A CatBoost‐Based Modeling Approach for Predicting End‐Point Carbon Content of Electric Arc Furnace

Abstract

Developing the prediction model of the end‐point carbon content of the electric arc furnace (EAF) is an effective way to reduce the adjustment frequency of liquid steel composition and shorten the smelting time. Previous data‐driven models lack effective handling of the missing values in EAF production data. This may be the main reason why model accuracy is difficult to improve. This article proposes a novel modeling method based on the CatBoost algorithm with two‐stage optimization. In the preprocessing session, empirical and empirical‐cumulative‐distribution‐based outlier detection (ECOD) methods are utilized to extract input features and reject outliers. The end‐point carbon content prediction model is built based on CatBoost. The generative adversarial imputation nets (GAIN) method is used in the first optimization stage to handle the missing values. In the second optimization stage, recursive feature elimination (RFE) is used to select the final features, and whale optimization algorithm (WOA) is used to optimize the parameters of the CatBoost model. After verification with actual production data, the two‐stage optimized CatBoost model demonstrates excellent performance compared with other methods, with an R2 of 0.903, mean absolute error of 0.021, root mean squared error of 0.043, and 90.34% hit ratio within ±0.05% error range.