Interpretable Machine Learning for Industrial Energy Forecasting—A Comparative Study of Random and Temporal Data Splits

Abstract

Industrial energy forecasting supports production scheduling, peak-load management, and energy-efficiency decisions, yet the reliability of reported performance depends heavily on whether the validation procedure reflects actual deployment. Using 35,040 fifteen-minute records from a Korean steel plant, this study constructs 34,368 leakage-free forecasting samples and compares random and temporally ordered splits across a persistence baseline, ridge regression, random forest, and XGBoost. TreeSHAP is further used to identify influential predictors and examine the stability of feature rankings. Random forest achieves the lowest mean absolute error under both protocols, with MAE values of 4.665 kWh for the random split and 4.171 kWh for the temporal split. However, temporal folds exhibit substantially greater variation, revealing uncertainty that would otherwise be obscured when observations from different operating periods are mixed. TreeSHAP shows that previous-period energy use, lagged reactive power, and time-of-day information are the dominant predictors, while the feature rankings remain highly consistent across the two protocols. These findings suggest that random splitting is useful for measuring average interpolation performance under a broadly stable distribution, whereas temporal validation provides a more deployment-relevant assessment of forward generalization.