Uniting neural network-based control and model predictive control: Application to a large-scale nonlinear process

Abstract

This work proposes a method to overcome the issue of nonlinear model predictive control (MPC) requiring practically infeasible computation times for large-scale systems. In particular, the use of Neural Networks (NN) to approximate nonlinear MPC calculated control actions in a real-time closed-loop implementation with externally enforced stability guarantees is explored. Using Lyapunov-based stability constraints, the reduced computational complexity of NNs paired with the ability to train using MPC that would be infeasible to apply in real-time systems (due to the use of a large prediction horizon to ensure good closed-loop performance) enables the training of an NN-based approximate control policy that directly substitutes MPC. With a stabilizing fallback controller available, this NN controller enables real-time stabilizing control of high-dimensional nonlinear systems. To demonstrate this, Aspen Plus Dynamics, a dynamic chemical process simulation software, is used to create a large-scale nonlinear chemical process example. Using an NN trained off of an offline MPC using a first-principles model and a large prediction horizon, a comprehensive study of the resulting closed-loop behavior is carried out to evaluate the closed-loop stability, performance, and robustness properties of the approach.