Machine-learning-assisted prediction of coke strength after reaction for coke plants

Coke strength after reaction (CSR) is a critical parameter in metallurgical applications, and its accurate prediction is essential for optimizing coal blends and coking processes. This research develops a data-driven method to model CSR by eight input variables, including moisture content, volatile matter, ash percentage, sulfur content, maximum fluidity, plastic layer thickness, mean maximum reflectance (MMR), and the basicity index. A dataset comprising 630 coal samples with diverse properties was analyzed using advanced techniques such as Pearson correlation analysis, the Monte Carlo outlier detection technique in data integrity assessment, and machine-learning models with five-fold cross-validation. Multiple algorithms were implemented, including random forests, decision trees, adaptive boosting, convolutional neural networks, support vector regression, multilayer perceptron-artificial neural networks, and an ensemble learning approach, with hyperparameter optimization and evaluation metrics like mean squared error, R 2 , and mean and average absolute relative error. The random forest model decisively outperformed all other contenders, demonstrating its superior predictive power through consistently high R 2 values and minimal error rates. Furthermore, Shapley additive explanations analysis revealed the influence of each input variable, with volatile matter having a predominantly negative effect on CSR, while features like MMR and moisture showed positive correlations. This systematic methodology underscores the importance of robust data assessment and machine-learning models in enhancing predictive accuracy for CSR.