Abstract
Accurate production forecasting in industrial hydrometallurgy is essential for process optimization yet is often hindered by the scarcity of extensive historical data. This study demonstrates the effectiveness of classical machine learning models as a data-efficient and interpretable alternative to complex deep learning methods for predicting total copper mass. We evaluated four models—Random Forest, Gradient Boosting, Decision Tree, and Linear Regression—using a methodology centered on two key strategies: synthetically expanding a limited 150-day dataset into 10,000 simulated cycles (approximately 1.5 million data points) via data augmentation, and engineering 10-day lag features to provide the models with a temporal perspective for a 10-step-ahead forecasting task. The results revealed exceptional predictive accuracy, with ensemble techniques proving superior. The Random Forest model emerged as the top performer, achieving an R² of 0.974, an MAE of 0.088, and an RMSE of 0.111, closely followed by Gradient Boosting (R² of 0.971). All models successfully captured the distinct 150-day cyclical dynamics of the production process, showing a near-zero phase lag (0.00 ± ≤0.05 days). While performance on new, independent data requires further validation, this work establishes a robust and transparent framework for developing reliable forecasting tools in data-limited industrial environments.