Because of the impact of extreme heat waves on society and biodiversity, their study is a key challenge. Physics applied weather forecast systems or climate models can be used to forecast their occurrence or predict their probability.
New research explores the use of deep learning architectures, trained using outputs of a climate model as an alternative strategy to forecast extreme heatwave occurrences. This new work will be useful for several key scientific goals which include the study of climate model statistics, study of the impact of climate change and should eventually be useful for forecasting. First some issues have to be addressed such as class size imbalance that is associated with rare event prediction, and assessing the potential benefits of transfer learning to address the nested nature of extreme events. The researchers trained a Convolutional Neural Network using 1000 years of climate model outputs, with large class undersampling and transfer learning. From measurements of the surface temperature and the 500 hPa geospatial height fields, the trained network achieves significant performance in forecasting the occurrence of long lasting heatwaves. They were able to predict them in three different levels of intensity and as early as 15 days ahead of the start of the event (30 days ahead of the end of the event).
The data used in the study was produced by the Planet Simulator (PlaSim) climate model. It solves the primitive equations for vorticity, divergence, temperature and surface pressure. Moisture is also included by transport of water vapor. The equations are solved using a spectral transform method. Unresolved processes, such as interactive clouds, radiation, moist and dry convection, large scale precipitation, boundary layer fluxes of latent and sensible heat and vertical and horizontal diffusion are parameterized. The model also simulated the combination with land surface scheme and ocean.