Evaluación de modelos para estimar insolación diaria en función de temperaturas máxima, mínima y precipitación del día
Loading...
Date
2008
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
En el presente estudio se probaron los sistemas ANFIS (Adaptive Network Based Fuzzy Inference System), para predecir la insolación diaria en Chillán, usando como variables explicativas de esta, las temperaturas máximas y mínimas del aire y la precipitación diaria, ya que estos sistemas pueden modelar fenómenos físicos no-lineales usando datos históricos. Estos sistemas fueron comparados con los modelos de Bristow & Campbell y De Jong & Stewart. Se crearon 32 modelos en total, y se seleccionaron los mejores de cada una de las tres familias de modelos (ANFIS, Bristow & Campbell y De Jong & Stewart), quedando 6 modelos para un análisis detallado, y se encontró que ellos entregan predicciones con un error medio absoluto (MAE) aproximado de 2,7-2 -1 MJ m día como promedio anual. Además, las predicciones de verano y primavera tienen un error porcentual absoluto (MAPE) cercano al 15%, mientras que para otoño e invierno éste es de un 50%, para los 6 modelos seleccionados. El modelo que presentó el mejor desempeño fue el modelo ANFIS que usa como variables explicativas la amplitud térmica del día y el día Juliano, con un radio de influencia en la agrupación de 0,75 (modelo C.6), el cual entrega predicciones con un MAE de 2,6 -2 -1 MJ m día como promedio anual. Finalmente, se crearon y evaluaron dos modelos de Ángstrom (que usan como variable explicativa el número de horas diarias), el recomendado por la FAO (para el cálculo de evapotranspiración) y uno para las condiciones específicas de Chillán. Se encontró que este último modelo tiene mejor desempeño, a nivel anual y estacional, que el modelo C.6, ya que entrega estimaciones de insolación con un MAE de 1,6 -2 -1 MJ m día como promedio anual
In this current study ANFIS systems (Adaptive Network Based Fuzzy Inference System) were proved, in order to predict the daily insolation in Chillán (Chile), using air maximum and minimum temperatures and daily rainfall as explanatory variables of them, since these systems can model physical non-lineal phenomena using historical data. These systems were compared with the models proposed by Bristow & Campbell, and De Jong & Stewart. Thirty two models were created and the best of each of three families of models were selected (ANFIS, Bristow & Campbell, and De Jong & Stewart), remaining 6 models for a detailed analysis, observing that they provide predictions with an approximate mean absolute error (MAE) of 2,7 -2 -1 MJm day as an annual average. Moreover, summer and spring predictions have a mean absolute percentage error (MAPE) close to 15%, while for autumn and winter this is close to 50% for the 6 selected models. The model that presented the best performance was the ANFIS model, which uses day thermal amplitude and the Julian Day as explanatory variables, with an influence radius in the cluster of 0,75 (Model C.6), which provides predictions with a MAE of 2,6-2 -1 MJm day as an annual average. Finally, two Ángstrom models (using the number of daily hours as explanatory variable) were created and evaluated, the first recommended by FAO (to calculate evapotranspiration) and the second for specific conditions in Chillán. The latter model presents a better performance, at the seasonal and annual level, than model C.6, since it provides insolation estimates with a MAE of 1,6 -2 -1 MJm day as an annual average.
In this current study ANFIS systems (Adaptive Network Based Fuzzy Inference System) were proved, in order to predict the daily insolation in Chillán (Chile), using air maximum and minimum temperatures and daily rainfall as explanatory variables of them, since these systems can model physical non-lineal phenomena using historical data. These systems were compared with the models proposed by Bristow & Campbell, and De Jong & Stewart. Thirty two models were created and the best of each of three families of models were selected (ANFIS, Bristow & Campbell, and De Jong & Stewart), remaining 6 models for a detailed analysis, observing that they provide predictions with an approximate mean absolute error (MAE) of 2,7 -2 -1 MJm day as an annual average. Moreover, summer and spring predictions have a mean absolute percentage error (MAPE) close to 15%, while for autumn and winter this is close to 50% for the 6 selected models. The model that presented the best performance was the ANFIS model, which uses day thermal amplitude and the Julian Day as explanatory variables, with an influence radius in the cluster of 0,75 (Model C.6), which provides predictions with a MAE of 2,6-2 -1 MJm day as an annual average. Finally, two Ángstrom models (using the number of daily hours as explanatory variable) were created and evaluated, the first recommended by FAO (to calculate evapotranspiration) and the second for specific conditions in Chillán. The latter model presents a better performance, at the seasonal and annual level, than model C.6, since it provides insolation estimates with a MAE of 1,6 -2 -1 MJm day as an annual average.
Description
Tesis presentada para optar al título de Ingeniero Civil Agrícola
Keywords
Lógica difusa, Conjuntos difusos, Redes neurales (ciencia de la computación)