A Hybrid L1–L2 Twin Support Vector Regression Model with Local Strategy for Basic Oxygen Furnace Steel‐Making Endpoint Prediction

The accurate prediction of the endpoint carbon temperature in basic oxygen furnace (BOF) steel‐making is the key to determining the final stage of steel‐making. In this work, a machine learning model for endpoint carbon content and temperature prediction in BOF is constructed and implemented. The model is designed with the improved twin support vector regression (TSVR) incorporating the local prediction strategy. The model proposed in this work is abbreviated as HLTSVR and is applied to a 260 t BOF steel‐making scenario. First, to ensure the quality and availability of the data, the BOF data collected by the sensors are processed by using K‐nearest neighbors imputation, interquartile range, and gray relational analysis. Second, a variant of TSVR (abbreviated as HLTSVR) was proposed by strictly followed the structural risk minimization principle and using a robust loss function, avoiding the inverse of matrix and extra kernel‐generated surface. Moreover, to tackle the low accuracy of global prediction models, a sub‐interval modeling strategy was designed, transforming the global prediction into a local prediction. Based on the simulation of the actual steel‐making production data, the HLTSVR model incorporating the local prediction strategy demonstrates the best performance in predicting endpoint carbon content and temperature among the six models.