Spatial-temporal graph neural networks for groundwater data.

Area of Science:

• Environmental Science
• Hydrology
• Data Science
• Machine Learning

Background:

• Groundwater level prediction is complex due to nonlinear and non-stationary data influenced by multiple factors.
• Traditional models face challenges in accurately capturing these complex dynamics.
• Existing methods struggle with data heterogeneity and missing values.

Purpose of the Study:

• To introduce and evaluate a novel application of spatial-temporal graph neural networks (ST-GNNs) for groundwater level prediction.
• To address the limitations of traditional models in handling complex hydrological data.
• To improve the accuracy and robustness of long-term groundwater forecasting.

Main Methods:

• Utilized a modified Multivariate Time Graph Neural Network (a type of ST-GNN).
• Integrated 395 groundwater level time series with auxiliary data (precipitation, evaporation, river stages, pumping data).
• Employed a graph-based framework to capture spatial interconnectivity and temporal dynamics.

Main Results:

• The ST-GNN model demonstrated significant improvements over traditional prediction methods.
• Achieved superior accuracy and robustness in long-term forecasting using both synthetic and measured data.
• Effectively handled missing data and minimized bias in groundwater level predictions.

Conclusions:

• ST-GNNs offer a powerful and effective approach for complex groundwater level prediction.
• The developed model represents a significant advancement in environmental and hydrological modeling.
• This methodology holds substantial potential for enhancing predictive capabilities in water resource management.