Abstract
This study analyzes the energy demand of the department of Huila between
2023 and 2025 using statistical and machine learning models, with the aim of
evaluating their predictive capacity and providing evidence for
understanding the recent dynamics of regional electricity consumption. The
database used includes daily and hourly electricity consumption series,
which allowed the examination of seasonal patterns, growth trends, and
demand variations associated with different temporal behaviors. The analysis
implemented ARIMA, SARIMAX, linear regression, Support Vector Regression
(SVR), Gaussian Process Regression (GPR), and Random Forest models, and
their performance was evaluated using error metrics such as MAE, RMSE, MAPE,
MASE, and R². The results show that traditional statistical models,
particularly ARIMA and SARIMAX, exhibited more limited performance in
dealing with the complexity and nonlinearity of the analyzed series. In
contrast, machine learning models achieved better levels of fit, with Random
Forest standing out, obtaining an RMSE of 14.137 units and an R² of 0.67,
positioning it as the best-performing model among those evaluated for hourly
forecasting. In this regard, the findings highlight the potential of machine
learning approaches to strengthen regional-scale energy demand forecasting
exercises.