Ensemble Runoff Forecasting Based on Multiple Machine Learning Methods

The accuracy of runoff forecasting is influenced by factors, such as the type of prediction model and parameter settings. To account for these uncertainties and leverage the strengths of different runoff forecasting models, in this study, we developed an ensemble forecasting model. It assigned combination weights to each model based on its prediction accuracy, improving the accuracy of short-term runoff forecasts. Using daily runoff data from the Dongqiaoyuan Hydrological Station in the Rongjiang River Basin, China, six runoff forecasting models were established by coupling ensemble empirical mode decomposition (EEMD) and empirical mode decomposition (EMD) with a Backpropagation Neural Network, convolutional neural network, and support vector machine. Two ensemble methods, namely, the optimal weighting method and stacking algorithm, were applied to the EMD- and EEMD-based coupled models, respectively, and their predictive performances were comparatively analyzed. Different data preprocessing methods significantly affected the performances of both the individual and ensemble models. The ensemble models had a considerable reduction in the root mean square error and mean absolute percentage error. The random forest stacking ensemble model showed the highest improvement and prediction accuracy. The model greatly enhances the prediction accuracy of large flows in the Rongjiang River Basin, meeting the needs for short-term forecasting. It can effectively support runoff prediction during extreme environmental changes in the river basin. Additionally, it serves as a valuable reference for enhancing ensemble prediction outcomes in runoff forecasting.