A good simulation of heatwave frequency variation by a machine-learning model

Against the background of global warming, the European region has experienced several severe extreme heatwave events in recent decades, which have had a huge socioeconomic and environmental impacts. It is therefore important for governments to improve the ability of climate models to predict heatwave events. However, the ability of most climate models in this regard is highly limited at present, mainly because they cannot simulate well the feedback between the atmospheric and boundary-layer physical quantities. Thus, we are still lacking in our understanding of the roles and relative contributions of these processes. It remains a great challenge to improve the ability of models to predict heat waves.

Recently, the team of Professor Jiaxiao Jing from Zhejiang University simulated the variations in summer heatwave frequency in eastern Europe by using the spatiotemporal decomposition method combined with the LightGBM machine-learning model and analyzed the contributions of multiple climate factors from the lower boundary layer (see figure, below). The results have been published in Atmospheric and Oceanic Science Letters.

“The climate factors selected in the machine-learning model include the sea surface temperature, soil moisture, snow, and sea ice in the previous winter, previous spring, and the simultaneous summer. Our results show that the LightGBM model can simulate well the variation in summer heatwave frequency in eastern Europe, and the SST factor contributes the most to the model simulation”, says Prof. Jia.

In addition, further study found that the best simulation results were obtained by using the climate factors in the previous winter.

“Clearly, the underlying surface climate factors, especially the sea surface temperature, are very important for improving the climate model’s prediction of the characteristics of summer heat waves in eastern Europe, but its related mechanisms need to be further studied”, adds Prof. Jia.