Towards Universal Thermal Climate Index Prediction via machine learning approaches

Abstract

Maintaining a proper outdoor thermal environment can encourage people to engage in healthy outdoor activities, reducing residential energy consumption. Urban designers and planners rely on different indexes to calculate and predict outdoor thermal environments, such as UTCI. Existing prediction models of UTCI focus on the relationship between environmental parameters, human perception, and personal factors. However, urban characteristics impacts on UTCI have not yet been embedded in UTCI prediction research. Thus, this study investigated 30 cities worldwide with diverse urban characteristics using ML methods to forecast the UTCI and develop a nuanced index of the relationship between the UTCI and urban characteristics. Specifically, this integrates physics-based parametric modeling using urban features and outdoor thermal comfort modeling with Honeybee, combined with ML techniques such as LSTM, Gaussian Process Regression, RF, KNN, DT, and ANN. Our results show that the ANN model achieved a notable level of precision with MSE=0.0008 and an $R^2$ Score=97%, demonstrating the robustness of ML in environmental modeling. The most critical variable of urban characteristics index to UTCI is ‘Average Volume’, and the model’s output is positively impacted by large SHAP values. Similarly, the ‘Green Space Ratio’ and ‘Average Height’ show a variety of impacts, indicating they affect UTCI estimations in different ways. Our study aims to support informed decision-making for large-scale sustainable city planning through a comprehensive data-driven model that enables more nuanced and precise global predictions of outdoor thermal comfort.