Modeling and Control of a Chemical Process Network Using Physics-Informed Transfer Learning

This work develops a physics-informed transfer learningframeworkfor modeling and control of a nonlinear process network with limitedtraining data. Unlike the conventional transfer learning method thattransfers the knowledge from one process to another process with similarconfigurations, the proposed transfer learning method is to developa machine learning model for the entire process network using theknowledge of some subsystems in the network. Specifically, based onthe machine learning models that have been developed for some subsystemsin the process network with sufficient training data, we develop atransfer-learning-based recurrent neural network (RNN) model for theentire process network by embedding the pretrained models in the overallRNN model, and utilizing physics-informed machine learning techniquesto improve the prediction accuracy by incorporating a priori process-structureknowledge and physical laws into the development of RNNs. Subsequently,transfer learning is used to reduce the computation time of characterizationof the region of attraction for model-based control using RNNs. Finally,two chemical process networks are used to illustrate the effectivenessof the proposed physics-informed transfer learning method.