Published: Apr 5, 2023 by Yin Yu
This work is in collaboration with Dr. Ashish Sharma’s group at University of Illinois Urbana-Champaign, and Dr. Zhi-Hua Wang at Arizona State University. It marks a new field of application for our Geometric Deep Learning approach, and demonstrates its strength in making fast, accurate, and interpretable online-predictions of complex dynamical system on graphs, such as a weather system.
Understanding the occurrence and propagation of regional heatwaves is of vital importance to mitigate the consequence of heat extremes. A low-cost, accurate, and timely prediction algorithm for regional heatwaves is desirable. In this study, we use the measurement data set collected from the ground weather stations, together with graph-based deep learning algorithm, to predict the occurrence of regional heatwaves in the lower 48 states of the U.S. The prediction model is trained by the daily weather observations from 91 weather stations and achieves over 90% accuracy in validation. In addition, we extract the spatiotemporal patterns of the climate dynamics learned by the deep learning model to facilitate the interpretation of the results. The proposed modeling framework can be applied to predict other types of extreme events, such as extreme precipitation, drought, and compounded events. The analysis of the model structure will also enhance our understanding of the causal inference between climate regions in the U.S. from a brand-new deep learning perspective.
For more details see our paper GRL2023.