Attention-Enhanced Sequential GAN for Reliable Groundwater Recharge Time-Series Augmentation

Abstract

Groundwater recharge modeling is critically hindered by the scarcity of long-term, high-resolution time-series data, limiting the robustness and generalization capability of predictive models. We propose the Attention-enhanced Sequential Generative Adversarial Network ( ) to synthesize high-fidelity, multivariate hydrological records, explicitly addressing the complex temporal dependencies required for groundwater dynamics. The architecture incorporates three key innovations: stabilization via the WGAN-GP objective for continuous learning; utilization of a pre-trained LSTM autoencoder to establish a meaningful latent space; and integration of a Self-Attention mechanism within the generative networks to effectively capture critical long-range dependencies, such as multi-year climatic cycles. A three-pronged evaluation demonstrated exceptional data quality: Statistical Fidelity confirmed the preservation of feature relationships, and Temporal Coherence validated the realism of sequential patterns. Crucially, the Predictive Utility was confirmed, with an auxiliary forecasting model trained on synthetic data achieving a Mean Absolute Error (MAE only 4.4% higher) than a model trained on real data. This provides a stable and effective generative approach for time-series augmentation, offering a viable path to developing reliable forecasting tools in data-scarce hydrological contexts.