Real-time dynamic estimation of carbon content in converter melt pool based on smelting mix data

The carbon content in converter molten steel is crucial to the impact of steel, and in view of the current situation that it is difficult to realize the real-time dynamic estimation of molten steel carbon content throughout the blowing process in the converter smelting process, this paper proposes a real-time dynamic estimation of carbon content in converter molten pool based on smelting mixing data. This method integrates scalar and temporal data from converter steelmaking via a cosine similarity metric strategy, constructing an attention-enhanced temporal convolutional network. It employs a genetic optimization algorithm to determine the optimal model parameters and introduces a sliding window mechanism to address the challenge of training variable-length sequences. The proposed model, named the Cosine Similarity-based Sliding Window Temporal Convolutional Network with Self-Attention (CS-TCN-Attention), achieves real-time dynamic estimation of carbon content in the molten steel bath. The results of ablation and comparison experiments show that under the same parameter settings, the model performs well in terms of prediction accuracy and goodness of fit, with the mean absolute error (MAE) reduced to 0.0736, the coefficient of determination ( R 2 ) improved to 0.9963, and the root mean square error (RMSE) reduced to 0.0879, which significantly improves the accuracy of real-time prediction of molten steel carbon content in the process of converter blowing and provides a reference for the time-series prediction in process industries.