Development of a novel modeling framework based on weighted kernel extreme learning machine and ridge regression for...

Area of Science:

• Hydrology
• Machine Learning
• Data Science

Background:

• Accurate streamflow forecasting is vital for water management and disaster preparedness.
• Existing machine learning models face challenges in predictor optimization, generalization across time horizons, and high-dimensional data analysis.

Purpose of the Study:

• To develop a novel hybrid machine learning framework, WKELM-R, for enhanced streamflow forecasting.
• To address limitations in predictor selection, model generalization, and handling complex hydrological time series.

Main Methods:

• A hybrid model (WKELM-R) combining ridge regression, locally weighted linear regression, and kernel extreme learning machine was developed.
• Data preprocessing involved multivariate variational mode decomposition (MVMD) for non-stationarity, Boruta-XGBoost for feature selection, and gradient-based optimizer (GBO) for parameter tuning.
• The model was applied to a North Dakota watershed for multi-step-ahead streamflow prediction.

Main Results:

• The WKELM-R model demonstrated high accuracy in streamflow forecasting for multiple time horizons (e.g., R=0.992, RMSE=0.426 at t+3; R=0.997, RMSE=0.249 at t+7; R=0.996, RMSE=0.304 at t+14).
• Performance was validated against existing models using multi-criteria decision-making (MCDM).
• The model effectively handled non-stationarity, selected optimal predictors, and optimized parameters.

Conclusions:

• The proposed WKELM-R framework offers a robust and effective solution for streamflow forecasting.
• The hybrid approach successfully integrates linear and nonlinear dynamics for improved prediction accuracy.
• This research advances hydrological prediction capabilities by overcoming key limitations of traditional machine learning methods.