Knowledge- and data-driven prediction of blast furnace gas generation and consumption in iron and steel sites

Abstract

Accurately predicting the blast furnace gas (BFG) generation and consumption is a big challenge for steel sites due to the low quality of industrial data and the variability of equipment operation states. To address this issue, a knowledge- and data-driven (KD) prediction method for BFG generation and consumption is proposed. This method employs dynamic time warping to describe the relationship between the historical database and the training set for BFG data streams with the same operation state, enabling the prediction model to learn the characteristics of BFG generation and consumption across all operation states. The method also considers the fluctuation characteristics of BFG generation and consumption, applying the 3σ criterion, mean substitution method, and Kalman filtering model to improve data quality. Additionally, by incorporating process knowledge, a rectified linear unit is proposed to refine the prediction results. The results demonstrate that the KD prediction model outperforms traditional data-driven models. Specifically, the KD–genetic algorithm (GA)–extreme gradient boosting (XGBoost) model delivers the best performance for BFG generation prediction across all operation states, with a mean absolute error (MAE) of 193.01 m³/min, symmetric mean absolute percentage error (SMAPE) of 0.22 %, mean absolute percentage error (MAPE) of 0.44 %, and R² of 0.9957. For BFG consumption prediction in the hot blast stove group, the KD–back propagation neural network (BPNN) model demonstrates superior performance, achieving MAE of 69.03 m³/min, SMAPE of 0.24 %, MAPE of 0.22 %, and R² of 0.9996.