(243f) Real-Time Machine Learning-Based Implementation of Model Predictive Control for Nonlinear Processes

In large-scale Industrial Control Systems, there is a persistent need for stable operation and profit optimization. The primary method to achieving this has been through the use of Model Predictive Control (MPC), a method of controlling systems using constrained nonlinear optimization methods to yield desirable control behavior [1]. A major limitation of this method is the computational complexity involved, which scales poorly with increasing system complexity. As the computation time rises, the minimum sampling time for state variables rises, worsening the quality and stabilizability of the system. This can potentially lead to a system being infeasible to control using MPC. Although this issue can partly be resolved through more generous tolerances or through novel optimizations, the difficulty of adapting these methods or finding methods that are sufficiently generalizable for broad use complicates the adoption at scale [2].

To combat these problems, we investigate the use of neural networks (NN) as a tool to improve performance. As opposed to existing research that looks at the use of NNs to model the system dynamics used within MPCs, our approach aims to bypass the computational complexity issues of the MPC by using an NN to approximate the MPC process entirely [3]. The results of this are tested on a non-linear process with miscellaneous variations in the networks complexity, training process, MPC cost functions, and data size to observe the consequences of the design choice. Using a backup stabilizing controller with Lyapunov constraints to ensure stability, it becomes possible to achieve lightweight real-time optimal control while maintaining stability guarantees without the need for specialized strategies during FNN construction or training as has been done in similar works [4].

[1] Forbes, M. G., Patwardhan, R. S., Hamadah, H., Gopaluni, R. B., 2015. Model Predictive Control in Industry: Challenges and Opportunities. In: 9th IFAC Symposium on Advanced Control of Chemical Processes ADCHEM 2015, 48(8), 531–538.

[2] Nouwens, S. A. N., de Jager, B., Paulides, M., Heemels, W. P. M. H., 2021. Constraint-adaptive MPC for large-scale systems: Satisfying state constraints without imposing them. In: 7th IFAC Conference on Nonlinear Model Predictive Control NMPC 2021, 54(6), 232–237.

[3] Ren, Y. M., Alhajeri, M. S., Luo, J., Chen, S., Abdullah, F., Wu, Z., Christofides, P. D., 2022. A tutorial review of neural network modeling approaches for model predictive control. In: Computers & Chemical Engineering, 165, 107956.

[4] Hertneck, M., Köhler, J., Trimpe, S., Allgöwer, F., 2018. Learning an Approximate Model Predictive Controller with Guarantees. In: CoRR, abs/1806.04167.