Spatial and Sequential Deep Learning Approach for Predicting Temperature Distribution in a Steel-Making Continuous Casting Process

Abstract

Continuous casting is the procedure of the successive casting for solidification of the steel, which contains several cooling processes along the caster to coagulate the molten steel. It is such a rule of thumb that strand surface quality and casting productivity is highly dependent on temperature control. A finite-difference method is one of estimating temperature distribution, yet it hinders the process control efficiently. Song, et al. suggest a multimodal deep learning approach for prediction of the temperature. However, sequential and transient phenomena of solidifying steel are not considered, which makes it difficult to estimate the sequential heat-transfer characteristics in the whole process of the steel concretion. Herein, a deep learning model is proposed to predict the temperature distribution by taking into account both transient and steady-state characteristics. The proposed model addresses both spatial and sequential information by incorporating a convolutional neural network (CNN) and a recurrent neural network (RNN). Our quantitative and qualitative results show considerable predictive performance improvement against baseline models. Furthermore, the proposed model is applicable in a real-world steel-making industry by providing real-time temperature prediction, whilst retaining a lower computational cost.