Modelagem em superfícies inclinadas das radiações global e difusa usando técnicas de aprendizado de máquina
Data
2018-05-30
Autores
Marques, Adriano de Souza [UNESP]
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
Neste trabalho é realizado um estudo para estimar a transmissividade da radiação global (Ktβh) e a fração difusa (Kdβh) incidentes em uma superfície com inclinação de 22,85° na base horária utilizando técnicas de aprendizado de máquina (TAM), a partir de dados obtidos no período de 1998 a 2001 em Botucatu/SP/Brasil. As estimativas foram realizadas usando uma série de combinações de variáveis astronômicas e geográficas por meio de três técnicas de redes neurais artificiais (RNA) do tipo Perceptron Multicamadas (MLP), Função de Base Radial (RBF) e Regressão Generalizada (GRNN) e do Sistema Adaptativo de Inferência Neuro Fuzzy (ANFIS). Como referência foram elaborados modelos estatísticos (ME) de regressão linear e polinomial. No Capítulo 1 as estimativas de (Ktβh) foram realizadas por combinações de variáveis medidas e calculadas a partir da irradiação global na superfície horizontal (HgH) e nas estimativas de (Kdβh) utilizou-se combinações de variáveis medidas e calculadas a partir de (HgH) e da irradiação global na superfície inclinada (Hgβ). No Capítulo 2 as estimativas de (Kdβh) foram realizadas por combinações de variáveis medidas e calculadas a partir das irradiações difusa (HdH) e global (HgH) obtidas na superfície horizontal. Os indicadores estatísticos r (correlação), RMSE(%) (precisão) e MBE(%) (exatidão) foram utilizados para avaliar os resultados das estimativas. No capítulo 1 os melhores resultados nas estimativas de (Ktβh) a partir das combinações realizadas com (HgH) foram: MLP - RMSE=3,73%; RBF - RMSE=3,99%; GRNN - RMSE=5,27%; ANFIS - RMSE=3,78% e ME - RMSE=6,65%. Nesse caso os indicadores de precisão mostram uma redução de aproximadamente 44% com o uso da técnica (MLP) em comparação ao modelo estatístico (ME). Nas estimativas de (Kdβh) a partir das combinações de (HgH) os melhores resultados foram: MLP - RMSE=21,69%; RBF - RMSE=25,43%; GRNN - RMSE=29,39%; ANFIS - RMSE=23,08% e - ME - RMSE=35,35%. Da mesma forma os indicadores de precisão mostram uma redução de aproximadamente 39% com o uso da técnica (MLP) em comparação ao modelo estatístico (ME). E nas estimativas de (Kdβh) a partir das combinações realizadas com (Hgβ) os melhores resultados foram: MLP - RMSE=20,32%; RBF - RMSE=21,95%; GRNN - RMSE=29,11%; ANFIS - RMSE=21,75% e ME - RMSE=36,48%. Os indicadores de precisão mostram uma redução de aproximadamente 44% com o uso da técnica (MLP) em comparação ao modelo estatístico (ME). No capítulo 2 as melhores estimativas de (Kdβh) a partir das combinações realizadas com (HdH) foram: MLP - RMSE=4,03%; RBF - RMSE=5,84%; GRNN - RMSE=10,85%; ANFIS - RMSE=4,15% e ME - RMSE=12,42%. Os indicadores de precisão mostram uma redução de aproximadamente 67% com o uso da técnica (MLP) em comparação ao modelo estatístico (ME). Nas estimativas de (Kdβh) a partir de (HgH) os melhores resultados foram: MLP - RMSE=21,69%; RBF - RMSE=25,43%; GRNN - RMSE=29,39%; ANFIS - RMSE=23,08% e ME - RMSE=35,35%. Os indicadores de precisão mostram uma redução de aproximadamente 39% com o uso da técnica (MLP) em comparação ao modelo estatístico (ME). Os resultados mostram que a técnica de rede neural artificial MLP apresentou os melhores índices em todas as estimativas de (Ktβh) e (Kdβh) com reduções significativas quando comparadas aos resultados obtidos com as estimativas obtidas com os modelos estatísticos. Pela análise dos resultados é possível observar que o uso das técnicas de aprendizado de máquina (TAM) nas combinações de variáveis propostas e com os dados obtidos de Botucatu/SP, se apresentam como alternativa aos modelos estatísticos (ME) para estimar as variáveis de (Ktβh) e (Kdβh).
