(14e) Phased Lstm-Based Model Predictive Control: Handling Asynchronous and Delayed Measurements

Delays and data loss are common problems faced during information exchange in Networked Control Systems (NCS), which can compromise the system stability [1]. Model Predictive Control (MPC) is among the plethora control strategies that can be used to handle time delays and data loss in NCS. Over the years, with the increasing availability of industrial data, there has been growing interests in adopting machine learning (ML) models for MPC usage, in particular, RNN-based MPC [2, 3]. However, in the presence of delays and data loss, the process state measurements may appear to be irregular, with missing data at certain sampling times. These irregularities pose challenges for RNNs in learning the system's dynamics [4]. Thus, several variants to the standard RNNs have been proposed. In particular, Phased Long Short-Term Memory (PLSTM), a modification of the standard Long Short-Term Memory (LSTM) unit, has shown promising results in processing irregularly sampled data [5].

In this work, we develop novel machine learning modeling and predictive control techniques for nonlinear chemical systems subjected to asynchronous and delayed measurements in both offline and online data collection. Specifically, a PLSTM network is used to learn the process dynamics amidst the irregularities in the data, during the offline training process. The generalizability of a ML model (i.e., model’s predictive ability on unseen data) is the key to analyzing the closed-loop performance of ML-based MPC. As the generalizability of PLSTM has yet to be explored, this work will theoretically study the generalization performance of PLSTM, on the basis of statistical machine learning theory, to better understand the capabilities of PLSTM models. The PLSTM model is then employed to forecast the evolution of states for a Lyapunov-based MPC (LMPC) that is designed to account for data loss and delays in real-time implementation. Finally, two chemical processes, including an extractive dividing wall column and a continuous stirred tank reactor, are used to demonstrate the effectiveness of PLSTM modeling and predictive control methods.

References:

[1] R. A. Gupta and M.-Y. Chow, "Overview of Networked Control Systems," in Networked Control Systems: Theory and Applications, F.-Y. Wang and D. Liu Eds. London: Springer London, 2008, pp. 1-23.

[2] Z. Wu, A. Tran, D. Rincon, and P. D. Christofides, "Machine-learning-based predictive control of nonlinear processes. Part II: Computational implementation," AIChE Journal, vol. 65, no. 11, p. e16734, 2019/11/01 2019.

[3] Y. Pan and J. Wang, "Model Predictive Control of Unknown Nonlinear Dynamical Systems Based on Recurrent Neural Networks," IEEE transactions on industrial electronics (1982), vol. 59, no. 8, pp. 3089-3101, 2012.

[4] P. B. Weerakody, K. W. Wong, G. Wang, and W. Ela, "A review of irregular time series data handling with gated recurrent neural networks," Neurocomputing, vol. 441, pp. 161-178, 2021.

[5] D. Neil, M. Pfeiffer, and S.-C. Liu, "Phased lstm: Accelerating recurrent network training for long or event-based sequences," Advances in neural information processing systems, vol. 29, 2016.