A data-driven model as a simulation platform for reinforcement-based dryness control of a pilot paper-making machine using active learning

The SteamDry project focuses on developing superheated steam drying technology (SSD) for materials such as paper, board, tissue, and nonwovens. With electrified CO2 emission-free manufacturing and a growing emphasis on renewable energy, SSD aims to reduce energy consumption of the drying process. The control system for the drying process will be designed to optimise product quality, production rates, and energy efficiency. This presents a multi-criteria optimisation challenge. The project utilises reinforcement learning rather than depending on traditional control methods, which can be conservative due to uncertainties. This method enables an agent to learn through trial and error while ensuring that actions taken are safe and prevent catastrophic outcomes.
To establish such strategies, we first identify a model of the existing drying technology at the VTT pilot plant as a simulation platform to test our toolchain. We examine two drying processes: impingement drying and through-air drying of paper. Various working points were assessed, and paper samples were collected. Specific paperweights and moisture levels were determined in the laboratory. Selection operator (Lasso) regression approach was initially employed to identify the model's relevant characteristics and to avoid overfitting. This method offers the advantages of enhanced predictive power and improved model interpretability. The second step involves assessing model uncertainty using a Gaussian regression method incorporating measurement uncertainties. These uncertainties are employed for the targeted improvement of the model through active learning. Moreover, this uncertainty will later on guide the learning process of the reinforcement-based energy optimal dryness control. This approach will be adapted when the new superheated steam drying system is installed at VTT.