Marco comparativo para el pronóstico de demanda eléctrica con machine learning y validación temporal rodante

Accurate load forecasting is essential for power system planning and operation, particularly under pronounced temporal variability and temporal drift. This study presents a reproducible comparative framework for machine learning models based on rolling-origin expanding validation, multihorizon evaluation, and an operational relative tolerance metric denoted as Tol. Four representative models are evaluated: EvoXGB, a sequential residual XGBoost ensemble; XGB; TabNet; and FT-Transformer. These models are applied to hourly active power forecasting in distribution substations within an Ecuadorian power system. To ensure a fair comparison when models exhibit differences in prediction coverage or temporal misalignment, the framework incorporates an explicit comparability audit based on temporal alignment and a common evaluation mask denoted as COMMONMASK, complemented the longest common contiguous block for the zoomed time-series visualization. For the representative substation, with metrics computed on the common set, XGB achieves the best performance, with R2= 0.993 for the short horizon and R2= 0.983 for the medium horizon, and RMSE values of 21.16 and 30.84 kW, respectively. EvoXGB remains competitive, whereas TabNet and FT-Transformer exhibit greater degradation in the medium horizon. The 90/10 holdout verification shows the expected performance decline associated with temporal drift while preserving the comparative ranking. Overall, the proposed framework provides a traceable benchmark for substation load forecasting and supports future extensions toward adaptive and hybrid forecasting approaches.