Analog forecasting of heat waves and cold spells using deep learning

Abstract:

Numerical weather prediction (NWP) models have been improving over the past decades but they require ever‐growing computing time and resources and still have sometimes difficulties with predicting some types of extreme weather events. As a result, there is an increasing interest in whether deep learning techniques can improve extreme weather prediction by improving the models and/or pre- or post-processing of the data. We will first discuss some of the recent advances in this area, and then we will introduce a data‐driven framework that is based on analog forecasting (prediction based on past similar patterns, i.e., the analogs) and employs a novel deep learning pattern‐recognition technique (capsule neural networks, CapsNets) and an impact‐based auto-labeling strategy. To provide a proof a concept, we use data from a large‐ensemble fully coupled Earth system model and train CapsNets on large‐scale circulation patterns and surface temperature to predict the geographical region of surface temperature extremes over North America several days ahead. Predicting 1–5 days ahead, CapsNets yield accuracies (recalls) of around 80% (88%), outperforming simpler techniques such as convolutional neural networks and logistic regression. Comparisons with forecast skills from state-of-the-art NWP models and potentials for further improvements will be discussed. The results show the promises of multivariate data‐driven frameworks for accurate and fast extreme weather predictions, which can potentially augment numerical weather prediction efforts in providing early warnings.