Evaluación de modelos de pronóstico de la demanda horaria de energía eléctrica en Colombia
Cargando...
Autores
Tipo de contenido
Editor
Document language:
Español
Fecha
Título de la revista
ISSN de la revista
Título del volumen
Documentos PDF
Resumen
The temporal characterization of electricity demand is a critical input for the operation and planning of the National Interconnected System (SIN). To identify predictive techniques capable of representing this behavior, several models are analyzed using hourly consumption series recorded between January 2022 and August 2025. The dataset incorporates day-type categorizations (weekday, Saturday, Sunday, and non-Sunday holiday) to capture operational patterns relevant to the series.
Four modeling approaches were implemented: a seasonal stochastic model (SARIMAX), a tree-based method (Random Forest), and two neural architectures—an LSTM recurrent network and a feed-forward network. All models were trained and validated under a uniform methodological protocol, enabling a consistent assessment of each approach’s ability to reproduce short-term variations and capture the fundamental structural patterns of hourly demand behavior.
The results indicate that, although SARIMAX provides an adequate representation of the series’ seasonal component and linear structure, the nonlinear models exhibit superior performance. In particular, the LSTM network achieved the lowest error magnitude, demonstrating its effectiveness in modeling complex temporal dependencies and its suitability for hourly forecasting applications in power systems with high operational variability. (Texto tomado de la fuente)
Abstract
The temporal characterization of electricity demand is a critical input for the operation and planning of the National Interconnected System (SIN). To identify predictive techniques capable of representing this behavior, several models are analyzed using hourly consumption series recorded between January 2022 and August 2025. The dataset incorporates day-type categorizations (weekday, Saturday, Sunday, and non-Sunday holiday) to capture operational patterns relevant to the series. Four modeling approaches were implemented: a seasonal stochastic model (SARIMAX), a treebased method (Random Forest), and two neural architectures—an LSTM recurrent network and a Feed Forward network. All models were trained and validated under a uniform methodological protocol, enabling a consistent assessment of each approach’s ability to reproduce short-term variations and capture the fundamental structural patterns of hourly demand behavior. The results indicate that, although SARIMAX provides an adequate representation of the series’ seasonal component and linear structure, the nonlinear models exhibit superior performance. In particular, the LSTM network achieved the lowest error magnitude, demonstrating its effectiveness in modeling complex temporal dependencies and its suitability for hourly forecasting applications in power systems with high operational variability.
Descripción
Palabras clave propuestas
Términos Normalizados
Aprendizaje automático (Inteligencia arifical); Demanda de energía eléctrica; Procesos estocasticos; 000 - Ciencias de la computación, información y obras generales::003 - Sistemas; 000 - Ciencias de la computación, información y obras generales::004 - Procesamiento de datos Ciencia de los computadores; 620 - Ingeniería y operaciones afines::621 - Física aplicada; 330 - Economía::333 - Economía de la tierra y de la energía