Learning groundwater dynamics across heterogeneous aquifers: A clustering-integrated spatiotemporal deep learning framework

Study region This study focuses on the Zhuoshui River Basin and its alluvial fan in central Taiwan, where groundwater is an important water resource under highly seasonal precipitation and heterogeneous hydrogeological conditions. Study focus We developed a Clustering-Spatiotemporal Integration LSTM (C-STIL) framework for basin-scale groundwater-level forecasting using daily records from 120 monitoring wells during 2013–2019. The framework combines K-means clustering, cluster standardization (CS), data integration (DI), and long short-term memory (LSTM) networks to account for spatial heterogeneity and temporal dependence. Model configurations were evaluated under direct and recursive multi-step forecasting strategies, and spatial transferability was tested by forecasting in regions excluded from training. New hydrological insights for the region Groundwater dynamics in the Zhuoshui River Basin are controlled by lateral hydrogeological variability and antecedent conditions. Although the dataset includes wells screened at different depths, most co-located wells were grouped into the same groundwater-response clusters, suggesting that lateral variability across the alluvial fan was more dominant than vertical differences for this forecasting task. DI improved temporal alignment by incorporating previous groundwater levels, whereas CS reduced regional magnitude bias; their combination produced the best performance. Direct forecasting was reliable up to 10 days, while recursive forecasting enabled longer-horizon projections from an initial groundwater state and future meteorological forcings. The untrained-region experiment further suggests transferable groundwater-response patterns within the basin.