in PSE. During the last decades, hybrid model structures have widely been used to replace complex equation systems that would arise from purely first-principle mechanistic models and thereby accelerate process simulations and/or optimizations [40,41]. In particular, numerous authors used surrogate models to replace thermodynamic property calculations in order to speed up process simulations/optimizations.…”

Was den Digitalen Zwilling zum genialen Kompagnon macht

“…Darum sind Methoden zu entwickeln, die mittels maschinellen Lernens aus Prozessdaten vorhandene Prozessmodelle ergänzen – ein Beispiel ist in zu finden – und Hinweise auf die Notwendigkeit weiterer Messstellen liefern. Die Kombination aus den mechanistischen Prozessmodellen mit den neu entwickelten Methoden zur Einbeziehung von Prozessdaten ermöglicht dann eine genaue Beschreibung des Gesamtprozessverhaltens und damit eine umfassende, verfahrenstechnische Optimierung.…”