This work focuses on applying machine learning modeling on predictive control of nonlinear processes with noisy data. We use long short-term memory (LSTM) networks with training data from sensor measurements 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 by reducing the impact of noise. 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 overfitting to noisy data. Furthermore, an LSTM network using co-teaching training method is proposed to further improve its approximation performance when noise-free data from a process model capturing the nominal process state evolution is available. A chemical process example is used throughout the study to illustrate the application of the proposed modeling approaches and demonstrate their open- and closed-loop performance under a Lyapunov-based model predictive controller with state measurements corrupted by industrial noise.

Electrochemical reaction processes attract increasing attention as a promising chemical process alternative to achieve green and sustainable chemical manufacturing due to its property that can produce chemical products without directly burning fossil fuels. Among various electrochemical reaction processes, CO\textsubscript{2} reduction provides a possibility to capture the CO\textsubscript{2} gas in the atmosphere and effectively relieve the global warming crisis. However, due to the complex, stochastic, and nonlinear nature of the electrochemical reactions, modeling an electrochemical process using first-principles is very challenging. To investigate the physical-chemical phenomena of the electrochemistry of reduction of CO\textsubscript{2}, the Laboratory of Electrochemical System Engineering at UCLA has developed a gastight rotating cylinder electrode (RCE) cell to decouple mass transport phenomena from the intrinsic kinetics of the electrochemical reactions. Specifically, this RCE cell allows manipulation of two variables, the applied potential and rotation speed of the cylindrical electrode, which determines the mass transport profile and reaction kinetics of the electrochemical reactions, respectively. This design further enables the development of an advanced process control system for the reactor to control the desired output states by adjusting the two key input variables.

However, the absence of the first-principle model for the electrochemical reactor is a significant challenge to develop the process control system. To this end, considering the complexity of the process, neural network (NN) models are developed and used in this thesis work to capture the input-output relationship of the reactor and provide a data-based alternative to the unavailable, first-principle models. On the other hand, in a data-driven approach, the performance of the NN model is determined by the quality of the training data. Therefore, the inevitable noise, caused by experimental uncertainty, can corrupt the collected data and negatively impact the NN modeling task. Motivated by this concern, various methods, such as the Monte Carlo dropout and co-teaching algorithm, are adopted in this thesis to improve the robustness of the NN models against noisy data. Eventually, this dissertation demonstrates an advanced model predictive (MPC) scheme based on experimental data-driven NN models that capture the input-output relationship of the electrochemical reactor. Furthermore, the Koopman operator method is adopted to perform on-line linearization to the NN model to improve the computational efficiency of the MPC and enable its real-time application to the experimental electrochemical reactor. Finally, simulations, open-, and closed-loop experiments are conducted to demonstrate the overall implementation and successful performance of the proposed NN models and MPC schemes.