Evaluating the Performance of Sequence-to-Sequence LSTM Model in Streamflow Modeling under the Beas River, India

Accurate streamflow forecasting is crucial for effectively managing the water related disasters. However, there exist several challenges as the data exhibit multitude of nonlinearity. For this endeavor, various physics based and machine-learning approaches have been investigated with the aim to improve the accuracy. Recently, deep learning techniques such as long short-term memory (LSTM) and sequence-to-sequence (seq2seq) have proven to yield better results capturing the nonlinear time series patterns. This paper focused on exploring the potential of LSTM-seq2seq model for the streamflow forecasting. To demonstrate the proposed method, the data such as daily rainfall, average temperatures streamflow from Beas River watershed, located in the Indian Himalayan region were used. The model performance was found to be good having the Nash-Sutcliffe efficiency of 0.98, correlation coefficient of 0.86, Kling-Gupta Efficiency of 0.88, and mean absolute error of 79.65 cumecs.