Attention-Based Convolutional Aggregation: An Efficient Model for Off-Gas Profile Forecasting and Dynamic Pre-Control of BOF Steelmaking

This study proved that the curves of carbon monoxide (CO), carbon dioxide (CO2), and CO + CO2 in the off-gas profile were forecastable, and realized a 32-s-ahead forecasting for them. It established a technical foundation for addressing the delay in off-gas profile display and for enabling pre-control in BOF steelmaking based on the forecasted curves’ features. First, a data pre-processing method was proposed based on the characteristics of the off-gas curves, where there are many samples, but each sample contains limited time-steps. It is termed the mixed-batch approach. The importance of the time series’ channels and time-steps were also analyzed by models with attention mechanism. Then, a deep-learning model is proposed to forecast the dynamic off-gas profile, named attention-based convolutional aggregation (ABCA). It incorporates artificial intelligence (AI) techniques, such as aggregation structures, causal dilation convolution, attention mechanisms, residual connections, etc. Its forecasting coefficient of determination (R2) values for the curves of CO, CO2, and CO + CO2 reached 0.9386, 0.8566, and 0.9428, respectively, while the mean squared errors (MSEs) values were 47.3884, 11.9314, and 54.3583, respectively. These results outperform the benchmark state-of-the-art (SOTA) models. Additionally, ABCA was implemented in a forecasting tool for external validation. The results of external validation showed that ABCA has good forecasting accuracy and robustness. What is more, approaches in four aspects of pre-control of BOF steelmaking process with forecasted off-gas profile were also provided as pre-control examples.