(182e) Reduced-Order Modeling and Predictive Control of Nonlinear Processes Using Machine Learning

Modeling and control of large-scale, complex nonlinear processes have been long-standing research problems in process systems engineering and remain critical challenges in the chemical industry. While machine learning (ML) methods have been extensively studied in recent years to model nonlinear processes and incorporated in machine-learning-based model predictive control (MPC), the computation performance of ML-based MPC (ML-MPC) for large-scale processes with high-dimensional input and output space is still unsatisfactory due to the complexity of ML models [1]. Therefore, model reduction techniques are required to improve the computational efficiency of solving ML-MPC for high-dimensional chemical processes. Autoencoder (AE) has been widely employed as a data-based dimensionality reduction method, which learns to map data in the original high-dimensional space to a latent space of a lower dimension [2]. While AE has been studied for state estimation and modeling distributed parameter systems in MPC [3,4], the incorporation of AE in ML-MPC for modeling and control of large-scale nonlinear chemical systems with stability analysis has not been studied.

In this work, we develop an AE-based reduced-order ML modeling framework for nonlinear chemical processes, and incorporate the model into MPC to improve the computational efficiency [5]. Specifically, an autoencoder is first developed for model dimension reduction by projecting the process states into a low-dimensional space using the data generated from open-loop simulations of the nonlinear system in the original high-dimensional space. Subsequently, the AE model is integrated with recurrent neural network (RNN) to capture the dominant dynamics of the nonlinear system using the low-dimensional data. The AE-based reduced-order RNN models are then used in the Lyapunov-based MPC, under which closed-loop stability is proved. Finally, a diffusion-reaction process is used to illustrate the effectiveness of the autoencoder-assisted reduced-order machine learning-based predictive control scheme.

[1] Vaupel, Y., Hamacher, N. C., Casparo, A., Kevrekidis, I. G., and Mitsos, A. (2020). Accelerating nonlinear model predictive control through machine learning. Process Control, 92, 261-270.

[2] Wang, W., Zhao, M., and Wang, J. (2019). Effective android malware detection with a hybrid model based on deep autoencoder and convolutional neural network. Amb. Intell. Human. Comput., 10, 3035-3043.

[3] Ellis, M. J., and Chinde, V. (2020). An encoder-decoder LSTM-based EMPC framework applied to a building HVAC system. Chem. Eng. Res. Des., 160, 508-520.

[4] Qing, X., Song, J., Jin, J., and Zhao, S. (2021). Nonlinear model predictive control for distributed parameter systems by time-space-coupled model reduction. AIChE, 67, e17246.

[5] Zhao, T., Y. Zheng, J. Gong, and Z. Wu, (2022). Machine Learning-Based Reduced-Order Modeling and Predictive Control of Nonlinear Processes. Chem. Eng. Res. & Des., 179, 435-451, 2022.