Groundwater level prediction is a critical operational function for water utilities managing aquifer-dependent supply systems, agricultural irrigation districts operating under groundwater rights frameworks, and environmental regulators monitoring aquifer sustainability under the EU Water Framework Directive and equivalent national frameworks. Traditional approaches rely on MODFLOW-based numerical groundwater models — physically-based finite difference models that simulate groundwater flow using Darcy's Law — which require extensive hydrogeological characterisation data (aquifer geometry, hydraulic conductivity distributions, storativity) that is expensive to obtain and often uncertain.
LSTM-based data-driven approaches offer an alternative pathway for aquifer systems where sufficient historical monitoring data exists — typically requiring a minimum of 5–10 years of groundwater level records at adequate spatial density — but where the hydrogeological characterisation required for reliable MODFLOW model construction is incomplete. The LSTM Seq2Seq architecture (encoder processes the historical observation sequence; decoder generates the forecast sequence) is well-suited to the multi-step-ahead prediction task (forecasting groundwater levels 30, 90, and 180 days ahead) required for operational water resources planning.
This benchmarking study evaluated LSTM-Seq2Seq models against calibrated MODFLOW-NWT models across 18 aquifer monitoring sites spanning six hydrogeological settings: unconsolidated alluvial, sandstone confined, carbonate karstic, fractured crystalline, volcanic basalt, and coastal saline intrusion. Training data for LSTM models comprised precipitation (CHIRPS), evapotranspiration (MODIS ET), temperature (ERA5), and groundwater level monitoring records at 15-minute to daily resolution. MODFLOW models were calibrated using PEST parameter estimation.
LSTM-Seq2Seq outperformed MODFLOW across 12 of the 18 sites at all forecast horizons, with the largest performance advantages in unconsolidated alluvial systems (RMSE improvement 34–51%) where abundant monitoring data was available and aquifer response to precipitation was rapid. In carbonate karstic systems, MODFLOW maintained a slight edge (RMSE 12% lower) attributable to the physics-based model's ability to represent preferential flow pathways through conduit networks — a flow mechanism that LSTM models represent only implicitly through empirical correlation.
The results support a tiered deployment strategy: LSTM-Seq2Seq as the primary predictive tool for alluvial and sandstone systems with >7 years of monitoring data, with MODFLOW retained for karstic and fractured crystalline systems where physically-explicit flow representation is required. Hybrid physics-informed neural network (PINN) approaches that embed Darcy flow constraints in the LSTM loss function are identified as the most promising research direction for bridging the performance gap in complex hydrogeological settings.