• closed-loop control
• control
• control process
• control systems
• deep learning
• feedback control
• gru (gated recurrent unit)
• model predictive control (mpc)
• optimization
• recurrent neural networks (rnns)
• system identification

This thesis explores the application of Neural Network-based Model Predictive Control (MPC) in the realm of automatic control systems, with a particular focus on addressing the challenges posed by nonlinear dynamics and input multiplicities.
The research journey begins with a comprehensive exploration of system identification principles, elucidating the distinguishing characteristics that differentiate linear and nonlinear systems. Building upon this foundation, the thesis introduces the concept of artificial neural networks, categorizing them into feed-forward and recurrent types, and providing an overview of existing typologies. Among feed-forward neural networks (FFNN), the Multilayer Perceptron (MLP) and Radial Basis Function (RBF) architectures are presented and studied. Similarly, among recurrent neural networks (RNN), the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) architectures are compared.
Subsequently, the thesis examines the potential applications of neural networks in automatic control, highlighting their versatility and efficacy in navigating complex control tasks. An overview of Model Predictive Control (MPC) follows, accentuating its unique features when integrated with neural networks. A practical application of MPC is explained and implemented, consisting of three parts: Disturbance Estimator, Steady-State Optimization, and Dynamic Optimization. Each part is accompanied by its mathematical model, providing a comprehensive understanding of the MPC implementation.
Moreover, the thesis delves deeper into the theoretical underpinnings of MPC, elucidating the principles behind each component and their interplay within the control system framework. The Disturbance Estimator is designed to predict and compensate for external disturbances that may affect the system's performance, ensuring robustness and stability. The Steady-State Optimization module aims to optimize control inputs to achieve desired steady-state conditions, while the Dynamic Optimization module adapts control actions dynamically to address changing system dynamics over time. Detailed mathematical models and algorithms are provided for each module, offering insights into their functionality and practical implementation.
The empirical investigation unfolds through an in-depth exploration of a case study involving a single Continuous Stirred Tank Reactor (CSTR) system. This reactor, governed by reversible first-order reactions, presents inherent complexities due to input multiplicities and nonlinear dynamics. Through meticulous comparative analysis, three distinct models—GRU, ARMAX, and SS—are constructed and evaluated against a benchmark control system. Notably, the Recurrent Neural Network (RNN) model emerges as a promising solution, effectively guiding the system to desired setpoints while reducing performance indices such as Integrated Square Error (ISE) and Total Variation (TV). However, the computational complexity of the RNN model results in longer simulation times.
The findings of this thesis offer valuable insights into the practical applicability of Neural Network-based MPC in managing complex reaction systems. By providing nuanced solutions for addressing nonlinear dynamics and input multiplicities, this research contributes to advancements in industrial process control and optimization. Additionally, it underscores the importance of leveraging neural network models in enhancing control performance and addressing challenges in automatic control systems.