Simulation of spring discharge using graph neural networks at Niangziguan Springs, China

Abstract

Over past decades, extensive groundwater development in karst aquifers has led to significant declines in groundwater levels and spring discharges. While the recent implementation of sustainable karst aquifer development policies appears to have lessened the declining trend of spring discharge, many aspects of karst aquifers remain to be explored. This paper applied two spectral-based graph neural networks (GNN) models of ChebNet and graph convolutional networks (GCN) for karst hydrological processes by considering the correlations among precipitation, spring discharge, and human impacts. This approach captures the spatial dependence of precipitation infiltration, groundwater propagation in heterogeneous karst aquifer and spring discharge from bared karst aquifer. This study proposed three graph structures (complete graph, information flow graph, and association graph of groundwater flow field) to depict relations between precipitation and spring discharge in the GNN models in Niangziguan Springs, China. The results show that the association graph of groundwater flow field is the optimal graph structure of GNN models. Based on the optimal structure, we investigated precipitation-driven-discharge models using ChebNet and GCN to predict the discharge. The results show that the high-order ChebNet is more adaptable to simulate the karst hydrological processes with nonlinear and nonstationary behaviors than that GCN. Moreover, the study confirms that the groundwater sustainable development policy has achieved a prominent contribution 2 m3/s in water conservancy in Niangziguan Springs catchment.