An upgraded neural network-based operational procedure for the Universal Thermal Climate Index (UTCI)

Extreme temperatures are the leading cause of climate-related mortality worldwide. To inform mitigation and adaptation strategies, it is crucial to have accurate feels-like temperature measures that quantify thermal stress on human physiology. The Universal Thermal Climate Index (UTCI) is among the most widely used feels-like temperature metrics in climate-health research, applicable in a large range of weather conditions. UTCI is also used by several national and international weather forecasting services to predict thermal stress and issue warnings. However, because of the high complexity of the full UTCI model and associated computational cost, it is operationally approximated by a high-order polynomial to increase computational efficiency.

Here, we demonstrate that a carefully trained and robustly tested neural network model calculates UTCI with significantly greater accuracy compared to the polynomial approximation used in the literature. The neural network model substantially outperforms the polynomial model with a similar computational cost, reducing the approximation error by 86%— from 2.78°C to 0.38°C— and thermal stress misclassification by 76%. We eliminate the need to exclude wind speeds above 17m/s from UTCI calculation, which currently limits the global application of the polynomial approximation. When applied to ERA5 reanalysis data, our model reveals a 25% operational difference in daily heat stress categorization between the two methods in Rome, Italy during the 2003 European heatwave. We provide our UTCI model as openly accessible software, as a more accurate way to calculate UTCI in operational procedures. The neural network UTCI model has the potential to enhance climate-health risk research and improve the accuracy of public weather warning systems.