Neural-UTCI: a neural network corrects the heat stress index and doubles the number of extreme heat days recorded in Rome in 2003

The polynomial approximation of the UTCI index is inaccurate under extreme conditions; its neural network version reduces the error by 87 per cent and increases the number of days of very high heat stress from 15 to 35.

Extreme temperatures are the leading cause of climate-related mortality worldwide, and meteorological services such as health research need to accurately estimate the heat stress experienced. The Universal Thermal Climate Index (UTCI) is one of the most widely used measures of "felt temperature", but it is usually calculated via a polynomial approximation that does not generalize well to extreme conditions.

Fewer errors, less calculation

Researchers including Louise Slater and Milan Klöwer present Neural-UTCI, a neural network that calculates the index more accurately and at a lower cost. The mean squared error drops from 2.78°C to 0.36°C, an 87% better, and the misclassification rate for heat stress drops from 5.3% to 1.7%. The model remains reliable for extreme cold stress, allowing it to be applied anywhere in the world.

The difference is not only academic: for the scorching summer of 2003 in Rome, the number of days of very high heat stress fell from 15 with operational products such as ERA5-HEAT to 35 with Neural-UTCI. The authors see this as a way to strengthen health risk assessments and public warnings, as global warming multiplies extreme episodes.