Enhanced prediction of end-point carbon content in electric arc furnaces using Bayesian optimised fully connected neural networks with early stopping

This study developed a Bayesian optimisation-enhanced fully connected neural network (BO-EFCNN) model with an early stopping mechanism to predict the end-point carbon content in electric arc furnace (EAF) steelmaking. Traditional approaches, including mechanistic and data-driven models, often encounter limitations such as the need for complex parameter tuning or extensive data. The proposed method integrates Bayesian optimisation to fine-tune hyperparameters and early stopping to prevent overfitting, ensuring effectiveness even with small sample sizes. Production data from a 100-ton EAF was preprocessed by eliminating outliers using the Isolation Forest algorithm and handling missing values with the K-nearest neighbour method. Relevant features were selected based on metallurgical principles and Pearson correlation analysis. The BO-EFCNN model was trained and evaluated using datasets of 500 and 5000 heats, respectively. Results demonstrated that the BO-EFCNN significantly outperformed benchmark models such as Support Vector Regression, Random Forest, Multivariate Linear Regression, and mechanistic models, particularly with smaller datasets. The model exhibited high accuracy and robustness, maintaining performance stability across different sample sizes. Testing over three consecutive months indicated the model's practical applicability, though a slight decline in accuracy over time suggests the need for periodic retraining. This research underscores the potential of BO-EFCNN for accurate and efficient EAF steelmaking predictions, especially in data-constrained environments.