Abstract: Through embedding a long short-term memory (LSTM) neural network in the encoder-decoder (ED) structure, this study constructs a LSTM-ED deep learning model and uses the Bayesian probabilistic forecasting processor to quantify flood forecast uncertainty. A probabilistic operational approach is developed, and the influence of rainfall forecast information on the probabilistic forecast performance is discussed. The new models are trained and validated using 6h rainfall and runoff series during 2010-2021 flood seasons in the interval basin between the Xiangjiaba reservoir and Three Gorges reservoir to forecast its floods for the forecast periods of 1 - 7 d. The results show the LSTM-ED model has a forecast accuracy higher than that of LSTM, achieving the Nash efficiency coefficients above 0.92 for the validation of 1 – 7 d forecast periods. The continuous ranking probability score values of the probabilistic forecasts are reduced by 26.8% - 32.7% relative to the mean absolute errors, effectively quantifying forecast uncertainties. The probabilistic forecasts could be further improved by considering rainfall forecast information so as to provide more reliable risk information for decision-making of reservoir scheduling.

Key words: flood forecast, deep learning, encoder-decoder structure, probabilistic forecasting, uncertainty analysis