Abstract: To explore the effect of deep learning algorithms combined with the multi-source data fusion method in watershed runoff prediction, a bidirectional Long Short-Term Memory (LSTM) neural network model and a data fusion algorithm of the ensemble Kalman filter are combined to construct runoff prediction models for five watersheds in the upper Hanjiang River. These models are verified using long-series hydrometeorological datasets from the study area and atmospheric circulation factor datasets. The results show that in the same prediction period, the models improve the prediction indexes and better capture the extreme values of runoff series in comparison with the traditional LSTM model. After the data fusion algorithm is used to join the atmospheric circulation factor datasets, the evaluation indexes of different watersheds can be further improved, and their time variations are more stable with a longer forecasting period. These prediction models are effective in improving deep learning-based runoff predictions.

Key words: runoff prediction, deep learning, bidirectional long short-term memory neural network, multi source data fusion, ensemble Kalman filtering