PROPOSAL FOR AN ARTIFICIAL NEURAL NETWORK TO ESTIMATE THERMAL COMFORT

Abstract

The study presents the development of a procedure using an Artificial Neural Network (ANN) capable of predicting a thermal comfort model. The data used in the research were made available in the ASHRAE database, which gathers climatological data from world scientific studies. For the ANN architecture, we used a Feed-Forward (FF) model with a three-stage triangular network, an Adam optimizer with a learning rate of 0.01, and a ReLu activation function in the three stages. The data was divided into 70% for training and 30% for testing, using a batch size of 512 and running 1,500 epochs. The input parameters used were the thermal resistance index of the clothes, individual metabolism rate, air temperature, relative air humidity, air velocity, and monthly external air temperature; the output parameter was thermal comfort. The neural network development was conducted in Python using existing libraries (Tensorflow and Keras) and was later made available on GitHub. The results found that an adjusted ANN presented an error of 13.7%, considered a good fit for estimating thermal sensation.