Estimation of mean radiant temperature across diverse outdoor spaces: A comparative study of different modeling approaches

The mean radiant temperature (T_mrt) is an important environmental parameter that affects the thermal comfort of human beings. However, both the measurement and estimation of T_mrt have some inherent challenges. This study evaluated the performance of five methods in estimating T_mrt, using data at 670 locations across 14 representative urban forms in Hong Kong. The evaluated methods include the customized globe thermometer method, recalibrated globe thermometer method, SOLWEIG simulation method, regression model, and neural network model. Values calculated from the integral radiation method were used as references for comparison. Results indicate that the customized and recalibrated globe thermometer methods and the SOLWEIG model consistently underestimate T_mrt throughout most of the day, with substantial errors observed at low sun elevations and sunlit sites. The regression model provides a moderate fit to the data. The deep neural network (DNN) model yields the highest estimation accuracy, with an R² of 0.878 and a root mean square error (RMSE) of 1.92 °C. To assess the generalizability of the DNN model, an additional dataset from Singapore is employed, including hourly meteorological data from 28 measurement stations over a two-year period. The DNN model demonstrates strong consistency between modeled T_mrt and the reference across most of the sites, affirming its effectiveness in estimating T_mrt in complex urban environments.