Empirical model of campus air temperature and urban morphology parameters based on field measurement and machine learning in Singapore

TL;DR

This study develops an empirical model using machine learning to predict campus air temperature in Singapore, highlighting the impact of urban morphology and greenery on Urban Heat Island mitigation.

Contribution

It introduces a novel combination of field measurements, GIS-based urban morphology parameters, and machine learning models, especially random forests, for accurate outdoor temperature prediction.

Findings

Random forests outperformed other models with 4% to 29% lower RMSE.

Greenery significantly mitigates Urban Heat Island effects.

High-rise buildings provide self-shadowing, reducing ambient temperature.

Abstract

The rising air temperature caused by Urban Heat Island (UHI) effect has become a problem for Singapore, it not only affects the thermal comfort of outdoor microclimate environment, but also increases the cooling energy consumption of buildings. As part of a multiscale and multi-physics urban microclimate model, weather stations were installed at 15 points within kent ridge campus of National University of Singapore (NUS) and continuously recorded the microclimate data from February 2019 to May 2019. A Geographical Information System (GIS) map and 3D model were constructed for extracting urban morphology parameters such as BDG, PAVE, WALL and HBDG. Through a site survey, SVF and GnPR were calculated. By using multi-criteria linear regression and machine learning, this research investigated five regression models for prediction of outdoor air temperature including linear regression (LR), k-nearest neighbours (KNN), support vector regression (SVR), decision tree (DT) and random forests (RF). The analysis of variables by best subsets regression showed greenery played crucial role in the mitigation of both daytime and night-time UHI. Pedestrian level wind flow was helpful in heat release in the daytime. High-rise buildings provided self-shadowing to reduce ambient air temperature but higher SVF was harmful to heat release in the night-time. For regression models, RF had the best predictive performance. Average RMSE of RF was reduced by 4% to 29% compared to linear regression. The learning curve indicated that the predictive power of LR could not be improved by additional data provision. In contrast, the downward trend in bias and variance suggested that RF can benefit from the training of big data. During the deployment of learning algorithms, RF continued to outperform other learning algorithms.