Modeling and Control of a Chemical Process Network Using Physics-Informed Transfer Learning
Ming Xiao, Zhe Wu
Abstract
This work develops a physics-informed transfer learning framework for modeling and control of a nonlinear process network with limited training data. Unlike the conventional transfer learning method that transfers the knowledge from one process to another process with similar configurations, the proposed transfer learning method is to develop a machine learning model for the entire process network using the knowledge of some subsystems in the network. Specifically, based on the machine learning models that have been developed for some subsystems in the process network with sufficient training data, we develop a transfer-learning-based recurrent neural network (RNN) model for the entire process network by embedding the pretrained models in the overall RNN model, and utilizing physics-informed machine learning techniques to improve the prediction accuracy by incorporating a priori process-structure knowledge and physical laws into the development of RNNs. Subsequently, transfer learning is used to reduce the computation time of characterization of the region of attraction for model-based control using RNNs. Finally, two chemical process networks are used to illustrate the effectiveness of the proposed physics-informed transfer learning method.