Deep Learning-Based Extreme Heatwave Forecast

Abstract

Because of the impact of extreme heat waves and heat domes on society and biodiversity, their study is a key challenge. We specifically study long-lasting extreme heat waves, which are among the most important for climate impacts. Physics driven weather forecast systems or climate models can be used to forecast their occurrence or predict their probability. The present work explores the use of deep learning architectures, trained using outputs of a climate model, as an alternative strategy to forecast the occurrence of extreme long-lasting heatwave. This new approach will be useful for several key scientific goals which include the study of climate model statistics, building a quantitative proxy for resampling rare events in climate models, study the impact of climate change, and should eventually be useful for forecasting. Fulfilling these important goals implies addressing issues such as class-size imbalance that is intrinsically associated with rare event prediction, assessing the potential benefits of transfer learning to address the nested nature of extreme events (naturally included in less extreme ones). We train a Convolutional Neural Network, using 1,000 years of climate model outputs, with large-class undersampling and transfer learning. From the observed snapshots of the surface temperature and the 500 hPa geopotential height fields, the trained network achieves significant performance in forecasting the occurrence of long-lasting extreme heatwaves. We are able to predict them at 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).

Heat Waves: Physical Understanding and Scientific Challenges

“…Therefore, there is not a single way to boost the skill of statistical subseasonal forecasts. Strategies include improvements in the signal‐to‐noise ratio through temporal and/or spatial data aggregation, or refined techniques to optimize the error losses or the limited sample size associated with extreme events (e.g., by using learning from less extreme events; Jacques‐Dumas et al., 2022; López‐Gómez et al., 2022; Vijverberg et al., 2020). ML methods have also been employed to discover subseasonal drivers of European high temperatures at different time leads (van Straaten et al., 2022), and windows of opportunity for enhanced subseasonal forecasts of HWs (Guigma et al., 2021).…”

“…Interpretability is a common issue in ML, and some methods require a large amount of input data for training and design choices (e.g., Dueben & Bauer, 2018). Approaches to tackle with the shortness of historical data include data augmentation techniques through sampling algorithms that multiply the number of observed events (e.g., Ragone & Bouchet, 2021) or dynamical‐statistical hybrid strategies (e.g., transfer learning from dynamical models to reanalysis data sets; e.g., Jacques‐Dumas et al., 2022). In addition to the selected architecture and hyper‐parameters, a predefined subset of predictors, regions and time leads is often required, which may not capture the true source of predictability due to the intrinsic non‐stationarity and intermittency of the drivers.…”

Machine learning analysis and future risk prediction of weather-sensitive cardiovascular disease mortality during summer in Tokyo, Japan

“…In this study, we approached the evaluation of mortality risk change under future climate conditions by incorporating TL and imbalanced learning (or resampling) 38,39 architectures. A similar attempt was used to successfully forecast extreme heatwaves 40 . Because the characteristics related T max to DRR in Osaka City (corresponding to "source or supporting data" in TL) except for population size were similar to that in Tokyo's 23 wards (corresponding to "target data" in TL) (Fig.…”

Regional Heatwave Prediction Using Graph Neural Network and Weather Station Data

“…Recent years have witnessed the emergence of machine learning (ML)‐based algorithms to predict atmospheric phenomena including temperature and HWs. Such models use training data from Global Climate Models (Jacques‐Dumas et al., 2022), climate reanalysis (Asadollah et al., 2022), and their combination (Jose et al., 2022). ML algorithms based on observations are rather limited primarily due to the spatial discontinuity of weather stations.…”