Research on Daily Runoff Simulation Based on VMD-CNN-LSTM

Abstract

Runoff simulation plays a crucial role in hydrological research and water resource management for scientific planning and flood control strategy formulation. To address runoff data non-stationarity, this study develops a VMD-CNN-LSTM ensemble model integrating Variational Mode Decomposition, Convolutional Neural Network, and Long Short-Term Memory network, aiming to enhance simulation accuracy and model generalization. Validated using 2008-2016 daily runoff data from the Wuding River Basin, the model demonstrates superior performance with training and testing period R2 values of 0.955 and 0.946, and Nash-Sutcliffe efficiency coefficients of 0.945 and 0.938 respectively, outperforming both standalone LSTM and CNN-LSTM models. Notably, the integrated model shows enhanced capability in peak runoff simulation while maintaining stable accuracy, confirming its robust generalization capacity for hydrological applications.