Publication: DISEÑO DE UN ALGORITMO DE EVALUACIOÓN DEL RENDIMIENTO DE MOÓDULOS FOTOVOLTAICOS EN OPERACIOÓN BASADO EN ANÁLISIS DE IMÁGENES
Loading...
Date
2021-01
Authors
CAVIERES ABARCA, ROBINSON DANIEL
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
La energía solar fotovoltaica es una herramienta que está encontrando su lugar en el mercado, pero aún es necesario mejorar las técnicas de monitoreo para lograr un salto definitivo a ser una alternativa comercialmente atractiva para los inversionistas.
El objetivo de esta investigación es desarrollar un algoritmo de monitoreo de paneles solares basado en análisis de imágenes, capaz de evaluar el rendimiento de múltiples módulos en operación simultáneamente basándose en una imagen RGB capturada en el espectro visible . El modelo diseñado debe ser simple, eficaz y capaz de aplicarse en múltiples condiciones.
Para lograr este objetivo, se desarrollan dos enfoques principales, uno basado en arquitectura de redes neuronales artificiales, capaz de procesar características extraídas manualmente de las imágenes y usar la información para predecir el rendimiento de los módulos fotovoltaicos como un valor de carácter continuo. El algoritmo desarrollado muestra resultados prometedores para ensuciamiento de acumulación natural, con un R2 de 0; 96 y un error cuadrático medio de 4; 87 . 10-5 para el conjunto de validación, pero no es efectivo para sombreamiento parcial ni ensuciamiento de acumulación artificial.
El segundo enfoque se basa en arquitectura de redes neuronales convoluciones, capaz de identificar, segmentar y analizar cada módulo fotovoltaico presente en una imagen y evaluar el rendimiento individualmente, como un valor categ´orico. Este nuevo enfoque es capaz de pronosticar correctamente el desempeñoo de módulos sujetos a ensuciamiento natural, artificial y sombreamiento parcial, alcanzando una precisión de un 73% sobre 8 categorías, y concentrando la mayor parte del error en categorías aledañas, logrando así una herramienta sólida para el diagnóstico de módulos fotovoltaicos
Photovoltaic solar power is a technology that is finding its place in global markets, but it is still necessary to improve monitoring approaches to achieve a definitive leap to being a commercially attractive alternative for investors. The goal of this research is to develop an algorithm for monitoring solar panels based on image analysis, capable of evaluating the performance of multiple modules in operation simultaneously. The designed model must be simple, effective and capable of being applied in multiple conditions. To achieve this goal, two main approaches are developed, one based on artificial neural network architecture, capable of processing characteristics extracted manually from the images and using the information to predict the performance of photovoltaic modules as a continuous value. The developed algorithm shows promising results for natural soiling, with an R2 of 0;96 and a mean square error of 4;87 .10-5 for the test set, but it is not effective for partial shading or artificial soiling. The second approach is based on convolutional neural network architecture, capable of identifying, segmenting and analyzing each photovoltaic module present in an image and evaluating the performance individually, as a categorical value. This new approach is capable of correctly predicting performance of modules subject to natural and artificial soiling and partial shading, scoring an accuracy of 73% over 8 categories, and concentrating most of the error in neighboring categories, thus achieving a solid tool for PV module diagnostics
Photovoltaic solar power is a technology that is finding its place in global markets, but it is still necessary to improve monitoring approaches to achieve a definitive leap to being a commercially attractive alternative for investors. The goal of this research is to develop an algorithm for monitoring solar panels based on image analysis, capable of evaluating the performance of multiple modules in operation simultaneously. The designed model must be simple, effective and capable of being applied in multiple conditions. To achieve this goal, two main approaches are developed, one based on artificial neural network architecture, capable of processing characteristics extracted manually from the images and using the information to predict the performance of photovoltaic modules as a continuous value. The developed algorithm shows promising results for natural soiling, with an R2 of 0;96 and a mean square error of 4;87 .10-5 for the test set, but it is not effective for partial shading or artificial soiling. The second approach is based on convolutional neural network architecture, capable of identifying, segmenting and analyzing each photovoltaic module present in an image and evaluating the performance individually, as a categorical value. This new approach is capable of correctly predicting performance of modules subject to natural and artificial soiling and partial shading, scoring an accuracy of 73% over 8 categories, and concentrating most of the error in neighboring categories, thus achieving a solid tool for PV module diagnostics
Description
Keywords
RGB , REDES NEURALES (Ciencia de la Computación) , EFICIENCIA ENERGETICA , FOTOVOLTAICO