Abstract

In the realm of process control, managing complex systems with limited prior knowledge presents significant challenges, particularly in environments where traditional mechanistic models are either unavailable or computationally prohibitive. This work explores the integration of deep transfer learning with system identification and Model Predictive Control (MPC) to develop control strategies that are both data-efficient and computationally streamlined. Initially, Long Short-Term Memory (LSTM) networks are employed to create approximate MPC controllers by leveraging transfer learning from a source system to a target system, demonstrating that transfer learning can achieve comparable performance to traditional MPC methods with reduced training data. Building upon this foundation, the research advances to incorporate transformer architectures combined with Physics-Informed Neural Networks (PINNs) to quantify the effects of transfer learning for system identification and surrogate control modeling in data-scarce environments. The physics-informed transfer learning approach not only improves predictive accuracy by up to 74% in system identification tasks but also enhances MPC performance by up to 98%, showcasing the impact of integrating domain-specific knowledge with advanced deep learning techniques. Furthermore, this study delves into the specific application of transfer learning for thickener control, which is a critical system in the minerals processing industry for water recovery and sustainable resource management. Given the highly nonlinear dynamics of thickeners, traditional control strategies often struggle to maintain stability and efficiency. By leveraging transfer learning combined with transformer architectures and PINNs, the research demonstrates a framework for surrogate modeling and MPC in thickener processes, achieving up to a 99% improvement in control performance under limited data conditions. This comprehensive approach addresses the dual challenges of data scarcity and model complexity, building a foundation for future scalable and adaptable solutions for industrial applications.

Degree

PhD

College and Department

Ira A. Fulton College of Engineering; Chemical Engineering

Rights

https://lib.byu.edu/about/copyright/