Author(s): Qingwen Deng; Dong Wang; Along Zhang

Linked Author(s):

Keywords: Hydrological risk assessment; Nonstationary extreme precipitation; Spatio-temporal graph neural network; Spatial heterogeneity

Abstract: Non-stationarity in extreme precipitation challenges traditional stationary assumptions in hydrologic design. While previous Generalized Additive Models for Location, Scale and Shape (GAMLSS) frameworks have identified drivers at station levels, they often overlook explicit spatial dependencies. This study proposes a Spatio-Temporal Graph Neural Network (ST-GNN) to characterize the spatial heterogeneity of non-stationary extremes and atmospheric circulation controls over China. By integrating graph convolutions with GAMLSS-derived physical benchmarks, the model identifies "non-stationary functional regions" that transcend traditional climate zones.

DOI:

Year: 2026