Predicting flood stages in watersheds with different scales using hourly rainfall dataset: A high-volume rainfall...

Area of Science:

• Hydrology and Water Resource Management
• Environmental Engineering
• Data Science and Machine Learning

Background:

• Accurate prediction of lake water levels and flood flows is critical for effective flood forecasting.
• Lake Carl Blackwell served as a case study, utilizing its extensive historical data.
• Understanding the contribution of both current and antecedent rainfall to water levels is essential.

Purpose of the Study:

• To develop and evaluate accurate methods for predicting instantaneous lake flood stages.
• To assess the predictive power of various rainfall features, including accumulated amounts over extended periods.
• To compare the performance of multiple regression algorithms for flood stage prediction.

Main Methods:

• Engineered hourly rainfall features, including accumulated amounts from present to 600 hours prior, and previous-day lake levels.
• Employed Random Forest Regression (RFR), Neural Network-Multi-layer Perceptron (NN-MLP), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost).
• Utilized linear models like Multiple Linear Regression (MLR), Principal Component Regression (PCR), and Partial Least Square Regression (PLSR) for comparison.

Main Results:

• RFR and NN-MLP achieved high prediction accuracy for lake flood stages (R² up to 0.95, MAE 0.11 ft, RMSE 0.21 ft).
• Non-linear algorithms (XGBoost, SVR) showed slightly lower but still strong performance.
• Linear regression models with dimension reduction exhibited the lowest accuracy.
• The approach demonstrated high accuracy and applicability for surface runoff and streamflow predictions across diverse watershed scales.

Conclusions:

• Machine learning models, particularly RFR and NN-MLP, are highly effective for predicting lake flood stages using accumulated rainfall data.
• The predictive power of rainfall data varies with watershed size, with earlier rainfall being more influential for larger basins.
• The developed methodology offers a robust and broadly applicable solution for hydrological forecasting across different scales.