Publication: PRONÓSTICO DE DEMANDA ELÉCTRICA UNIVARIADA A CORTO PLAZO MEDIANTE APROXIMACIONES ESTADÍSTICAS Y DE INTELIGENCIA ARTIFICIAL. CASO APLICADO A OPERADOR DE SISTEMA DE TRANSMISIÓN FRANCÉS
Date
2018
Authors
ABALLAY LEIVA, BASTIÁN ALEXIS
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
El pronóstico de carga a corto plazo (STLF, por sus siglas en inglés) juega un papel
fundamental en la planificación y operación eficiente de los sistemas de energía. Los
pronósticos a corto plazo precisos ayudan con las decisiones sobre programación de
unidades, transferencia energética, planes de mantenimiento y respuesta a la demanda.
Diversos modelos han sido desarrollados para obtener pronósticos precisos, sin embargo,
pocos se encuentran disponibles gratuitamente para cualquier practicante. En el presente
trabajo se comparan enfoques estadísticos y de inteligencia artificial para el pronóstico de
la demanda eléctrica cuyo horizonte es un día en adelante. Para este fin se utilizan paquetes
de software gratuito que facilitan el modelamiento de series de tiempo y la especificación
de modelos estadísticos así como también no-lineales. El análisis comparativo se enfoca
en técnicas de pronóstico univariado que pueden establecerse como punto de referencia
para modelos más complejos. Para proporcionar un análisis integrado, se realiza Análisis
de Datos Exploratorio (EDA) y las visualizaciones necesarias para comprender los datos
son entregadas. Los métodos son revisados y comparados por tipo de técnica utilizando
la base de datos proveída de manera libre por el sistema de transmisión francés RTE. La
codificación estacional determinística para la serie de carga se compara con el enfoque de
diferenciación estacional. Se considera la función de autocorrelación y los procedimientos
de preprocesamiento de información mutua para llevar a cabo la selección de variables
a utilizar en los modelos de inteligencia artificial. En los experimentos numéricos, el
promedio de la media de error absoluta de los mejores modelos por técnica revisada fue
inferior al 3 %. El modelo Holt Winters con Estacionalidad Doble supera a todos los
modelos considerando un año entero como período de prueba. Los modelos de inteligencia
artificial logran mayor precisión cuando la doble diferenciación estacional es utilizada en las
etapas de preprocesamiento. Este estudio puede ser de utilidad tanto para los operadores del
sistema, así como también para los practicantes que buscan una introducción al problema
de STLF, centrándose en modelos disponibles al alcance de la mano.
Short-term load forecasting (STLF) plays a fundamental role in the e cient planning and operation of power systems. Accurate short-term forecasts help with decisions regarding to unit commitment, economic dispatch, maintenance plans and demand response. Several models have been developed to obtain accurate forecasts, however, few of them are freely available to any practitioner. In this work, statistical and artificial intelligence approaches for one day-ahead electricity demand forecasting are compared. To this end, we use free software environment packages that facilitate time series modelling and non-linear model specification. We focus our comparative analysis to univariate forecast techniques that can be established as benchmark for more complex models. To provide an integrated analysis, Exploratory Data Analysis (EDA) is performed and the necessary visualizations to understand the data are provided. All methods are reviewed and compared among each technique using the RTE French database. Deterministic seasonal encoding for the load series is compared to the seasonal di erencing approach. Autocorrelation function and mutual information preprocessing procedures are considered to perform feature selection of the artificial intelligence input variables. In the numerical experiments, the average mean absolute percent errors of the best models per technique reviewed were less than 3 %. Double seasonal Holt Winters outperforms all models considering one year as test period. Artificial intelligence models were more accurate when double seasonal di erencing was used in the preprocessing stages. This study should be useful to system operators as well as practitioners looking for an introduction to the STLF problem with focus on models at hand.
Short-term load forecasting (STLF) plays a fundamental role in the e cient planning and operation of power systems. Accurate short-term forecasts help with decisions regarding to unit commitment, economic dispatch, maintenance plans and demand response. Several models have been developed to obtain accurate forecasts, however, few of them are freely available to any practitioner. In this work, statistical and artificial intelligence approaches for one day-ahead electricity demand forecasting are compared. To this end, we use free software environment packages that facilitate time series modelling and non-linear model specification. We focus our comparative analysis to univariate forecast techniques that can be established as benchmark for more complex models. To provide an integrated analysis, Exploratory Data Analysis (EDA) is performed and the necessary visualizations to understand the data are provided. All methods are reviewed and compared among each technique using the RTE French database. Deterministic seasonal encoding for the load series is compared to the seasonal di erencing approach. Autocorrelation function and mutual information preprocessing procedures are considered to perform feature selection of the artificial intelligence input variables. In the numerical experiments, the average mean absolute percent errors of the best models per technique reviewed were less than 3 %. Double seasonal Holt Winters outperforms all models considering one year as test period. Artificial intelligence models were more accurate when double seasonal di erencing was used in the preprocessing stages. This study should be useful to system operators as well as practitioners looking for an introduction to the STLF problem with focus on models at hand.
Description
Keywords
PRONOSTICA DE CARGA A CORTO PLAZO , ESCENARIOS DIARIOS , SELECCION DE VARIABLES , ANALISIS DE DATOS EXPLORATORIO , PRECISION DE PRONOSTICO