Big data is a cornerstone component of the fourth industrial revolution, which calls onengineers and researchers to fully utilize data in order to make smart decisions and enhance the efficiency of industrial processes as well as control systems. In practice, industrial process control systems typically rely on a data-driven model (often linear) with parameters that are determined by industrial/simulation data. However, in some scenarios, such as in profit-critical or quality-critical control loops, first-principles concepts that are based on the underlying physico-chemical phenomena may also need to be employed in the modeling phase to improve data-based process models. Hence, process systems engineers still face significant challenges when it comes to modeling large-scale, complicated nonlinear processes. Modeling will continue to be crucial since process models are essential components of cutting-edge model-based control systems, such as model predictive control (MPC).

Machine learning models have a lot of potential based on their success in numerousapplications. Specifically, recurrent neural network (RNN) models, designed to account for every input-output interconnection, have gained popularity in providing approximation of various highly nonlinear chemical processes to a desired accuracy. Although the training error of neural networks that are dense and fully-connected may often be made sufficiently small, their accuracy can be further improved by incorporating prior knowledge in the structure development of such machine learning models. Physics-based recurrent neural networks modeling has yielded more reliable machine learning models than traditional, fully black-box, machine learning modeling methods. Furthermore, the development of systematic and rigorous approaches to integrate such machine learning techniques into nonlinear model-based process control systems is only getting started. In particular, physics-based machine learning modeling techniques can be employed to derive more accurate and well-conditioned dynamic process models to be utilized in advanced control systems such as model predictive control. Along with Lyapunov-based stability constraints, this scheme has the potential to significantly improve process operational performance and dynamics. Hence, investigating the effectiveness of this control scheme under the various long-standing challenges in the field of process systems engineering such as incomplete state measurements, and noise and uncertainty is essential. Also, a theoretical framework for constructing and assessing the generalizability of this type of machine learning models to be utilized in model predictive control systems is lacking.

In light of the aforementioned considerations, this dissertation addresses the incorporation ofprior process knowledge into machine learning models for model predictive control of nonlinear chemical processes. The motivation, background and outline of this dissertation are first presented. Then, the use of machine learning modeling techniques to construct two different data-driven state observers to compensate for incomplete process measurements is presented. The closed-loop stability under Lyapunov-based model predictive controllers is then addressed. Next, the development of process-structure-based machine learning models to approximate large, nonlinear chemical processes is presented, with the improvements yielded by this approach demonstrated via open-loop and closed-loop simulations. Subsequently, the reliability of process-structure-based machine learning models is investigated in the presence of different types of industrial noise. Two novel approaches are proposed to enhance the accuracy of machine learning models in the presence of noise. Lastly, a theoretical framework that connects the accuracy of an RNN model to its structure is presented, where an upper bound on a physics-based RNN model’s generalization error is established. Nonlinear chemical process examples are numerically simulated or modeled in Aspen Plus Dynamics to illustrate the effectiveness and performance of the proposed control methods throughout the dissertation.