River discharge prediction using wavelet-based artificial neural network and long short-term memory models: a case study of Teesta River Basin, India

Discharge prediction with novel machine learning techniques are highly encouraged for its effective response in flood forecasting. This study establishes discharge forecasting models based on artificial neural networks (ANN) and long short-term memory (LSTM) networks at three hydrological stations, Teesta Bazaar, Domohani, and Mekhliganj, within the Teesta river basin, India, for different lead times. Wavelet transform was linked with ANN and LSTM to develop two hybrid models: the wavelet-based artificial neural network (WANN) and the wavelet-based long short-term memory (WLSTM) models. The selection of input variables for the WANN model was carried out through cross-correlation statistics of the discharge data from 2001 to 2017, available only for the monsoon season. Wavelet decomposition of the discharge data culminated in various details and approximation components, which were arranged in specific combinations to generate multiple sub-series. These sub-series were employed to formulate the WANN and WLSTM models. Both LSTM and WLSTM models applied Encoder-Decoder-LSTM (En-De-LSTM) architecture for those stations with multivariate inputs. Furthermore, the Coefficient of Determination (R-2), Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) have been utilized to assess the accuracy of WANN, LSTM, and WLSTM models and compare them. The LSTM networks are characterized by long-term dependencies, memorizing the necessary information, and discarding redundant data. This research highlights that WANN and WLSTM models are superior alternatives for long-term discharge prediction, which can assist the strategy makers in mitigating flood.