Data-driven and echo state network-based prediction of wave propagation behavior in dam-break flood

The computational prediction of wave propagation in dam-break floods is a long-standing problem in hydrodynamics and hydrology. We show that a reservoir computing echo state network (RC-ESN) that is well-trained on a minimal amount of data can accurately predict the long-term dynamic behavior of a one-dimensional dam-break flood. We solve the de Saint-Venant equations for a one-dimensional dam-break flood scenario using the Lax-Wendroff numerical scheme and train the RC-ESN model. The results demonstrate that the RC-ESN model has good prediction ability, as it predicts wave propagation behavior 286 time-steps ahead with a root mean square error smaller than 0.01, outperforming the conventional long short-term memory (LSTM) model, which only predicts 81 time-steps ahead. We also provide a sensitivity analysis of prediction accuracy for RC-ESN's key parameters such as training set size, reservoir size, and spectral radius. Results indicate that the RC-ESN is less dependent on training set size, with a medium reservoir size of 1,200-2,600 sufficient. We confirm that the spectral radius has a complex influence on the prediction accuracy and currently recommend a smaller spectral radius. Even when the initial flow depth of the dam break is changed, the prediction horizon of RC-ESN remains greater than that of LSTM.HIGHLIGHTSA machine learning model for predicting wave propagation in dam-break floods is presented. The proposed RC-ESN model well predicts wave propagation 286 time-steps ahead. The prediction ability of the RC-ESN model significantly outperforms the LSTM model. The model is not sensitive to the training dataset size but is influenced by the spectral radius.