Probabilistic forecasting of flood processes based on hybrid deep learning models

The traditional artificial neural network model cannot quantify the uncertainty of flood forecasting and is unable to consider the temporal correlation of flood process forecasting in multi-time continuous forecasting. In this paper, a XAJ -LSTM -EDE-M DIN model is constructed by fusing the Xinanjiang (XAJ) model, the long short-term memory ( XAJ-LSTM-EDE) neural network based on the exogenous input encoder-decoder structure, and the mixture density network (MDN) to achieve probabilistic forecasting of the flood process. The model transforms the point estimates generated by the decoding process into the estimates of conditional probability distributions while considering the temporal correlation of the forecasted flood. The loss function is further established u-sing the maximum likelihood estimation method, and the model parameters are trained by an adaptive moment estimation algorithm. The study results in the two river basins of Lushui and Jianxishow that the model can effectively reflect the uncertainty of the forecast flood without reducing the forecast accuracy of the XAJ-LSTM-EDE model, and obtain reasonable and reliable confidence intervals and excellent probabilistic forecast performance. It provides more risk information for decision-making such as reservoir flood control and scheduling, and also provides a reference for studying the application of deep learning in probabilistic flood forecasting. © 2023 China Water Power Press. All rights reserved.