Research on optimal selection of runoff prediction models based on coupled machine learning methods

Abstract

Runoff fluctuations under the influence of climate change and human activities present a significant challenge and valuable application in constructing high-accuracy runoff prediction models. This study aims to address this challenge by taking the Wanzhou station in the Three Gorges Reservoir area as a case study to optimize various prediction models. The study first selects artificial neural network (ANN) and support vector machine (SVM) as the base models. Then, it evaluates and selects from three time-series decomposition methods. Time-Varying Filter-based Empirical Mode Decomposition (TVF-EMD), Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), and Variational Mode Decomposition (VMD). Subsequently, these decomposition methods are coupled with optimization algorithms, including Whale Optimization Algorithm (WOA), Grasshopper Optimization Algorithm (GOA), and Sparrow Search Algorithm (SSA), to construct various hybrid prediction models. The results indicate that: (1) The single prediction model LSTM demonstrated higher prediction accuracy compared to BP and SVM; (2) The VMD-LSTM model outperformed the CEEMDAN-LSTM and TVF-EMD-LSTM models. Compared to the single LSTM model, the Nash-Sutcliffe Efficiency (NSE) and Pearson's correlation coefficient (R) of the VMD-LSTM model were improved by 15.06% and 6.82%, respectively; (3) Among the machine learning prediction models coupled with various methods, the VMD-SSA-LSTM model achieved the highest accuracy. Compared to the VMD-LSTM model, the NSE and R values of the VMD-SSA-LSTM model were further increased by 13.09% and 4.26%, respectively. Employing a "decomposition-reconstruction" strategy combined with robust optimization algorithms enhances the performance of machine learning prediction models, thereby significantly improving the runoff prediction capabilities in watershed hydrological models.