Document Type : Research Paper
Authors
Associate Professor (Corresponding Author), Department of Watershed Management Engineering, Faculty of Natural Resources, Tarbiat Modares University, Noor, Iran.
Abstract
Introduction:
Modeling groundwater level is challenging due to its complex, nonlinear, spatiotemporal nature. This study aims to present and evaluate a distributed deep learning approach to predict groundwater levels in the Koshk-Fahan aquifer (Sefidrud basin). By combining graph neural networks with convolutional and LSTM structures, we seek to capture spatial and temporal dependencies and improve prediction accuracy over conventional methods.
Materials and Methods:
Four graph-based deep learning structures were used: Graph Neural Network (GNN), Graph Attention Network (GAT), Graph Convolutional Network (GCN), and the hybrid GCN-LSTM. Each piezometric well was a node; the adjacency matrix used inverse distance weighting. Input data included hydro-climatic variables, geology, pumping characteristics, and spatial distances. Evaluation indices: R², RMSE, MAE, NSE, KGE.
Results and discussion:
The results indicated that while the GNN model could partially reconstruct the general trends of groundwater level changes, it lacked accuracy in representing extreme behaviors and temporal dependencies. The GAT model showed a slight improvement over GNN but remained limited in extracting deep temporal patterns. In contrast, the GCN model demonstrated better performance in identifying spatial dependencies, leading to a significant improvement in evaluation metrics. The best performance was achieved by the GCN-LSTM model, which effectively represented both spatial and temporal features simultaneously, showing the highest overlap with observed data. This model reached an R2 of 0.896, RMSE of 0.336, and MAE of 0.269, indicating its high accuracy in predicting groundwater levels.
Conclusion:
Hybrid graph-sequential architectures, especially GCN-LSTM, are highly effective for modeling complex aquifer hydrodynamics. This model predicted groundwater levels with high accuracy and outperformed purely graph-based models. Therefore, it is recommended for operational groundwater prediction and sustainable water management, serving as a novel framework to support better decision-making in water resources.
Keywords
)