Study on performance of Machine-Learning-Based prediction model for slab reheating furnaces

Based on 8297 data samples from a 1450 mm hot-strip mill reheating furnace in a Chinese steel plant, XGBoost and LSTM models using four sets of input variables had been developed and used to predict furnace discharge temperature, energy consumption of per ton steel, and oxidation burn rate. It was found out that the LSTM model performed well in predicting furnace discharge temperature and oxidation burn rate, with coefficients of determination (R²) exceeding 0.95. The XGBoost model was superior in predicting energy consumption, achieving R² values above 0.94 and stable prediction trends. Comparative analysis indicated that LSTM was more reliable for predicting time-dependent parameters (such as discharge temperature and oxidation burn), while XGBoost provided higher accuracy for static parameters (such as energy consumption). Further investigation revealed that LSTM effectively captures time-related patterns due to its gated mechanism. In contrast, XGBoost performed better on static features due to its ability to optimize feature combinations. Based on these findings, a hybrid LSTM-XGBoost model was proposed. In this combined model, LSTM deals with time-series data, and XGBoost processes static data. Applying the combined model to furnace control can further improve prediction accuracy. This study provides theoretical guidance and data support for optimizing reheating furnace operations and enhancing resource efficiency in the steel industry.