Long-term Prediction of Saltwater Intrusion Based on Sequence Learning Framework

Area of Science:

• Environmental Science
• Hydrology
• Water Resource Management

Background:

• Saltwater intrusion threatens water resources and safety in estuarine regions.
• Current prediction methods lack sufficient lead time and accuracy.
• Accurate prediction of chloride exceedance hours is crucial for water management.

Purpose of the Study:

• To develop a long-term prediction framework for monthly hours exceeding the chloride threshold (250 mg/L).
• To enhance saltwater intrusion prediction accuracy and lead time using sequence learning and extreme-value statistics.

Main Methods:

• Developed a sequence-to-sequence prediction framework leveraging periodic characteristics of saltwater intrusion.
• Incorporated extreme-value statistical features to improve prediction of high-value intrusion events.
• Evaluated model performance using Nash-Sutcliffe Efficiency (NSE) and compared with other machine learning models.
• Investigated the impact of spatial correlation between stations on prediction accuracy.

Main Results:

• The sequence learning framework achieved high predictive accuracy with a 12-month lead time (NSE values of 0.817 and 0.798).
• Extreme-value features showed strong correlations with exceedance hours, comparable to mean-based features.
• The framework significantly outperformed Random Forest, Gated Recurrent Unit, and Long Short-Term Memory models, improving NSE by over 0.5.
• Joint prediction incorporating spatial correlation further increased annual NSE by 0.128.

Conclusions:

• The developed framework effectively captures periodic and high-value saltwater intrusion characteristics.
• Demonstrated strong long-term predictive capability for monthly hours of chloride exceedance.
• Offers a reliable tool for water resource regulation and management in estuarine areas facing saltwater intrusion.