Daily streamflow forecasting using hybrid long short-term memory model

Abstract

Characterized by heterogeneity, complexity and non-stationary, streamflow forecasting has always been a challenge in hydrological sciences. In this study, a multiscale wavelet decomposition method with long short-term memory model (WLSTM) is developed to handle the daily streamflow forecasting. Discrete wavelet transform (DWT) is employed to extract multiscale features, which are then simulated by long short-term memory models (LSTMs), respectively. The outputs of the different scales LSTMs are finally reconstructed toward the forecasting results. Seven years daily streamflow sequences of three tributaries and one mainstream in the upper reaches of the Yangtze River are investigated by the WLSTM models. For comparison, standard LSTM, and traditional ANN are chosen for the same streamflow forecasting task. Experimental results show that the proposed hybrid model is better than other comparison models in terms of evaluation indicators. Considering that large floods often occur in the Yangtze River Basin, the performance of flood forecasting is also analyzed and the result shows that forecasting errors of WLSTM are more concentrated, which means WLSTM outperforms traditional ANN and LSTM for the extreme flood forecasting.