Dynamic endpoint forecasting in BOF process: a JITL-BP approach integrated with steel grade coding

To achieve the goal of real-time prediction of the basic oxygen furnace (BOF) endpoint and improving the model prediction result by using steel grade information, this paper proposes a composite strategy of target encoding algorithm combined with JITL-BP model. The study found that applying target encoding to steel grade information produced input features with high correlation. The features will significantly improve the model performance. The real-time prediction performance of the JITL-BP model is continuously improved as more samples are added, and it performs much better than BP and SVR when dealing with low-quality data. As the sample input is increased, fluctuations in the JITL sample library gradually decrease, indicating that effective samples are being continuously accumulated. At the same time, the MAE value steadily decreases, indicating continuous improvement in prediction accuracy. This suggests that the JITL model combined with target encoding can effectively enhance the model’s performance in a real-time prediction environment. For the JITL-BP model, the hit rate in predicting the endpoint carbon content within ±15% is 81.27%, while the hit rate for predicting the endpoint phosphorus content within ±15% is 95.13%.