Two-step hybrid model for monthly runoff prediction utilizing integrated machine learning algorithms and dual signal decompositions

Abstract

Runoff is pivotal in water resource management and ecological conservation. Current research predominantly emphasizes enhancing the precision of machine learning-based runoff predictions, with limited focus on their physical interpretability. This study introduces an innovative two-step hybrid runoff prediction framework tailored for the headwater region of the Yellow River Basin (YRB) to improve prediction accuracy and elucidate the runoff modeling process. The framework integrates machine learning techniques with dual signal decomposition approaches, incorporating diverse hydrometeorological and geographic indicators. Long Short-Term Memory (LSTM) and eXtreme Gradient Boosting (XGBoost) algorithms were employed to predict monthly runoff generation in sub-basins delineated by the Soil and Water Assessment Tool (SWAT), which were subsequently integrated using a Recurrent Neural Network (RNN) for monthly runoff concentration prediction. Results indicate that the proposed models delivered superior prediction performance compared to the SWAT model (R 2 = 0.86, NSE = 0.85), with the LSTM-based two-step hybrid model (R 2 = 0.90, NSE = 0.90) outperforming the XGBoost-based model (R 2 = 0.89, NSE = 0.88). The dual decomposition method, integrating seasonal-trend decomposition based on loess (STL) and successive variational mode decomposition (SVMD), demonstrated exceptional efficacy in addressing the complexities of hydrometeorological time series. Models decomposed by STL-SVMD exhibited the highest average R 2 and NSE values, as well as the lowest RMSE and MAE values in sub-basin runoff calculations. The low standard deviations of performance metrics further underscored the stability of these models across all sub-basins. This study demonstrates the efficacy of the proposed two-step hybrid model for simulating physical runoff processes in the headwater region of the YRB, providing valuable insights for regional hydrological cycle research and hydro-ecological security. • Propose a novel hybrid runoff generation and concentration forecasting framework. • Explore the performance of different decomposition and machine learning models. • The LSTM-based two-step hybrid model exhibited remarkable prediction accuracy. • Stepwise prediction combining geographic and hydrometeorological information increases the model's physical interpretability. • Dual signal decomposition improves accuracy by effectively managing complex time series.

Topics & Concepts