End-to-end process optimization in polymer extrusion lines using model predictive control and multi-task learning

Due to the inherent complexity of modern polymer extrusion lines caused by nonlinear, dynamic behavior with numerous influential factors, disturbances lead to higher scrap rates and downtimes. Unfortunately, conventional approaches to address this challenge rely on independent models and local control loops, neglecting the multistage characteristic of extrusion lines and therefore, cannot obtain a global optimum. Consequently, we propose a predictive control design based on a nonlinear and autoregressive multi-task learning model covering the entire extrusion line including local control loops and essential quality measurements. Training on an extended prediction horizon successfully addresses accumulating prediction errors and measurement noise. Due to its capability to predict state trajectories for up to 60 minutes, the proposed methodology enables effective model predictive control. Additionally, using an efficient method for adaptive error compensation based on previous error trajectories increases robustness substantially while the hierarchical control architecture supports global optimization and efficient local control loops. The proposed control design is validated on two polymer extrusion lines. The results show that changing operating points and optimizing process states, such as melt temperature, can be reliably achieved in spite of process disturbances. In comparison to baseline production periods, dominant oscillations are successfully damped by 62 %, reference values are closely followed and process variations are reduced by 41 % to 63 % improving product quality notably. A comparative analysis of disturbance rejection capabilities shows superior results while empirical analyses of stability and runtime further improve its closed-loop applicability. As a consequence, the proposed methodology for predictive control allows to steer extrusion processes accurately, address upcoming issues in advance and generally leads to more efficient processes.