Data-Driven Iron Ore Planning : An Exploration of Machine-Learning Models in Operational Mine Planning

Sammanfattning

Efficient production planning in underground mining requires accurate estimation of operational performance. This thesis investigated whether Machine Learning models could be used to predict loading cycle time in LKAB’s underground iron ore mine in Kiruna using historical operational data. Several regression-based and ensemble learning models were evaluated, including Ridge regression, LASSO, Random Forest and XGBoost. The modeling workflow included data extraction, feature engineering, filtering of operational interruptions and chronological validation. Additional experiments were conducted to evaluate temporal robustness, generalization across mining areas and aggregation of predictions to daily planning metrics. The results showed that filtering extreme cycle times significantly improved predictive performance, where a seven-minute time cap produced the best results. Among the evaluated models, XGBoost achieved the strongest overall predictive accuracy at event level. However, the study also demonstrated that prediction of individual loading cycles remained challenging due to substantial operational variability. When event-level predictions were aggregated to daily operational metrics, predictive performance improved substantially, particularly for total daily loading workload. This suggests that Machine Learning models may provide greater practical value for aggregated operational forecasting and production planning than for precise prediction of individual loading events. Feature importance analysis and segment-based exploration further indicated that loading performance was influenced by spatial, temporal and operational factors such as transport distance, previous operational conditions, shift and mining level. Overall, the results demonstrate that Machine Learning methods can complement existing production planning practices by providing data-driven operational insights and planning-relevant workload estimates in underground mining systems.