This work focuses on machine learning modeling and predictive control of nonlinear processes using noisy data. We use long short-term memory (LSTM) networks with training data corrupted by two types of noise: Gaussian and non-Gaussian noise, to train the process model that will be used in a model predictive controller (MPC). We first discuss the LSTM training with noisy data following a Gaussian distribution, and demonstrate that the standard LSTM network is capable of capturing the underlying process dynamic behavior. Subsequently, given that the standard LSTM performs poorly on a noisy dataset from industrial operation (i.e., non-Gaussian noisy data), we propose an LSTM network using Monte Carlo dropout method to reduce the over-fitting to noisy data. Furthermore, an LSTM network using co-teaching training method is proposed to further improve its approximation performance when clean data from a nonlinear model capturing the nominal process state evolution is available.

This work addresses the design of distributed model predictive control (DMPC) systems for nonlinear processes using machine learning models to predict nonlinear dynamic behavior. Specifically, sequential and iterative distributed model predictive control systems are designed and analyzed with respect to closed-loop stability and performance properties. Extensive open- loop data within a desired operating region are used to develop Long Short-Term Memory (LSTM) recurrent neural network models with a sufficiently small modeling error from the actual nonlinear process model. Subsequently, these LSTM models are utilized in Lyapunov- based DMPC to achieve efficient real-time computation time while ensuring closed-loop state boundedness and convergence to the origin. Using a nonlinear chemical process network exam- ple, the simulation results demonstrate the improved computational efficiency when the process is operated under sequential and iterative DMPCs while the closed-loop performance is very close to the one of a centralized MPC system.