(458f) Employing Weight Sharing in Recurrent Neural Network Construction for Model Predictive Control of Nonlinear Processes

Machine learning (ML) algorithms have received increasing interest in engineering disciplines due to their ability to capture complex relationships and nonlinearities, with recurrent neural networks (RNN) being investigated widely in the context of dynamical systems due to them being inherently designed to capture time-dependencies in time-series data [1,2]. Nevertheless, their adaptation in practical scenarios is oftentimes hampered by the exorbitant amount of training data required to construct accurate ML models since installation of sensors can be restricted in terms of costs as well as the type of variable being measured. Furthermore, the majority of ML modeling is carried out for individual processes using historical data obtained from the particular process of concern and the resulting model cannot then be used to describe a broader class of processes, even those with similar configurations, which is often the case in process systems.

Transfer learning (TL) is a technique involving the migration of knowledge from a data-rich source process to a data-scarce target process. This method has been proven to be effective in overcoming limited data availability when modeling chemical processes and enables the transfer of established physical insights to new processes [3,4]. Specifically, transfer learning entails adjusting a pre-trained model, initially fitted to a data-abundant source process, to accommodate a target process with less voluminous data. This method seeks to improve both learning efficiency and training accuracy compared to the basic approach of training a machine learning model for the target process from scratch without leveraging any prior knowledge, essentially treating machine learning modeling as a completely opaque process. In this study, we introduce a transfer learning-based RNN architecture, specifically employing a weight-sharing RNN model, to integrate physical process knowledge into RNN modeling and training. Following this, the proposed weight-sharing RNN model is integrated into the design of a model predictive controller (MPC) to offer future state predictions for an optimization problem aimed at improving process performance in terms of closed-loop stability and setpoint tracking. Finally, the RNN-MPC is applied to an example in chemical process control to compare its closed-loop performance compared to controllers utilizing a conventional RNN (C-RNN) model.

References:

[1] Wong, W.C., Chee, E., Li, J., Wang, X., 2018. Recurrent neural network-based model predictive control for continuous pharmaceutical manufacturing. Mathematics 6 (11), 242.

[2] Zheng, Y., Wang, X., Wu, Z., 2022. Machine learning modeling and predictive control of the batch crystallization process. Ind. Eng. Chem. Res. 61, 5578–5592.

[3] Amabilino, S., Pogány, P., Pickett, S.D., Green, D.V., 2020. Guidelines for recurrent neural network transfer learning-based molecular generation of focused libraries. J. Chem. Inf. Model. 60 (12), 5699–5713.

[4] Xiao, M., Hu, C., Wu, Z., 2023. Modeling and predictive control of nonlinear processes using transfer learning method. AIChE J. 69 (7), e18076.