Detecção de outliers e preenchimento de lacunas em medidas das irradiâncias solares espectrais usando princípios físicos e técnicas de inteligência artificial
Carregando...
Arquivos
Data
2022-12-12
Autores
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
A radiação solar de ondas curtas que incide sob a superfície terrestre é composta espectralmente pelas componentes ultravioleta (290 a 400nm), visível (400 a 700nm) e infravermelha (700 a 4000nm). O conhecimento sobre a sua disponibilidade na superfície terrestre permite que órgãos públicos e instituições acadêmicas planejem adequadamente projetos energéticos no Brasil. Para isso, se torna necessário a implantação de sistemas terrestres de medição, que compreendem, naturalmente, valores errôneos causados por adversidades técnicas e operacionais. Neste sentido, o objetivo deste estudo foi implementar um procedimento de controle de qualidade, para identificar valores errôneos, e outro para preencher lacunas das irradiâncias solares ultravioleta, visível e infravermelha sub-horárias (W/m²), coletadas em Botucatu (SP) - Brasil no período de 2001 a 2006. O procedimento de controle de qualidade foi criado em função de três princípios físicos e um intervalo de confiança estatística. Em sequência, o procedimento de preenchimento de lacunas foi implementado em função de um modelo de Machine Learning e avaliado a partir do coeficiente de determinação (R²), erro médio absoluto (MAE e rMAE), e da raiz do erro quadrático médio (RMSE e rRMSE). Foi possível detectar a partir do procedimento de controle de qualidade 1,003%, 0,989% e 0,806% de medições errôneas e isoladas das irradiâncias solares ultravioleta, infravermelha e visível, respectivamente, com relação aos totais medidos. O modelo de preenchimento de lacunas obteve métricas, com relação a sua capacidade de estimar valores, de R² de 0,99565, rMAE de 2,849% e rRMSE de 4,821% para a irradiância solar infravermelha, bem como R² de 0,99562, rMAE de 2,886% e rRMSE de 4,875% para a visível e, por fim, R² de 0,99137, rMAE de 3,923% e rRMSE de 6,889% para a ultravioleta. A partir deste estudo, foi possível avaliar que o primeiro e o terceiro princípio físico foram o mais e o menos restritivos do controle de qualidade, respectivamente, ao detectar medidas errôneas e isoladas, consideradas outliers por se afastarem consideravelmente das demais. Constatou-se, também, que os modelos de Machine Learning obtiveram resultados semelhantes as modelagens de Escobedo et al. (2011) para resoluções temporais horárias e diárias, o que possibilita utilizá-los para preencher lacunas das irradiâncias solares espectrais.
The shortwave solar radiation that falls on the Earth's surface is spectrally composed of ultraviolet (290 to 400nm), photosynthetically active (400 to 700nm) and near infrared (700 to 4000nm) components. The knowledge about its availability on the Earth's surface allows public agencies and academic institutions to plan energy projects in Brazil. For this, it is necessary to implement terrestrial measurement systems, which naturally include erroneous values caused by technical and operational adversities. In this sense, the aim of this study was to implement a quality control procedure, to identify erroneous values, and to fill gaps in sub hourly ultraviolet, photosynthetically active, and near-infrared solar irradiance (Wm-2 ), collected in Botucatu (SP) - Brazil from 2001 to 2006. The quality control procedure was created based on three physical principles and a statistical confidence interval. The gap-filling procedure was implemented based on a Machine Learning model and evaluated based on the coefficient of determination (R2), mean absolute error (MAE and rMAE), and the root mean squared error (RMSE and rRMSE). It was possible to detect from the quality control procedure 1.003%, 0.989% and 0.806% of erroneous and isolated measurements of ultraviolet, near infrared and photosynthetically active solar irradiances, respectively, in relation to the total measured. The gap-filling model obtained metrics, regarding its ability to estimate values, of R2 of 0.99565, rMAE of 2.849% and rRMSE of 4.821% for near infrared, as well as R2 of 0.99562, rMAE of 2.886% and rRMSE of 4.875% for the photosynthetically active and R2 of 0.99137, rMAE of 3.923% and rRMSE of 6.889% for the ultraviolet solar irradiances. From this study it was possible to assess that the first physical principle was the most restrictive, and the third physical principle the least restrictive, of quality control, when detecting erroneous measurements. All measurements flagged as erroneous or isolated are considered outliers because they are considerably different from the others. Finally, it was found that the gap-filling models obtained results similar to the models by Escobedo et al. (2011) for hourly and daily temporal resolutions, which makes it possible to use them to estimate measurements of spectral solar irradiance.
The shortwave solar radiation that falls on the Earth's surface is spectrally composed of ultraviolet (290 to 400nm), photosynthetically active (400 to 700nm) and near infrared (700 to 4000nm) components. The knowledge about its availability on the Earth's surface allows public agencies and academic institutions to plan energy projects in Brazil. For this, it is necessary to implement terrestrial measurement systems, which naturally include erroneous values caused by technical and operational adversities. In this sense, the aim of this study was to implement a quality control procedure, to identify erroneous values, and to fill gaps in sub hourly ultraviolet, photosynthetically active, and near-infrared solar irradiance (Wm-2 ), collected in Botucatu (SP) - Brazil from 2001 to 2006. The quality control procedure was created based on three physical principles and a statistical confidence interval. The gap-filling procedure was implemented based on a Machine Learning model and evaluated based on the coefficient of determination (R2), mean absolute error (MAE and rMAE), and the root mean squared error (RMSE and rRMSE). It was possible to detect from the quality control procedure 1.003%, 0.989% and 0.806% of erroneous and isolated measurements of ultraviolet, near infrared and photosynthetically active solar irradiances, respectively, in relation to the total measured. The gap-filling model obtained metrics, regarding its ability to estimate values, of R2 of 0.99565, rMAE of 2.849% and rRMSE of 4.821% for near infrared, as well as R2 of 0.99562, rMAE of 2.886% and rRMSE of 4.875% for the photosynthetically active and R2 of 0.99137, rMAE of 3.923% and rRMSE of 6.889% for the ultraviolet solar irradiances. From this study it was possible to assess that the first physical principle was the most restrictive, and the third physical principle the least restrictive, of quality control, when detecting erroneous measurements. All measurements flagged as erroneous or isolated are considered outliers because they are considerably different from the others. Finally, it was found that the gap-filling models obtained results similar to the models by Escobedo et al. (2011) for hourly and daily temporal resolutions, which makes it possible to use them to estimate measurements of spectral solar irradiance.
Descrição
Palavras-chave
Radiação solar espectral, Detecção de medições errôneas, Controle de qualidade, Modelagem, Spectral solar radiation, Detection of erroneous measurements, Quality control, Machine learning, Modeling