Machine learning-based modeling for precise runoff forecasts in hydrological systems.

Area of Science:

• Hydrology
• Environmental Science
• Computational Science

Background:

• Accurate daily runoff forecasting is vital for informed decisions regarding rainfall, flooding, and drought.
• Traditional hydrological models often struggle with seasonal data and require manual hyperparameter tuning.
• Existing methods include standalone, integrated, and manually tuned models, with limitations in handling complex seasonal patterns.

Purpose of the Study:

• To develop and evaluate hybrid models for enhanced daily runoff prediction.
• To integrate Extreme Gradient Boosting (XGB), Adaptive Boosting (Adb), and Gene Expression Programming (GEP) with Particle Swarm Optimization (PSO).
• To compare the performance of optimized hybrid models against standalone algorithms.

Main Methods:

• Integration of standalone models (XGB, Adb, GEP) with Particle Swarm Optimization (PSO).
• Development of hybrid models: PSO-XGB, PSO-Adb, and PSO-GEP.
• Evaluation using performance metrics: RMSE, MAE, NSE, R², PBIAS, and a-20 index.
• Utilized nine years of rainfall data for model training and validation.

Main Results:

• The PSO-XGB hybrid model achieved the highest accuracy with R² = 0.96.
• Optimized hybrid models significantly outperformed their standalone counterparts.
• A user-friendly Software Interface (SI) was developed for practical engineering applications.
• A GEP-based equation was derived for practical runoff magnitude estimation.

Conclusions:

• Hybrid models, particularly PSO-XGB, offer superior accuracy for daily runoff forecasting.
• The developed Software Interface facilitates the real-world application of these advanced hydrological models.
• The study provides a robust framework for improving hydrological predictions and water resource management.