Enhancing Nickel Matte Grade Prediction Using SMOTE-Based Data Augmentation and Stacking Ensemble Learning for Limited Dataset

To address the limited data availability and low predictive accuracy of nickel matte grade models in the early stages of facility operation, this study introduces a unique stepwise prediction methodology that integrates data augmentation and ensemble learning, specifically tailored for limited industrial datasets. Predicting matte nickel grade accurately is critical for nickel sulfate production, a key precursor in cathode manufacturing. However, in newly adopted facilities, operational data are scarce, posing a major challenge for conventional machine learning models that require large, well-balanced datasets to generalize effectively. Moreover, the nonlinear dependencies between raw material composition, operational conditions, and metallurgical reactions further complicate the prediction task, often leading to high errors in traditional regression models. To overcome these challenges, this study introduces an innovative approach that integrates feature engineering, Gaussian noise augmentation, SMOTE regression, and a stacking ensemble model, using XGBoost (2.0.3) and CatBoost (1.2.7). First, input variables were refined through feature engineering, followed by data augmentation to enhance dataset diversity and improve model robustness. Next, a stacking ensemble framework was implemented to mitigate overfitting and enhance predictive accuracy. Finally, SHAP, an XAI technique that quantifies the impact of each input variable on the model’s predictions based on cooperative game theory, was employed to interpret key process variables, offering deeper insights into the factors influencing nickel grade. The experimental results demonstrate a substantial improvement in prediction accuracy, with the R2 coefficient increasing from 0.3050 to 0.9245, alongside significant reductions in RMSE, MAE, and MAPE. The proposed methodology not only enhances predictive performance in data-scarce industrial environments but also provides an interpretable framework for real-world process optimization. These findings validate its applicability to nickel matte operations, offering a scalable and explainable machine learning approach for metallurgical industries with limited data availability.