Intrinsic Controls on the Spatiotemporal Predictability of Groundwater Levels in Arid Regions: Insights From Dynamic Graph Learning

Accurate groundwater level prediction is fundamental to achieving sustainable groundwater resource management and protecting groundwater-dependent ecosystems in arid regions. Numerical simulation and machine learning approaches have been widely applied for groundwater level prediction. However, obtaining sufficient external observational data is often challenging in arid areas, which significantly limits the construction of numerical models and the application of deep learning models that depend on multi-source inputs. To enable groundwater level prediction under data-limited conditions and to elucidate the intrinsic drivers of groundwater level dynamics in arid regions, this study proposes a spatiotemporal graph neural network (RD-STGNN) framework that integrates residual decomposition with a dynamic graph learning mechanism. In addition, the influences of hydrogeological conditions on groundwater connectivity are systematically analyzed. Performance evaluation using data from 18 groundwater monitoring wells in the West Liao River Basin, China, demonstrates that this spatiotemporal graph neural network framework outperforms multiple baseline models, including time-series models and static graph neural networks. It maintains high accuracy (NSE > 0.86) when predicting groundwater levels up to two months ahead. In addition, the temporal variability of inter-well dependencies captured by the dynamic graph reflects the evolving hydraulic connectivity and recharge processes within the regional groundwater system. The results indicate that dynamic spatial dependencies inferred from groundwater level variations provide a data-driven representation of hydraulic connectivity, highlighting the intrinsic role of lateral groundwater recharge in shaping groundwater level predictability in arid regions. This study provides a novel data-driven framework for groundwater level prediction in arid regions and reveals the hydrogeological significance of dynamic graph structures in spatiotemporal graph neural networks.