Self-Attention-Based Convolutional Parallel Network: An Efficient Multi-Input Deep Learning Model for Endpoint Prediction of High-Carbon BOF Steelmaking

Abstract

In this work, a data-driven model for endpoint prediction of basic oxygen furnace (BOF) steelmaking based on both tabular features (information about hot metal, scrap, additives, blowing practices) and time series (curves of off-gas profiles, sonar slagging, and blowing practices) was developed and implemented. The model was designed with the following distinctive artificial intelligence (AI) characteristics: convolutional neural networks, patching embedding, wavelet decomposition, a parallel structure, a self-attention mechanism, a collaborative attention mechanism, and so on. The model presented in this work is named the self-attention-based convolutional parallel network (SabCP) and was applied to high-carbon steelmaking scenarios. SabCP predicts the endpoint of molten steel temperature (Temp) and chemistry (contents of carbon (C), phosphorus (P), and sulfur (S)). For training, validation, and testing, historical data from 13,656 heats were collected. The testing results show that the mean absolute errors (MAEs) of SabCP for temperature and the contents of carbon, phosphorus, and sulfur are 6.374 °C, 7.192 × 10⁻³, 2.390 × 10⁻³, and 2.224 × 10⁻³ pct, respectively, while the mean square errors (MSEs) are 67.345, 1.132 × 10⁻⁴, 1.306 × 10⁻⁵, and 1.298 × 10⁻⁵, respectively, which are lower than those of other published models with same dataset. Relevant importance analyses for tabular features, time series time steps, and channels are also performed. SabCP has been implemented in a prediction module, and the practical results show its strong robustness and generalizability. This model provides significant feasibility for fully eliminating the conventional physical temperature, sampling, and oxygen test (TSO test), which may greatly decrease the cost of BOF steelmaking.