Daily Streamflow Prediction and Uncertainty Using a Long Short-Term Memory (LSTM) Network Coupled with Bootstrap

Long short-term memory (LSTM) models with excellent data mining ability have great potential in streamflow prediction. The parameters and structure of the LSTM model, which should be completely determined in an explanatory manner based on the observed datasets, have a significant impact on the model performance. Due to the limitations and uncertainty in the observed datasets, the uncertainty in daily streamflow prediction needs to be quantitatively assessed. In this work, LSTM models are used to predict daily streamflow for two stations in the Mississippi River basin in Iowa, USA, and the performance of LSTM models with different parameters and inputs is investigated to demonstrate the process of determining the optimal parameters. The results show that the LSTM model with optimized parameters and an optimized structure performs the best among the four data-driven models, and the model with selected predictors (inputs) performs better than that without selected predictors. Moreover, the bootstrap method is employed to generate different realizations of the observed datasets that are used for developing LSTM models; thus, the prediction streamflow values from different LSTM models are finally used for uncertainty analysis in daily streamflow prediction. LSTM can be a promising tool for daily streamflow prediction. When LSTM is combined with Bootstrap method, reliable uncertainty quantification of streamflow prediction is also provided.