BOF endpoint forecasting via informer architecture for multivariate time series data

For Basic Oxygen Furnace (BOF) steelmaking, machine learning models are employed to predict endpoint carbon, sulfur, phosphorus contents, and endpoint temperature. Based on four traditional machine learning models of linear regression (LR), ridge regression (RR), random forest regression (RFR), and support vector machine (SVM), it is found that the optimized training set does not improve model performance. To address the issue of excessive data optimization, this study innovatively introduces the Informer model into the BOF process, as it can learn directly from raw time process data without the need for extensive preprocessing, while capturing complex long-term dependencies in production sequences. Four traditional machine learning models, long short-term memory (LSTM), gated recurrent unit (GRU), and Informer model were trained and compared based on an unoptimized original dataset. The Informer model demonstrated superior performance, achieving a significant quantitative improvement over traditional models (which typically show hit rates of 55–60%): The probability of the prediction error of the end-point carbon, sulfur, and phosphorus contents being within ±15% reaches over 80%, and for endpoint temperature within ±3% was 92%. The data fluctuation pattern of Informer is highly consistent with that of the original data. The adjustment of the internal parameters of the Informer model has little impact on the model performance, and it has strong generalization ability. In addition, compared with other models, the Informer model can simultaneously predict all endpoint parameters with significantly less computational time.