Physics-informed local model networks for the coating weight of hot-dip galvanized steel strips

In continuous hot-dip galvanizing, there is a strong economic incentive to reduce zinc consumption by improving coating weight control. In this context, the long transport delay between the jet-wiping process and coating weight measurements is a major challenge and motivates the use of model-based control strategies, e.g., feedforward control or internal model control. These strategies rely heavily on the accuracy of the underlying model. Thus far, most control applications have used a simple power law model of the coating weight. However, the power law typically cannot cover the entire operating range of a hot-dip galvanizing plant with a single set of parameters. The power law is a simplified version of a physics-based model and usually represents a good local approximation of the jet-wiping process. Therefore, this work explores integrating the power law into control-oriented machine learning models. It is shown that the physics-based knowledge of the process can be systematically incorporated into both a neural network and a hierarchical local model tree (HILOMOT). The proposed physics-informed models are then compared to existing models from the literature, using measurement data from an industrial hot-dip galvanizing plant. This comparison demonstrates how incorporating the power law can enhance a machine learning model both in prediction accuracy and the number of parameters required, with the physics-informed HILOMOT model outperforming all other considered models.