In this work, a study was carried out to estimate the transmissivity of the global radiation (Ktβh) and the diffuse fraction (Kdβh) incident on a surface with slope of 22.85 ° in the hourly basis using machine learning techniques (MLT), from data obtained from 1998 to 2001 in Botucatu / SP / Brazil. The estimates were made using a series of combinations of astronomical and geographic variables by means of three artificial neural network (ANN) techniques such as MultLayer Perceptron (MLP), Radial Basis Functions Networks (RBF) and Generalized Regression Neural Network (GRNN) Adaptive Neuro Fuzzy Inference System (ANFIS). Statistical models (SM) of linear and polynomial regression were elaborated as reference. In Chapter 1 estimates of (Ktβh) were performed by combinations of variables measured and calculated from global horizontal surface irradiation (HgH) and estimates of (Kdβh) combinations of variables measured and calculated from (HgH) and global radiation on the sloped surface (Hgβ). In Chapter 2 estimates of (Kdβh) were performed by combinations of variables measured and calculated from the diffuse (HdH) and global (HgH) irradiances obtained on the horizontal surface. The statistical indicators r (correlation), RMSE (%) (precision) and MBE (%) (accuracy) were used to evaluate the results of the estimates. In Chapter 1 the best results in the estimates of (Ktβh) from the combinations performed with (HgH) were: MLP - RMSE = 3.73%; RBF - RMSE = 3.99%; GRNN - RMSE = 5.27%; ANFIS-RMSE = 3.78% and SM - RMSE = 6.65%. In this case the precision indicators show a reduction of approximately 44% with the use of the technique (MLP) in comparison to the statistical model (SM). In the estimates of (Kdβh) from the combinations of (HgH) the best results were: MLP - RMSE = 21.69%; RBF - RMSE = 25.43%; GRNN - RMSE = 29.39%; ANFIS - RMSE = 23.08% and SM - RMSE = 35.35%. Likewise, the precision indicators show a reduction of approximately 39% with the use of the technique (MLP) in comparison to the statistical model (SM). And in the estimates of (Kdβh) from the combinations performed with (Hgβ) the best results were: MLP - RMSE = 20.32%; RBF - RMSE = 21.95%; GRNN - RMSE = 29.11%; ANFIS - RMSE = 21.75% and SM - RMSE = 36.48%. The precision indicators show a reduction of approximately 44% with the use of the technique (MLP) in comparison to the statistical model (SM). In Chapter 2 the best estimates of (Kdβh) from the combinations performed with (HdH) were: MLP - RMSE = 4.03%; RBF - RMSE = 5.84%; GRNN - RMSE = 10.85%; ANFIS - RMSE = 4.15% and SM - RMSE = 12.42%. The precision indicators show a reduction of approximately 67% with the use of the technique (MLP) in comparison to the statistical model (SM). In the estimates of (Kdβh) from (HgH) the best results were: MLP - RMSE = 21.69%; RBF - RMSE = 25.43%; GRNN - RMSE = 29.39%; ANFIS - RMSE = 23.08% and SM - RMSE = 35.35%. The precision indicators show a reduction of approximately 39% with the use of the technique (MLP) in comparison to the statistical model (SM). The results show that the artificial neural network MLP technique presented the best indexes in all estimates of (Ktβh) and (Kdβh) with significant reductions when compared to the results obtained with the estimates obtained with the statistical models. By the analysis of the results it is possible to observe that the use of the machine learning techniques (MLT) in the combinations of proposed variables and the data obtained from Botucatu / SP, are presented as an alternative to the statistical models (SM) to estimate the variables of (Ktβh) and (Kdβh).
