Deep generative modeling of groundwater level dynamics using a conditional LSTM-WGAN-GP architecture

This study focuses on the Doroud–Borujerd Plain aquifer in Lorestan Province, western Iran. Monthly groundwater level data from 25 monitoring wells were used for model development and evaluation. The region is characterized by increasing water demand, climate variability, and hydrogeological heterogeneity that have intensified pressure on the aquifer system. This study aims to develop and evaluate a novel deep learning framework, termed Conditional LSTM-WGAN-GP, that integrates Long Short-Term Memory (LSTM) networks with a Conditional Wasserstein Generative Adversarial Network using Gradient Penalty (WGAN-GP) for monthly groundwater level forecasting. The proposed architecture is designed to simultaneously capture temporal dynamics and probability distributions by conditioning the generator on historical observations. Predictive performance was assessed using four widely adopted hydrological indicators: Root Mean Square Error (RMSE), Coefficient of Determination (R²), Nash–Sutcliffe Efficiency (NSE), and Kling–Gupta Efficiency (KGE). In addition to statistical metrics, time-series comparisons, scatter plots, and density distribution analyses were employed for comprehensive evaluation. During the independent testing period, the proposed model achieved a mean RMSE of 1.55 m, R² of 0.69, NSE of 0.47, and KGE of 0.70 across the 25 wells. The model demonstrated satisfactory performance in 13 out of 25 wells (NSE ≥ 0.50), although performance varied across locations due to local hydrogeological heterogeneity. The results demonstrated that the proposed model successfully reproduced the temporal dynamics and statistical distribution of groundwater levels across most wells. Stable convergence of the adversarial training process confirmed the effectiveness of the Gradient Penalty strategy in improving optimization stability, while the LSTM component effectively captured seasonal and long-term dependencies. The close agreement between observed and simulated groundwater levels indicates that integrating temporal sequence learning with adversarial distribution learning provides a robust framework for groundwater forecasting in the region. • A novel Conditional LSTM-WGAN-GP framework was developed for groundwater forecasting. • Conditional adversarial learning improved temporal consistency and prediction reliability. • Gradient Penalty stabilized GAN training and prevented mode collapse. • The proposed model accurately preserved both groundwater dynamics and data distributions.