Pedestrians' urban thermal comfort: A machine learning assessment through transect walks

As the effect of climate change increases, combined with the urban exodus, the relevance of thermal comfort becomes more and more evident, becoming an urgent need for all. This research presents a methodology for data collection of personal, microclimatic, and morphological variables relevant to the evaluation of thermal comfort in an urban environment. A dataset was collected and processed for a total of 200 surveys in 6 different transect walks in Seville, Spain, half of them in the historical city centre and the other half outside of it. Following the visualization of the data, a descriptive analysis of the main variables was carried out, showing the differences between the UTCI index and the real perceived thermal comfort, as well as the improvement associated to vegetation and spatiality. Additionally, the relevance of both air temperature and radiation on perceived thermal comfort was stated, showing how these two variables are intertwined regarding personal preferences. After this analysis, 3 reliable machine learning algorithms were used to predict the expected comfort: Random Forest, XGBoost and a Multilayer Perceptron. This was achieved with the R language, obtaining results of 0.6 accuracy for thermal comfort prediction in all cases, showing therefore the complexity of the problem.