(545c) Improving Computational Efficiency of Machine Learning-Based Distributed Predictive Control of Nonlinear Processes Using Feature Selection

While machine-learning-based model predictive control (ML-MPC) for nonlinear processes has been studied in many recent works, the computational cost of ML-MPC for large-scale nonlinear processes with high-dimensional input and output space remains a challenge due to the complexity of both the control problems and the ML models. To reduce the complexity of the control problem, ML-based distributed model predictive control (DMPC) with inter-controller communication has been proposed in [1] to improve computation efficiency. However, since full-state information is required for all the controllers in DMPC [2], a machine learning model that accounts for the states of the entire system is needed, which leads to increased model complexity and computation time in both offline training and online prediction. Therefore, to further improve the computational efficiency of ML-MPC, model reduction techniques such as feature selection (e.g., [3, 4]) will be utilized to develop reduced-order ML models in DMPC.

Motivated by the above considerations, we develop reduced-order ML models for large-scale nonlinear processes using feature selection and active learning, and incorporate the ML models within DMPC scheme to improve its computational efficiency. Specifically, sensitivity analysis is first utilized to reduce model dimension by identifying the important connections between model outputs and inputs. The results from sensitivity analysis will be used to develop reduced-order recurrent neural network (RNN) models that predict the process dynamics based on the important input features only. Additionally, to further reduce the model complexity in terms of sample complexity and training time, active learning is employed to enrich the training set by iteratively identifying the most informative training data that could efficiently improve model performance [5]. Subsequently, DMPC systems are designed using reduced-order RNN models to stabilize the nonlinear system at the steady state. Finally, the effectiveness of the reduced-order modeling and predictive control method using sensitivity analysis and active learning is demonstrated using a chemical reactor-reactor-separator process example.

[1] Chen, S., Wu, Z., Rincon, D., & Christofides, P. D. (2020). Machine learning‐based distributed model predictive control of nonlinear processes. AIChE Journal, 66, e17013.

[2] Christofides, P. D., Scattolini, R., de la Pena, D. M., and Liu, J. (2013). Distributed model predictive control: A tutorial review and future research directions. Computer & Chemical Engineering, 51, 21-41.

[3] Dhal, P., and Azad, C. (2022). A comprehensive survey on feature selection in the various fields of machine learning. Applied Intelligence, 52, 4543-4581.

[4] Zurada, J. M., Malinowski, A., and Cloete, I. (1994). Sensitivity analysis for minimization of input data dimension for feedforward neural network. IEEE, 6,447-450.

[5] Qiu, J., Wu, Q., Ding, G., Xu, Y., and Feng, S. (2016). A survey of machine learning for big data processing. EURASIP Journal in Signal Processing, 1-16.