Aquifer-differentiated process-guided GRU for multi-timescale groundwater-level prediction

Groundwater-level prediction supports water-resource management in overexploited aquifers, yet conventional data-driven models often struggle to represent pumping disturbances and contrasting shallow- and deep-aquifer responses. Using groundwater levels, meteorological variables, and agricultural pumping data from four shallow and four deep wells in Guantao County, North China Plain, we developed an aquifer-differentiated process-guided gated recurrent unit model for daily and dekadal (10-day) prediction. Compared with a conventional gated recurrent unit model, the process-guided model reduced group-mean root-mean-square error by 7.2% and 6.1% for shallow and deep wells at the daily scale and by 4.0% and 2.1% at the dekadal scale, respectively. Deep-well performance depended on the pumping-response window, with 60-day averaged pumping providing a stable configuration. These incremental gains show that incorporating pumping disturbances and aquifer-specific response differences as process guidance can improve groundwater-level representation and provide a modeling approach for groundwater management in agricultural abstraction areas.