Coupled intelligent prediction model for medium- to long-term runoff based on teleconnection factors selection and...

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Accurate medium- to long-term runoff forecasting is crucial for water resource management, flood control, and ecological restoration.
• Traditional statistical models struggle with the nonlinearity, nonstationarity, and multi-source data inherent in runoff processes, limiting prediction accuracy.
• Existing methods often overlook complex interactions between numerous influencing factors, leading to incomplete and unreliable forecasts.

Purpose of the Study:

• To develop and evaluate novel coupled intelligent prediction models for enhanced medium- to long-term runoff forecasting.
• To integrate Random Forest (RF) with Support Vector Regression (SVR) and Multilayer Perceptron Regression (MLPR) to leverage their complementary strengths.
• To assess the performance of these coupled models in the Yalong River Basin (YLRB) for practical water management applications.

Main Methods:

• Developed two coupled models: RF-SVR and RF-MLPR, combining RF's data dimensionality reduction with SVR/MLPR's nonlinearity handling.
• Utilized MLPR's deep learning capabilities for extracting complex latent information, particularly beneficial for long-term predictions.
• Tested models in the Yalong River Basin, evaluating performance across various forecast horizons and hydrological stations.

Main Results:

• Atmospheric circulation indices showed a one-month lag effect on YLRB runoff, offering insights for scheduling and prevention.
• Coupled models effectively reduced collinearity and redundancy, significantly improving prediction accuracy across all forecast periods.
• RF-MLPR model demonstrated superior performance over RF-SVR, with notable improvements in Nash-Sutcliffe efficiency (NSE) and R2 metrics, especially for longer forecast horizons.

Conclusions:

• The coupled RF-SVR and RF-MLPR models offer substantial improvements in hydrological forecasting accuracy compared to single models.
• The RF-MLPR model shows particular promise for long-term forecasting due to its advanced deep learning structure.
• These models provide practical tools for water resource management, flood control, and drought mitigation, with broad applicability in similar hydrological regions.