Modelling the energy footprint of urban form in Istanbul: A big data and machine learning approach

There are numerous studies examining the impact of morphological characteristics on the energy performance of buildings. However, there are a limited number of studies that investigate energy consumption in relation to urban forms across metropolitan areas using big data and AI techniques. This study aims to evaluate the energy performance of different urban forms based on their morphological characteristics in Istanbul, employing AI driven methods. In the study, the energy performance certificates of 36,901 buildings, selected through a data preprocessing process, were used to accurately represent different residential typologies. Among eight different machine learning techniques that tested both linear and nonlinear relationships among variables, the most successful results were obtained from techniques based on decision trees that capture nonlinear relationships. Among these techniques, the Random Forest model demonstrated the highest predictive performance, achieving an R-squared score of 77%. The influence of variables was further analyzed using explainable artificial intelligence (XAI), revealing that total construction area and building compactness were the most influential variables in model predictions. This approach also enabled the identification of critical threshold values of these variables affecting energy consumption. Additionally, geometric variables were employed as predictor variables in AI models to estimate energy consumption for 713,648 buildings lacking available data. In another phase of the study, clustering analysis was performed to classify 51,862 building blocks into seven distinct urban form types. The estimated energy consumption data of buildings were integrated with the classified building block data to systematically examine the energy performance of different urban forms. As a result, among the seven different urban forms, ’open skyscraper’ was identified as the most efficient in terms of energy performance, while ’irregular compact low-rise’ was determined to be the least efficient. The study contributes to supporting research in energy-efficient urban planning through the application of AI techniques in both analysis and modelling.