Proposta de rede neural artificial para previsão do conforto térmico utilizando o ASHRAE Global Thermal Comfort DataBase II

• Ano de defesa: 2022
• Autor(a) principal: Santos, Fernanda Marcielli
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Mato Grosso; Brasil; Instituto de Física (IF); UFMT CUC - Cuiabá; Programa de Pós-Graduação em Física Ambiental
• Palavras-chave em Português: Sensação térmica; Aprendizagem de máquina; Climatologia urbana; CNPQ::CIENCIAS EXATAS E DA TERRA::FISICA; Thermal sensation; Machine learning; Urban climatology
• Link de acesso: http://ri.ufmt.br/handle/1/5695

This work aims to develop an Artificial Neural Network (ANN) to predict a thermal comfort model. The ASHRAE database was used, which gathers climatological data from worldwide scientific works, compiled from 1985 to 2018. The ANN was trained using the parameters of the database and its architecture was of the Feed-Forward (FF), type with three triangular layers, NAdam optimizer with a learning rate of 0.01, ReLu activation function in the three layers, training/test percentage of 70:30, batch size of 512 and epochs of 1500. For the input we used: thermal resistance index clothing (clo), individual metabolism rate (met), air temperature (ºC), relative air humidity (%), air velocity (m/s) and monthly external air temperature (ºC); and an output parameter: thermal sensation. The trained ANN presented an error of 13.7%, with at least 100 neurons. The thermal sensation estimate was evaluated by analysis of variance (ANOVA) and showed a good fit. The correlation coefficient (R) was close to 1, indicating a high degree of correlation, while the determination coefficient was significant for the model, considering the low standard error of estimation. The F test value greater than 1 demonstrates improvement due to the regression model fit being much greater than its variation within the model. From the calibration and validation of the ANN, it was possible to verify a good accuracy of the model. The results showed that the ANN was able to make assertive predictions about the thermal sensation, which leads to the conclusion that the model showed a capacity for learning and generalization of satisfactory results.