In this study, artificial neural networks are adopted to perform multi-step predictions of the power consumed by the refiner of a thermo-mechanical pulping process specialized in medium-density fiberboard production. In this way, the obtained model can be integrated within a model-based control. The refining process is characterized by a large number of variables, and artificial neural networks are a well-established methodology for multivariate data processing, able to identify the non-linear hidden relationship between monitored variables. Both a Long Short-Term Memory network, with stability guarantees, and a Transformer one are implemented due to their ability to model the evolution of dynamical systems. Simulation results prove both models' multi-step prediction capabilities.

A two-layer hierarchical control scheme entirely based on Model Predictive Control (MPC) is proposed for the control of a compressed air network. The high-level exploits the air demand prediction and - through a hybrid MPC - defines the optimal unit commitment and compressor operating points, minimizing the network energy consumption over a long time horizon. At the low-level, compressors are controlled independently to track the references received from the upper layer in the presence of actuation constraints. The proposed control solution can be applied to different network configurations. Simulation results prove the capabilities of the strategy when compared to traditional techniques used nowadays in industry.