Abstract:
Polybutylene succinate (PBS) is a biodegradable plastic known for its strength and versatility in various applications. This research presents a data-driven approach to simulate temperature control in a semi-batch reactor during polymerization, the performance of the proposed approaches was compared against conventional controllers, including PID control and first-principles model MPC control. The study developed neural network model-based predictive control (NNMPC) and multiple neural network model-based predictive control (Multi-NNMPC), using Python and Tensorflow. Neural network models were trained by using a wide range of dynamic data with varying numbers of neurons in hidden layers to investigate the process dynamics under different model complexities. Under nominal conditions, 50 neuron NNMPC demonstrated the most efficient complexity among the tested structures, exhibiting an Integral of Absolute Error (IAE) value of 2,104.77, 20 neuron Multi-NNMPC provided slightly improved performance as IAE reduced to 2,030.52 and the control action trended duplicating MPC control. These approaches addressed the failure of PID control, which caused overshoot and inefficient setpoint tracking. The PID control resulted in polymer over-specification, with the molecular weight reaching almost 14,000 and an IAE value of 3,271.83. In contrast, the optimal temperature control approach of the 50 neuron NNMPC could perform tight temperature control and yield the desired properties of the polymer, significantly outperforming PID control. This research also considers uncertain conditions, including the interference of white noise and model mismatch, all control approaches successfully handled the noise and maintained temperature isothermally, the 50 neuron NNMPC exhibited less aggressive valve movement than PID control, enhancing control action and leading to increased robustness and reduced utility consumption. When model mismatch was introduced to represent reactor fouling, reducing the overall heat transfer coefficient by 30%, the 50 neuron NNMPC achieved faster convergence of control variable to setpoints compared to other controllers. It yielded an IAE of 2,892.41, while MPC showed an IAE of 3,009.59. Moreover, the neural network model demonstrated the ability to learn highly nonlinear dynamics efficiently, enabling the prediction of optimal manipulated variables up to 5 to 20 times faster than a mathematical model using the Sequential Least Squares Programming (SLSQP) method.
