Groundwater depth prediction based on CNN-GRU-attention model

Area of Science:

• Hydrology and Water Resources
• Artificial Intelligence
• Environmental Science

Background:

• Groundwater is a vital freshwater source in arid regions, facing threats from low precipitation and droughts.
• Effective groundwater level prediction is crucial for sustainable water resource management.

Purpose of the Study:

• To develop and evaluate a hybrid deep learning model for accurate groundwater depth forecasting.
• To assess the model's performance under data scarcity and drought scenarios.

Main Methods:

• A spatiotemporal analysis of groundwater depth dynamics was conducted.
• A hybrid CNN-GRU-Attention model was developed, integrating CNN, GRU, and attention mechanisms.
• Input variables (evaporation, precipitation, temperature, extraction) were selected using the Shannon entropy method.

Main Results:

• The CNN-GRU-Attention model significantly outperformed other methods, achieving high accuracy (MAE: 0.4-0.5, RMSE: 0.5-0.6, R²: 0.8-0.9).
• The model maintained superior predictive performance even with 10-25% data reduction.
• A reduction of 42 million m³ in extraction was identified as optimal for groundwater stability during drought.

Conclusions:

• The CNN-GRU-Attention model offers a robust framework for groundwater depth prediction in arid environments.
• The findings provide valuable insights for implementing effective groundwater management strategies, particularly during drought periods.