Temporal Representation versus Model Complexity in Appliance Energy Forecasting: An Explainable and Robust Machine Learning Study

Introduction: Reliable short-term energy forecasting supports energy monitoring, demand management, and operational planning in building-energy systems. However, forecasting performance can depend as much on temporal representation and recent consumption history as on model complexity, while many benchmark studies emphasize algorithm comparison without systematically examining these effects. This study develops an explainable, leakage-aware forecasting framework using the UCI Appliances Energy Prediction dataset, treated as a high-frequency building-energy monitoring benchmark rather than a direct industrial process dataset.

Methods: Ten-minute appliance-energy observations were modeled using chronological data partitioning, persistence and regression baselines, tree-based machine-learning models, and temporally engineered lag and rolling-history features. Performance was assessed using MAE, RMSE, MAPE, and R². Robustness was examined through feature-group ablation, multi-horizon forecasting, rolling-origin evaluation, peak-demand analysis, compact-feature modeling, and external validation. Model interpretation used SHAP and permutation-based importance to quantify the contribution of recent consumption history, environmental variables, and time-related predictors.

Results: The enhanced experiments showed that recent appliance-consumption history was the dominant source of predictive information. XGBoost produced the strongest fixed-split predictive performance overall, while Ridge regression remained highly competitive and persistence was difficult to outperform on some error criteria. Ablation and interpretability analyses consistently showed substantial deterioration when consumption-history features were removed, and the main conclusion remained stable across multiple forecasting horizons and rolling-origin evaluations. Peak-demand and external-validation analyses further tested the robustness of the findings beyond a single split.

Conclusion: The study demonstrates that temporal feature representation and recent consumption history can be at least as important as increasing model complexity in short-horizon appliance-energy forecasting. By combining leakage-aware validation, explainability, and robustness analysis, the framework provides transparent evidence for energy-monitoring and demand-planning decisions. Sustainability relevance is operationalized through forecasting accuracy and peak-demand reliability as measurable decision-support outcomes, while direct cost, energy-saving, or emissions reductions are not claimed.