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

Saltwater intrusion poses serious threats to water ecosystems, agricultural production, industrial activities, and especially drinking water safety in estuarine regions. Addressing saltwater intrusion through water resources regulation requires accurate predictions of its duration of impact, namely the hours of chloride exceedance. However, existing saltwater intrusion predictions suffer from short lead times and low accuracy. This study develops a sequence learning framework for the long-term prediction of monthly hours exceeding the chloride threshold (250 mg/L). A sequence-to-sequence prediction paradigm is developed based on the periodic characteristics of saltwater intrusion. Statistical features, particularly extreme-value features, are incorporated to enhance the representation of high-value saltwater intrusion processes. This framework enables reliable prediction of monthly hours exceeding the chloride threshold, with a lead time of 12 months. The results show that extreme-value-based statistical features have correlations with exceedance hours that are comparable to, or even higher than, those based on mean values. Under a 12-month lead time, the sequence learning framework achieved Nash-Sutcliffe Efficiency (NSE) values of 0.817 and 0.798 at the two representative stations. Compared with Random Forest, Gated Recurrent Unit, and Long Short-Term Memory models, the framework improved NSE by more than 0.5 for both the full year and the dry season evaluation. When incorporating the spatial correlation between stations, the joint prediction approach increased the annual NSE by 0.128 relative to single-station prediction. Overall, the framework effectively captured the periodic characteristics of saltwater intrusion and high-value during the dry season, demonstrating strong long-term predictive capability.