Estimativa da irradiação solar global pelos modelos Hargreaves-Samani, Bristow-Campbell, regressões múltiplas e técnicas MVS e RNA no município de Botucatu - SP
Carregando...
Data
2021-01-29
Autores
Paes, Angélica Castilho
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Estadual Paulista (Unesp)
Resumo
No presente trabalho realizou-se um estudo comparativo entre os modelos de estimativas da irradiação solar global (HG) diária: Hargreaves-Samani (HS); Bristow e Campbell (BC); uma equação de regressão múltipla não linear (RM) desenvolvida neste trabalho e duas técnicas de Aprendizado de Máquina (Máquinas de Vetores de Suporte - MVS e Redes Neurais Artificiais - RNA). A base de dados usada no trabalho foi medida em Botucatu no período dos anos de 1996-2011. Os modelos foram estimados com a base de dados total, sendo que para validação dos modelos foram usados duas bases de dados, denominados ano típico e ano atípico, criados a partir da base total dos dados. O modelo estatístico HS foi estimado pelo método dos mínimos quadrados, utilizado para ajustes lineares, e os modelos BC e RM foram estimados por algoritmos de regressão não linear que utilizam métodos iterativos para a estimativa dos modelos, enquanto que as técnicas, MVS e RNA, foram treinadas com entradas de 4 variáveis meteorológicas totalizando 4 combinações no programa WEKA. Os resultados obtidos na validação dos modelos mostram que: os modelos estatísticos HS, BC podem estimar HG com coeficientes de correlação da ordem de r = 0,84 , e os valores de rMBE e rRMSE variando de +/- 3,0% e 20,0% - 25,0% respectivamente para anos típicos e atípicos; o modelo RM pode estimar HG com valor de r da ordem de 0,960, e valores de rMBE e rRMSE variando de +/- 4,0% e 11,0% - 14,0% respectivamente para anos típicos e atípicos; os indicativos estatísticos obtidos para a combinação 4, a de melhor desempenho, pode estimar HG com valores de r = 0,965 e r = 0,960 e valores para MVS (rMBE = +/- 0,5 %; rRMSE = 12,0 %) e RNA (rMBE = +/- 1,5; rRMSE = 12,0 % para anos típico e atípico). A comparação entre os desempenhos dos modelos HS, BC, RM e as técnicas RNA e MVS na combinação 4, sugere o uso do modelo de RM, quando os dados de RI e UR estejam disponíveis, por ser um método mais prático que o uso das técnicas de aprendizagem de máquinas.
In this paper is described a comparative study between models for estimating daily global solar radiation (HG): Hargreavis-Samani (HS), Bristow-Campbell (BC), a non-linear multiple regression equation developed in this work and two learning machine techniques (Support Vector Machine - MVS and Artificial Neural Network - RNA). The database used in this work was measured in Botucatu in the period of years of 1996-2011. The models were generated and trained with the total database, whereas for model validation were used two fictitious years, denominated typical and atypical, created from the total database. The statistical models HS was estimated by the least saqures method, used for linear regressions, and the BC and RM models were estimated by algorithms for non linear regression that, use iteratives methods to estimate, MVS and RNA, were trained with input of four meteorological variables totalizing four combinations on the WEKA software. The results show that: the statistical models HS, BC can estimate HG with coefficient of correlation of the range of r = 0,84 and the values of rMBE and rRMSE varying from +/- 3,0% and 20,0% - 25,0%, respectively, for typical and atypical years; the multivariate equation can estimate HG with values of r of the order of 0,960, and values of rMBE and rRMSE varying from +/- 4,0% and 11,0% - 14,0%, respectively, for the typical and atypical years; the statistical indicative obtained for combination 4, the with the best performance, can estimate HG with values of r = 0,965 and r = 0,960 and values for MVS (rMBE = +/- 0,5%, rRMSE = 12,0%), and RNA (rMBE = +/- 1,5%; rRMSE = 12,0%, for typical and atypical years). The comparision between the performance of the HS, BC, RM models and RNA and MVS techniques in the conbination 4, suggest the use of the RM model, when the RI and UR datas are availables, for being a method more pratical than the use of the learning machine techniques.
In this paper is described a comparative study between models for estimating daily global solar radiation (HG): Hargreavis-Samani (HS), Bristow-Campbell (BC), a non-linear multiple regression equation developed in this work and two learning machine techniques (Support Vector Machine - MVS and Artificial Neural Network - RNA). The database used in this work was measured in Botucatu in the period of years of 1996-2011. The models were generated and trained with the total database, whereas for model validation were used two fictitious years, denominated typical and atypical, created from the total database. The statistical models HS was estimated by the least saqures method, used for linear regressions, and the BC and RM models were estimated by algorithms for non linear regression that, use iteratives methods to estimate, MVS and RNA, were trained with input of four meteorological variables totalizing four combinations on the WEKA software. The results show that: the statistical models HS, BC can estimate HG with coefficient of correlation of the range of r = 0,84 and the values of rMBE and rRMSE varying from +/- 3,0% and 20,0% - 25,0%, respectively, for typical and atypical years; the multivariate equation can estimate HG with values of r of the order of 0,960, and values of rMBE and rRMSE varying from +/- 4,0% and 11,0% - 14,0%, respectively, for the typical and atypical years; the statistical indicative obtained for combination 4, the with the best performance, can estimate HG with values of r = 0,965 and r = 0,960 and values for MVS (rMBE = +/- 0,5%, rRMSE = 12,0%), and RNA (rMBE = +/- 1,5%; rRMSE = 12,0%, for typical and atypical years). The comparision between the performance of the HS, BC, RM models and RNA and MVS techniques in the conbination 4, suggest the use of the RM model, when the RI and UR datas are availables, for being a method more pratical than the use of the learning machine techniques.
Descrição
Palavras-chave
Radiação solar, Modelo de Hargreaves-Samani, Modelo de Bristow e Campbell, Técnicas de aprendizado de máquinas, Solar radiation, Hargreaves-Samani model, Bristow-Campbell model, Machine Learning Techniques