Interpretable multivariable forecasting of underground loading and haulage equipment productivity using advanced time-series models

Underground loading and haulage productivity reflects interactions among production, equipment utilization, delays, and maintenance. This study developed an interpretable multivariable framework for forecasting next-calendar-day equipment productivity at an underground mine in Pataz, Peru. We processed 15,955 operational events, yielding 742 equipment-day observations from two dump trucks and three load–haul–dump machines. Equipment-specific series were chronologically divided into training (70%) and test (30%) sets. SARIMAX, XGBoost, Temporal Convolutional Network (TCN), and Temporal Fusion Transformer models were optimized using four expanding-window folds and compared with naïve persistence. The TCN recorded the lowest aggregate errors: mean absolute error (MAE) = 5.865 t/h, root mean squared error (RMSE) = 11.250 t/h, symmetric mean absolute percentage error (sMAPE) = 38.78%, and mean absolute scaled error (MASE) = 1.257, reducing benchmark MAE and RMSE by 10.4% and 22.0%, respectively. However, model differences were not statistically significant. Predictive performance was moderate ( R 2 = 0.53 ), and the TCN tended to smooth abrupt changes, overpredict near-zero productivity, and underpredict some extreme values. Shapley Additive Explanations (SHAP) identified temporal trend, material moved, and recent equipment utilization as the most influential predictors. Cross-conformal intervals achieved 91.60% empirical coverage against the nominal 95%. Although retained because it recorded the lowest aggregate errors, the TCN should be interpreted alongside uncertainty intervals, equipment status, and operational expertise.