Publication: Predicción a corto plazo de radiación solar
Loading...
Official URL
Full text at PDC
Publication Date
2019
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
La predicción de los niveles de irradiación solar a corto plazo es relevante para obtener la energía que se va a producir y hacer el mejor uso. Existen modelos capaces de proporcionar con cierto grado de satisfacción esta predicción. Además de ser precisa, conviene que sea eficiente en recursos, reduciendo los requisitos al máximo posible del aprendizaje máquina. El enfoque de este trabajo es una exploración para hacer modelos más eficientes, intentando mejorar la predicción si es posible haciendo uso de modelos especializados.
Los datos de irradiación con los que se trabaja provienen de la red ubicada en Oahu, Hawai con diecisiete estaciones de medida durante un periodo de diecinueve meses con una frecuencia de un segundo. Este trabajo busca modelos creados con una red neuronal a partir de subconjuntos de todos estos datos mediante la aplicación de distintos métodos de análisis de afinidad entre los valores de irradiación, considerando por un lado la elección manual de los modelos especializados y por otro lado una selección de subconjuntos de datos automática, reduciendo así el impacto en los resultados finales de este trabajo de como son elegidos los modelos.
Los resultados de los modelos especializados elegidos con un conjunto representativo de datos no mejoran la calidad de la predicción pero se pueden tener en cuenta por conseguir resultados similares con menos recursos. Además de la obtención una mayor cantidad de datos correspondiendo a cada modelo, hacer uso de otros criterios para seleccionar modelos especializados serían líneas de posible mejora en la predicción.
Predicting solar irradiance levels in the short term is important to have an accurate idea of the energy to be obtained and make the best use of it. There are models able to achieve this result to a certain extent. Is important to have both an accurate and efficient model, reducing as much as possible its machine learning requirements. The idea of this work is finding more efficient models, even more accurate by defining specialised models. The data used here comes from an Oahu, Hawai based network consisting of seventeen measuring stations registering each one of them the solar irradiance at one second intervals. This work creates models from a neural network after selecting subsets of the data according to the results of applying some affinity analysis on the data. Selecting one where the final definition of the specialised model is made manually based on the aforementioned analysis, and one where the different models are automatically created by the irradiation values, the fact of how the models are chosen has less impact on the final result of the study. Even if the results of the selected specialised models do not improve the quality of the prediction compared to the original generic model, those having a sufficient amount of data are highly recommandable to use because they perform almost as well as the original model with less resources. In addition to increasing the amount of data for each model, studying other criteria to define models, particularly when is manually, might be ways to improve solar prediction.
Predicting solar irradiance levels in the short term is important to have an accurate idea of the energy to be obtained and make the best use of it. There are models able to achieve this result to a certain extent. Is important to have both an accurate and efficient model, reducing as much as possible its machine learning requirements. The idea of this work is finding more efficient models, even more accurate by defining specialised models. The data used here comes from an Oahu, Hawai based network consisting of seventeen measuring stations registering each one of them the solar irradiance at one second intervals. This work creates models from a neural network after selecting subsets of the data according to the results of applying some affinity analysis on the data. Selecting one where the final definition of the specialised model is made manually based on the aforementioned analysis, and one where the different models are automatically created by the irradiation values, the fact of how the models are chosen has less impact on the final result of the study. Even if the results of the selected specialised models do not improve the quality of the prediction compared to the original generic model, those having a sufficient amount of data are highly recommandable to use because they perform almost as well as the original model with less resources. In addition to increasing the amount of data for each model, studying other criteria to define models, particularly when is manually, might be ways to improve solar prediction.
Description
Trabajo de Fin de Máster, Universidad Complutense, Facultad de Informática, Departamento de Arquitectura de Computadores y Automática, Curso 2018/2019