Aprendizado de máquina aplicado à análise de desempenho térmico de edificações

Abstract

The increasing energy demand for the thermal conditioning of buildings transcends the environmental sphere, also reaching social and economic dimensions, since the need to maintain a comfortable indoor environment entails significant financial costs for families. It is essential to prioritize the thermal performance of buildings, ensuring the lowest possible energy consumption and user comfort. Thus, the present work aims to develop machine learning (ML) models to predict the thermal performance of buildings based on the thermophysical properties of the materials used in their envelope. For this purpose, models based on decision trees were applied to two social housing units. A database with 35,280 combinations of construction systems was built (35 types of walls, 21 types of floors, and 48 types of roofs). Thermal performance assessments were conducted using the metrics of Cooling/Heating Load (CT) and Degree-Hours (GH), based on energy simulations performed in the EnergyPlus software, considering the climate of three cities: Curitiba, São Paulo, and Belém. Initially, using all instances for model training, the performance was satisfactory. Subsequently, when the models were tested with combinations involving new construction systems, a decrease in performance was observed. Despite this, the initial results indicate a good predictive capacity for combinations with known systems, which reduces the need for so many energy simulations. Transfer learning was also explored, to test the models when dealing with a building different from the training data. The metrics showed a significant decrease, highlighting a possible insufficiency of representative variables that could allow for better generalization by the models. It is concluded that the use of ML techniques constitutes a promising approach to promoting greater efficiency and effectiveness in building design, representing a field with wide potential for development and application.