In this work, a study was carried out to estimate the transmissivity of the global radiation (Ktβh) and the diffuse fraction (Kdβh) incident on a surface with slope of 22.85 ° in the hourly basis using machine learning techniques (MLT), from data obtained from 1998 to 2001 in Botucatu / SP / Brazil. The estimates were made using a series of combinations of astronomical and geographic variables by means of three artificial neural network (ANN) techniques such as MultLayer Perceptron (MLP), Radial Basis Functions Networks (RBF) and Generalized Regression Neural Network (GRNN) Adaptive Neuro Fuzzy Inference System (ANFIS). Statistical models (SM) of linear and polynomial regression were elaborated as reference. In Chapter 1 estimates of (Ktβh) were performed by combinations of variables measured and calculated from global horizontal surface irradiation (HgH) and estimates of (Kdβh) combinations of variables measured and calculated from (HgH) and global radiation on the sloped surface (Hgβ). In Chapter 2 estimates of (Kdβh) were performed by combinations of variables measured and calculated from the diffuse (HdH) and global (HgH) irradiances obtained on the horizontal surface. The statistical indicators r (correlation), RMSE (%) (precision) and MBE (%) (accuracy) were used to evaluate the results of the estimates. In Chapter 1 the best results in the estimates of (Ktβh) from the combinations performed with (HgH) were: MLP - RMSE = 3.73%; RBF - RMSE = 3.99%; GRNN - RMSE = 5.27%; ANFIS-RMSE = 3.78% and SM - RMSE = 6.65%. In this case the precision indicators show a reduction of approximately 44% with the use of the technique (MLP) in comparison to the statistical model (SM). In the estimates of (Kdβh) from the combinations of (HgH) the best results were: MLP - RMSE = 21.69%; RBF - RMSE = 25.43%; GRNN - RMSE = 29.39%; ANFIS - RMSE = 23.08% and SM - RMSE = 35.35%. Likewise, the precision indicators show a reduction of approximately 39% with the use of the technique (MLP) in comparison to the statistical model (SM). And in the estimates of (Kdβh) from the combinations performed with (Hgβ) the best results were: MLP - RMSE = 20.32%; RBF - RMSE = 21.95%; GRNN - RMSE = 29.11%; ANFIS - RMSE = 21.75% and SM - RMSE = 36.48%. The precision indicators show a reduction of approximately 44% with the use of the technique (MLP) in comparison to the statistical model (SM). In Chapter 2 the best estimates of (Kdβh) from the combinations performed with (HdH) were: MLP - RMSE = 4.03%; RBF - RMSE = 5.84%; GRNN - RMSE = 10.85%; ANFIS - RMSE = 4.15% and SM - RMSE = 12.42%. The precision indicators show a reduction of approximately 67% with the use of the technique (MLP) in comparison to the statistical model (SM). In the estimates of (Kdβh) from (HgH) the best results were: MLP - RMSE = 21.69%; RBF - RMSE = 25.43%; GRNN - RMSE = 29.39%; ANFIS - RMSE = 23.08% and SM - RMSE = 35.35%. The precision indicators show a reduction of approximately 39% with the use of the technique (MLP) in comparison to the statistical model (SM). The results show that the artificial neural network MLP technique presented the best indexes in all estimates of (Ktβh) and (Kdβh) with significant reductions when compared to the results obtained with the estimates obtained with the statistical models. By the analysis of the results it is possible to observe that the use of the machine learning techniques (MLT) in the combinations of proposed variables and the data obtained from Botucatu / SP, are presented as an alternative to the statistical models (SM) to estimate the variables of (Ktβh) and (Kdβh).
Descrição
Palavras-chave
Irradiação solar, Redes Neurais Artificiais, ANFIS, Estimativas, Solar irradiation, Artificial neural networks, Estimates