Comparative Framework For Electric Demand Forecasting With Machine Learning and Rolling Temporal Validation

Hourly load forecasting at substation level is a key input for planning and operating distribution networks, especially under increas- ing variability driven by electrification and renewable integration. This contribution develops a comparative framework between two representa- tive approaches for tabular data: an XGBoost model with evolutionary hyperparameter tuning (EvoXGB) and the deep tabular network Tab- Net. Both models are evaluated on real active-power time series from an anonymized distribution substation of an Ecuadorian power system (Substation A). The evaluation follows a rolling-origin expanding vali- dation scheme with short- and medium-term horizons, complemented by an independent 90/10 hold-out test. The feature set includes hourly lags, lagged moving averages and calendar variables. Standard statistical met- rics (MAE, RMSE, R2) are combined with an absolute-tolerance metric %Tol, defined as the fraction of predictions whose absolute error remains within a threshold τ in kW. Results show that, under a realistic tempo- ral evaluation protocol, EvoXGB reduces RMSE by about 20–30% with respect to TabNet in both rolling backtesting and in the hold-out pe- riod, while increasing the proportion of forecasts with error below 10 kW by 6–9 percentage points. These gains are obtained with a significantly lower training cost. The proposed framework illustrates that a carefully calibrated gradient-boosting model remains a very competitive option compared to deep tabular networks for substation-level load forecasting.