Physics-Informed Neural Networks for Industrial Applications: A Case Study in Thermal Power Prediction of an AHU for a Topcoat Process

Modelling approaches for industry are typically classified into data-driven and physics-based models. The first type often struggles to interpret the complexities of physical systems and require enough data to capture patterns, while the second type depends on detailed information, which might be only partially available or subject to uncertainty. The recent methodology of Physics-Informed Neural Networks (PINNs) is investigated as a hybrid approach that integrates neural network flexibility with the definition of physical constraints. The work focuses first on a review of PINNs applications in industry, highlighting their advantages over traditional methods in terms of efficiency and versatility. Then, methods and tools for PINN implementation are described and a general mathematical formulation is provided. In order to validate the model, a case study concerning the prediction of the thermal power of an industrial Air Handling Unit (AHU) is proposed, detailing the implementation process. The results show good performance compared to traditional data-driven techniques, highlighting the main advantages of this hybrid approach, such as the efficiency even with simple physical models and a smaller amount of training data. Future research directions include the implementation of variants of PINNs with different neural network models or different forms of integration of physical knowledge.