Evaluación del desempeño de modelos LSTM y XGBoost en la predicción de la demanda eléctrica del sistema ecuatoriano

Accurate short-term electricity demand forecasting is essential for the technical and economic operation of the Ecuadorian power system. This paper presents a comparison between Long Short-Term Memory (LSTM) neural networks and the XGBoost algorithm for short-term load forecasting, incorporating exogenous variables such as apparent temperature and national holidays. Hourly demand data were obtained from the CENACE database starting in 2021, and meteorological data were sourced from the Open-Meteo satellite platform. A recursive single-step forecasting strategy was implemented for a 24-hour prediction horizon. Results show that the LSTM model achieved the highest accuracy, significantly outperforming XGBoost. The study concludes that incorporating exogenous variables improves forecasting performance and that LSTM provides a reliable approach for short-term load prediction to support national power system planning.