Random Forest Model for Salinity Prediction in the Nakdong River Estuary Reflecting Spatio-temporal Lag Effects
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
<i>The Nakdong River Estuary, characterized by ecological diversity and economic importance, faces challenges from salinity intrusion owing to estuary bank operations. This study aimed to develop a rapid and accurate prediction system for salinity concentrations to support effective water resource management. Time series data from 2021 were analyzed to identify correlations between salinity and environmental factors. Granger causality analysis determined that discharge, seawater inflow, tidal level, and rainfall significantly influence salinity, with optimal temporal lags for each variable. A random forest model, combined with a sliding window approach, was used to forecast salinity for short-term (1–6 h) and medium-term (12–24 h) horizons. The model incorporating lagged variables achieved high accuracy in short-term predictions (coefficient of determination > 0.9), effectively capturing the physical travel time of water. Medium-term predictions also showed improved performance compared with baseline models, though prediction uncertainty increased with longer horizons due to environmental variability. Incorporating lagged variables based on physical and statistical analysis enhances the accuracy of salinity forecasts in the Nakdong River Estuary. These findings support the development of data-driven management strategies for estuarine environments.</i>
Abstract
The Nakdong River Estuary, characterized by ecological diversity and economic importance, faces challenges from salinity intrusion owing to estuary bank operations. This study aimed to develop a rapid and accurate prediction system for salinity concentrations to support effective water resource management. Time series data from 2021 were analyzed to identify correlations between salinity and environmental factors. Granger causality analysis determined that discharge, seawater inflow, tidal level, and rainfall significantly influence salinity, with optimal temporal lags for each variable. A random forest model, combined with a sliding window approach, was used to forecast salinity for short-term (1–6 h) and medium-term (12–24 h) horizons. The model incorporating lagged variables achieved high accuracy in short-term predictions (coefficient of determination > 0.9), effectively capturing the physical travel time of water. Medium-term predictions also showed improved performance compared with baseline models, though prediction uncertainty increased with longer horizons due to environmental variability. Incorporating lagged variables based on physical and statistical analysis enhances the accuracy of salinity forecasts in the Nakdong River Estuary. These findings support the development of data-driven management strategies for estuarine environments.