Uso de redes neurais artificiais para estimação do voto de percepção térmica a céu aberto em clima tropical de savana

Abstract

The thermal perception knowledge in open-air environments requires the application of field interviews. These campaigns are time-consuming and have a high execution cost since it is necessary to collect data at different times of the year and carry out a large number of interviews to characterize the individuals' thermal adaptation. To optimize this stage, the application of the Artificial Neural Networks (ANN) technique is envisaged, which enables the data integration and the models creation which are capable of predicting human behavior in response to environmental characteristics. This study aims to evaluate the accuracy of ANN application in estimating declared thermal perception for open-air environments located in a tropical savanna climate region. The methodology uses a secondary database of thermal perceptions declared by local respondents, obtained simultaneously with meteorological variables. The ANN was developed using the MATLAB application, with structure and learning algorithm-specific configurations. Estimations with the modeled neural network involved nine different combinations of input variables for learning, focusing on determining the best accuracy with the available database. The best results are shown when the network is trained using anthropometric, individual, and meteorological variables and the data collection location as inputs, obtaining root mean square errors of 0.115 and a correlation coefficient of 0.783. We also found that gender was the input variable that least affected the accuracy of the network when it was excluded from training.