Hybrid multi-module for short-term forecasting of river runoff considering spatio-temporal features

ABSTRACT Runoff forecasting is crucial for water resources management, demanding precise models. This study proposes a runoff forecasting model that utilises a hybrid variational mode decomposition (VMD), convolutional neural network (CNN), and long short-term memory network (LSTM) based on the attention mechanism (AM) to enhance the accuracy and stability of runoff forecasting. The volatility of the runoff sequence was significantly reduced by the VMD. The AM focused on extracting the most critical information from the features. The VMD–CNN–AM–LSTM model, using the two-stage decomposition forecasting framework, was used to predict daily runoff at the Jianli (JL) hydrological station in the section from Yichang to JL from 1 October 2006 to 30 October 2022. This model outperformed the model without the hybrid of VMD and/or AM in forecasting runoff, with a root mean square error value of 646.160, a mean absolute error value of 424.124, a mean absolute percentage error value of 2.54%, and an R2 value of 0.9933. Model stability was assessed using the bias-variance, which was found to be significantly more stable than the model without VMD and AM. The VMD and AM optimises runoff forecasting at the target station by utilising upstream stations’ runoff. This improves the accuracy and stability of the model, providing technical support for water resources planning and management.