Fusion of data-driven and physical modelling for hot-rolling mill force prediction

Abstract

Accurate prediction of rolling force is pivotal for process optimisation and quality control within the steel industry. Conventional prediction models, which are primarily based on physical mechanisms, often struggle to achieve both high precision and strong robustness when confronted with complex process parameters and multi-source heterogeneous data. To address this limitation, this paper proposes a hybrid hot-rolling force prediction method that integrates a data-driven approach with a physical model, optimised by a Laplace-approximated neural additive model (LA-NAM). First, a data acquisition module was established, incorporating data derived from mechanistic analysis, data generated by a finite-element model and an augmented sample dataset produced by a generative adversarial network. Concurrently, a deep neural network was constructed as the primary prediction model. The LA-NAM was then utilised to provide interpretability and uncertainty estimation, further optimising the model's parameters. Experimental results demonstrate that, compared to conventional physical models or standalone data-driven methods, the proposed hybrid approach exhibits significant advantages in terms of prediction accuracy, adaptability to complex conditions and compatibility with physical mechanisms. This research not only furnishes an effective intelligent prediction tool for the hot-rolling process but also offers a viable methodology for academic research and engineering practice involving multi-source data fusion and small-sample conditions.