Modelo de Demanda Eléctrica Mediante Ciencia de Datos y Machine Learning: Caso Sistema Eléctrico del Ecuador

Electricity Demand Model Using Data Science and Machine Learning: Case Study of the Ecuadorian Electricity System

Abstract

In this study, machine learning and statistical models were compared to examine their performance on multi-source data from 2001 to 2025. This data includes climatic, socioeconomic, and available energy values ​​for Ecuador, with the latter being the target variable. The SARIMAX, XGBoost, and LightGBM models were evaluated in a temporal validation comparison. The results show SARIMAX outperforming the others (R²=0.987; MAPE<1%), demonstrating autocorrelation, seasonality, and trends. This shows that machine learning models, excluding hybrid models, do not follow drastic changes, resulting in low levels of accuracy. The collinearity between billed and available energy is also evident, consistent with the system's structural constraints, with time dynamics being the predominant factor in the model's complexities.