A long short-term components neural network model with data augmentation for daily runoff forecasting

Abstract

Forecasting daily runoff is of great importance to the allocation of water resources and flood prevention. Many existing methods utilize identical networks to learn the long-term dependencies and the short-term ones. In addition, the importance of data augmentation in a deep network is ignored. In order to attain more accurate and reliable runoff forecasts, this paper proposes a novel framework that designs two different components in nonlinear part to learn the long-term data and the short-term ones, respectively. A long short-term components neural network (LSTCNet) is presented to verify the effectiveness of the framework. Meanwhile, we introduce AR model to capture the linear dependencies. Furthermore, considering that the daily runoff data are unstable and change frequently and sharply in flood season, a linear interpolation method that focuses on the peak values is used to enhance the stability of hydrological data. Experimental results of LSTCNet, the multivariate adaptive regression spline (MARS), the long short-term memory neural networks (LSTM), the attention-mechanism-based LSTM model (AM-LSTM) and the CAGANet model show that LSTCNet achieves the best performance in accurate daily runoff prediction. The LSTCNet’s numerical values of mean absolute error (MAE), root mean square error (RMSE), the Nash-Sutcliffe effciency (NSE), correlation coefficient (CC), and Willmott’s index (WI) can reach 0.32, 1.50, 0.997, 0.999 and 0.999, respectively.