Enhancing groundwater level forecasting using LSTM: integrating hydrogeological variability for improved accuracy (case study: Dez Plain, Iran)

The performance of data driven models for groundwater level forecasting varies widely across hydrogeological settings. Conventional practice often responds with increasingly complex model architectures, while overlooking the inherent predictability of the underlying system. This study challenges that approach and argues that hydrogeological stability, largely shaped by anthropogenic pressure, is the primary control on forecast accuracy. Fixed structure Long Short Term Memory (LSTM) and Support Vector Regression (SVR) models were applied to groundwater level data from five observation wells in Iran’s Dez Plain, representing conditions from hydrologically stable to highly unstable. System stability was quantified using three complementary indices—Maximum Annual Fluctuation (MaxAF), Average Annual Fluctuation (AvgAF), and the Average September–October Difference (ASODA)—which describe both the magnitude and persistence of groundwater level variability. Results from 100 bootstrap iterations show that in stabilized conditions, LSTM achieved low absolute error (RMSE ≈ 0.25 m), whereas the static SVR failed to reproduce observed dynamics (NSE < 0). In contrast, under pumping dominated and highly unstable conditions, LSTM produced high NSE (NSE > 0.90) but with large absolute errors (RMSE ≈ 0.60 m). This discrepancy reveals a “Variance Paradox,” where strong normalized performance metrics mask limited practical accuracy in high variance systems. We identify a physical “Predictability Barrier,” expressed as a monotonic increase in forecast error with hydrogeological instability. The findings support a shift from model optimization toward system diagnosis. A tiered management framework is proposed, guided by qualitative interpretation of forecast uncertainty: narrow Confidence Intervals (CIs) indicate stable, predictable systems (Tier 1), while broad CIs signal unstable conditions requiring physical restoration (Tier 2).