A Cross‐Attentive Gated Fusion Model for Steel Billet Tapping Temperature Prediction in Heating Furnace

Highly accurate prediction of steel billet tapping temperature is an essential technology for improving quality of steel products. However, the mechanism model lacks adaptability to the complex production environment. A data‐driven model typically utilizes a single kind of input feature, which fails to establish sophisticated connections between input attributes and prediction objects. Therefore, a novel billets tapping temperature prediction model called cross‐attentive gated fusion network (CA‐GFNet), which integrates static and dynamic data from the heating furnace, has been designed. In the proposed model, a long short‐term memory neural network is leveraged to process dynamic variables, a multilayer perceptron is employed to process static variables, a cross‐attention module is developed to establish correct connections between two types of features and generate interactive features, and a gated fusion module is used to suppress redundant interactive features. Results show that the hit rate of CA‐GFNet is 17.6% higher than models using a single type of feature in the predicted error range −5, 5 °C. Additionally, the discrepancies in hit rates across three error ranges are within 0.5% when input variables are replaced by a new dataset from other furnaces. These experiments validate the superiority of the proposal in prediction accuracy and generalization.