Комплексная система глубокого обучения для прогнозирования запасов подземных вод в условиях гидрологической изменчивости

Ensemble deep learning framework for groundwater storage forecasting under hydrological variability

This study uses deep learning models to present an advanced methodology for forecasting groundwater levels. The primary objective is to estimate monthly streamflow at various gauging stations, analyze long-term groundwater storage trends from 1986 to 2022, and predict future groundwater storage (GWS) for 2028. The majority of research relies on single-model forecasts, without considering regional hydrological variability or integrating minimal-data contexts, despite the increasing use of deep learning models in hydrology. By employing an ensemble deep learning (DL) architecture that combines Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BiLSTM), Stacked Long Short-Term Memory (SLSTM), and Gated Recurrent Unit (GRU), this study closes that gap by accurately predicting groundwater storage over the Middle Mahanadi Basin utilizing Hargreaves–Samani potential evapotranspiration (PET) estimate and SCS-CN runoff. Results reveal that the Ensemble DL model consistently outperforms individual models across all gauging stations, offering the most accurate predictions of GWS changes. This model’s integration of multiple techniques allows it to capture complex patterns and mitigate errors, particularly in regions with high variability. The analysis of seasonal trends reveals that the post-monsoon season exhibits increased groundwater storage, whereas the pre-monsoon and monsoon seasons display a declining trend. In 2004, there was a decrease in GWS across most stations out of 8 stations, likely due to reduced rainfall and increased water extraction, with slight recoveries observed in 2016 and 2022. In conclusion, the Ensemble DL model emerges as the region’s most reliable tool for groundwater forecasting, offering valuable insights for effective water resource planning and management, particularly in drought-prone areas. In drought-prone basins with limited data, the model provides a dependable tool for groundwater management and performs better than individual DL models at every station.