Forecasting heat stress using data-driven model outputs

Heat stress represents a major risk to human health, making the development of advanced warning systems essential for safeguarding individuals and communities. Data-driven models, such as FourCastNet, PanguWeather, and GraphCast, provide rapid, accurate, and publicly accessible forecasts of meteorological variables. However, these models do not provide all the variables required to calculate thermal stress indices, such as the Universal Thermal Climate Index (UTCI). To address this limitation, this study proposes a method to estimate the UTCI for southern South America using a subset of variables available from these data-driven models. First, feature selection techniques were applied, including stepwise selection and a wrapper evolutionary approach based on the Probabilistic Coral-Reef Optimization with Substrate Layers algorithm (PCRO-SL). These techniques were used to identify key variables both at the individual grid point level and within homogeneous regions of the UTCI, defined through k-means clustering. The selected variables were then incorporated into various regression and classification models, ranging from simple linear methods to the advanced Light Gradient Boosting Machine (LGBM). The performance of these models was evaluated against the ground-truth UTCI data provided by ERA5-HEAT. Results show that the combination of PCRO-SL and LGBM yielded the most accurate UTCI estimates. Key variables identified included 2-meter temperature, specific humidity, and low-level wind components. Finally, using forecasts of these selected variables from FourCastNet, PanguWeather, and GraphCast, the method was applied to estimate the UTCI during a heatwave. Forecasts up to three days show good agreement between the observed and modeled thermal stress category. Future work will explore improvements through post-processing techniques for the meteorological variables provided by data-driven models.

 

Acknowledgments: This work was supported by the SAFETE project, which has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 847635 (UNA4CAREER). This work has also been partially supported by the project PID2023-150663NB-C21 of the Spanish Ministry of Science, Innovation and Universities (MICINNU), and by the EU-funded H2020 project CLINT (Grant Agreement No. 101003876).