Prediction of non-stationary daily streamflow series based on ensemble learning: a case study of the Wei River Basin, China

Time series decomposition methods are widely used in runoff prediction. However, they often face challenges such as boundary effects and a focus on simulation rather than prediction, which limits their practical utility. This study employs a combination of two-stage stepwise decomposition algorithms—EEMD-VMD, LMD-VMD, SSA-VMD, and STL-VMD—along with financial time series feature derivation techniques, including differencing and growth rate, to enhance the accuracy and usability of runoff prediction. Bayesian optimization, employing the Tree-structured Parzen Estimator, was applied to optimize hyperparameters for XGBoost, LightGBM, and a Blending model that integrates both models. Three new data-driven models—TSD-XGBoost, TSD-LightGBM, and TSD-XGBoost-LightGBM-BayesianRidge (TSD-Blending)—were developed to improve daily runoff prediction performance. An empirical analysis was conducted for the Linjiacun and Weijiabao hydrological stations in the Wei River Basin in China, predicting daily runoff from 1 to 7 days in advance. The main results were: (1) The main results were: (1) The application of TSD technology significantly enhanced the prediction accuracy of the three base models based on two-stage stepwise decomposition. In the runoff prediction for Linjiacun with a one-day lead time, the EEMD-VMD-LightGBM model, after integration of TSD technology, improved its NSE from 0.45 to 0.92, representing an increase of 104.44%. Concurrently, NRMSE and PPTS (5%) decreased from 1.43 and 46.55 to 0.56 and 9.31, respectively, indicating reductions of 60.84% and 80%. (2) Of the two-stage stepwise decomposition models employing TSD, the EEMD-VMD method demonstrated the best performance. After integration of the EEMD-VMD decomposition method, the TSD-XGBoost model achieved NSE values of 0.9927, 0.9933, 0.9698, 0.9575, 0.9931, 0.9559, and 0.9883 for predictions with one- to seven-day lead times in the Linjiacun test set. (3) In comparing the prediction performance of the three base models, TSD-XGBoost demonstrated a higher median and narrower interquartile range, signifying its high stability. Conversely, the data distribution of the TSD-Blending model was more dispersed, failing to achieve the expected enhancements in accuracy and stability, thereby indicating a potential overfitting issue. Overall, the EEMD-VMD-XGBoost model exhibited considerable robustness and effectiveness in predicting highly non-stationary and nonlinear daily runoff sequences. The results of this study provide new insights into runoff prediction and offer robust support for the development and application of hydrological models.