Mountain flood forecasting in small watershed based on loop multi-step machine learning regression model.

Area of Science:

• Hydrology and Water Resources
• Environmental Science
• Data Science and Machine Learning

Background:

• Mountain floods in small watersheds are frequent, sudden, and destructive disasters.
• Traditional forecasting methods exhibit high error rates for hourly predictions.
• Accurate and real-time water level forecasting is crucial for mitigating flood impacts.

Purpose of the Study:

• To improve the accuracy and real-time performance of water level forecasting in small watersheds.
• To integrate multi-dimensional disaster-causing factors for enhanced prediction.
• To reduce process errors in Machine Learning (ML) models for flood forecasting.

Main Methods:

• Extracted disaster-causing information and integrated hydrological, meteorological, and geographical factors.
• Employed a short-term prediction window and a loop multi-step input method for ML regression models.
• Developed and compared non-ensemble (Linear Regression, Support Vector Machine Regression, k-Nearest Neighbors Regression) and ensemble (Random Forest Regression, Gradient Boosting Regression) ML models.

Main Results:

• Loop multi-step ensemble ML regression models demonstrated higher accuracy compared to general ML models.
• Ensemble models showed significantly lower time consumption in forecasting.
• The integrated approach effectively reduced ML model process errors for mountain flood prediction.

Conclusions:

• Loop multi-step ensemble ML regression models offer a superior approach for mountain flood forecasting in small watersheds.
• The method provides accurate and time-efficient predictions, outperforming traditional techniques.
• Integrating diverse data sources and advanced ML techniques is key to effective flood disaster management.