Neural network implementation of model predictive control with stability guarantees

This work explores the use of supervised learning on data generated by a model predictive controller (MPC) to train a neural network (NN). The goal is to create an approximate control policy that can replace the MPC, offering reduced computational complexity while maintaining stability guarantees. Through the use of Lyapunov-based stability constraints, an MPC can be designed to guarantee stability. Once designed, this MPC can be used to generate a dataset of various state-space points and their resulting immediate optimal control actions. With the MPC dataset representing an optimal control policy, an NN is trained to function as a direct substitute for the MPC. The resulting approximate control policy can then be applied in real-time to the process, with stability guarantees being enforced through post-inference validation. If, for a given set of sensor readings, the NN yields control actions that violate the Lyapunov stability constraints used in the MPC, the control action is discarded and replaced with stabilizing control from a fallback stabilizing controller. This control architecture is applied to a benchmark chemical reactor model. Using this model, a comprehensive study of the stability, performance, robustness, and computational burden of the approach is carried out.