Multi-head attention driven aggregation-differentiation network for probabilistic groundwater depth forecasting and depth-stratified early warning: A multi-site efficient framework

Accurate groundwater depth (GWD) forecasting is critical for sustainable water management, yet remains challenging due to the reliance on multi-source data and inadequate uncertainty-guided warning strategies. This study presents a Multi-Head Attention driven Aggregation-Differentiation Network (MHA-ADNet), a novel efficient framework that requires only historical GWD records for probabilistic multi-site forecasting. MHA-ADNet decouples the learning process into an aggregation module, which extracts shared temporal features across monitoring wells by multi-head attention, and parallel differentiation modules that capture site-specific characteristics, with integrated quantile regression for uncertainty assessment. Validated on data from 43 wells in Eastern China, the framework achieves superior performance across multi-step predictions (7-day R 2 ≥0.89, 14-day R 2 ≥0.85, 30-day R 2 ≥0.72), consistently outperforming baseline models. It demonstrates higher reliability in deep inland aquifers and facilitates depth-stratified early warning strategies using prediction intervals. MHA-ADNet provides a practical tool for probabilistic groundwater forecasting and adaptive warnings, making it particularly valuable for management in data-scarce regions. • Probabilistic prediction using multi-head attention and aggregation-differentiation. • MHA-ADNet outperforms baseline models with higher accuracy and lower uncertainty. • MHA-ADNet shows fast recovery for independent, gradual changes in deep groundwater. • Short-term predictions rely on recent 3 steps and long-term on extended periods. • Prediction intervals enable depth-stratified early warnings from 7 to 30 days ahead.