Grey Box Modeling of a Packed-Bed Regenerator Using Recurrent Neural Networks

摘要

A data-driven modeling approach for a pilot scale Packed-Bed Regenerator is examined and insights are generalized. Training data is generated with a one dimensional physical simulation model, which covers a wide variety of operation conditions including full load and partial load behavior. The NARX Recurrent Neural Network architecture is used to create a model that is able to describe the complex behavior of the regenerator. A grey box modeling approach is proposed that utilizes feedback state variables and incorporates knowledge about the internal behavior of the device. Using this approach, the behavior of the Packed-Bed Regenerator can be described accurately with multi-step ahead predictions. This work presents a first step towards data-driven modeling of dynamic processes in industrial applications. In addition to the presentation of important modeling key points for the proposed grey box model, important steps regarding data preprocessing are identified and insights in the applicability of different Neural Network architectures are discussed.