DESIGN OF MODEL REFERENCE ADAPTIVE CONTROLLER ON BATCH DISTILLATION COLUMN WITH DATA-DRIVEN SYSTEM MODELING USING XGBOOST MACHINE LEARNING

• Main Author: Amalia, Hayati
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79082
• Institution: Institut Teknologi Bandung

The Indonesian government has prioritized the chemical industry as one of the five key manufacturing sectors for ongoing development. This is due to the crucial role of chemical products as essential raw materials for other industrial sectors, such as textiles and plastics. However, in the chemical industry, several challenges in process technology need to be addressed. One particular aspect that receives special attention is the distillation process, utilized for material separation, purification, and waste product disposal. Hence, the development of batch distillation column technology has become increasingly vital. One of the primary issues discussed is how to accurately produce the desired concentration of distillate, irrespective of the nonlinear properties and complex dynamics behaviour. Accurate system modeling is essential for designing an appropriate control system to control the system in accordance with the specified requirements.. However, finding a mathematical model for highly nonlinear systems is generally challenging. Furthermore, designing control systems for nonlinear systems is a significant challenge in dynamic system control. Conventional controllers often struggle to adapt to dynamic changes and uncertainties frequently encountered in such systems. This study attempts to do a data-driven model-free system identification in a batch distillation column using the XGBoost machine learning algorithm. The optimal model structure and parameters are sought to serve as the basis for controller design. The control system is built using the concept of a Model Reference Adaptive Control (MRAC) system integrated with a multilayer perceptron neural network. The neural network parameters continuously adapt based on the error between the reference model response and the system model. The best XGBoost model structure is obtained when using ten input features, which are combinations of input and target delayed by 5 time steps and 4 time steps, respectively, with gamma set to 100 and lambda to 200. By using this combination of parameters, the Mean Absolute Error (MAE) values obtained are 0.222 when applied to the training dataset, 0.247 for testing dataset 1, and 0.425 for testing dataset 2. The designed adaptive model reference control system has been able to track the given setpoint, albeit with a slower settling time compared to the reference model settling time. Specifically, it took 30 seconds to settle when provided with a setpoint of 75%, 128 seconds for a setpoint of 85%, and 390 seconds for a setpoint of 95%, where the reference model settling time was 15 seconds. The best response without overshoot and a steady-state error of 0.3% is achieved when using a neural network architecture with 2 hidden layers, resulting in an MAE of 0.777888. When varying the number of neurons in these two hidden layers, the optimal response without overshoot and a steady-state error of 0.5% is attained with 64 neurons in hidden layer 1 and 16 neurons in hidden layer 2, yielding an MAE of 0.4291.