Graph Neural Networks for Enhancing Ensemble Forecasts of Extreme Rainfall

TL;DR

The paper tackles the challenge of unreliable extreme rainfall forecasts by post-processing ensemble outputs with a graph neural network that respects spatial station topology. It introduces a tail-aware distributional regression network that models precipitation as a mixed distribution with a discrete mass at zero and a generalized Pareto tail above a threshold $u$, trained via CRPS and operating on a log-transformed space $h(y)=\log(y+\varepsilon)$. Spatial dependencies are captured through a graph over stations, using a permutation-invariant DeepSet embedding for ensemble features and a Graph Isomorphism Network with Edge features (GINE) to predict station-wise distribution parameters. On the EUPPBench dataset, the method improves tail calibration and spatial coherence for extreme rainfall compared to baselines, with notable gains at longer lead times and for extreme-event metrics.

Abstract

Climate change is increasing the occurrence of extreme precipitation events, threatening infrastructure, agriculture, and public safety. Ensemble prediction systems provide probabilistic forecasts but exhibit biases and difficulties in capturing extreme weather. While post-processing techniques aim to enhance forecast accuracy, they rarely focus on precipitation, which exhibits complex spatial dependencies and tail behavior. Our novel framework leverages graph neural networks to post-process ensemble forecasts, specifically modeling the extremes of the underlying distribution. This allows to capture spatial dependencies and improves forecast accuracy for extreme events, thus leading to more reliable forecasts and mitigating risks of extreme precipitation and flooding.

Figures (2)

• Figure 1: The left panel shows the precipitation rates on April 29, 2018 with a highlighted extreme precipitation occurrence of about 70mm over 6 hours. The two right plots show the threshold exceedance probability $\mathbb{P}[Y>25\text{mm}]$ for our model and the ensemble prediction. In contrast to the ensemble, our model assigns a nonzero probability ($0.63\%$) to the corresponding event.
• Figure 2: The figure compares a sample prediction of the different modeling methods. In our proposed approach, the low precipitation scenarios are modeled via a discrete point mass that accounts for the event of no precipitation, while the GPD models the tail behavior over a certain threshold.