Application of wavelet-based multivariate long short-term memory models in prediction of stage for Teesta River, India

Abstract

Listen

Translate article

ABSTRACT River stage prediction is indispensably a challenging task in flood-prone river basins to disseminate accurate early warning in advance. In this study, multivariate wavelet-based long short-term memory (WLSTM) models have been developed to predict river stage at six gauging stations of the Teesta River basin in India for 1, 3, and 5-day lead time, the comparison of which has been done with long short-term memory (LSTM) models. Various combinations of wavelet decomposed components were utilized to form different sub-series that were fed as input in WLSTM models. In terms of statistical indicators, both the models yielded exceptionally good results, but the root-mean-square error values of the WLSTM model for 1- and 3-day lead time were minimal compared to the LSTM model. However, the accuracy of the LSTM model in longer lead time prediction is noticeable. Specifically, the WLSTM model predicted the peak stage values more precisely compared to the LSTM model, indicating the potential of wavelet analysis to capture the variations and periodicities of the data by removing the noise. Though the WLSTM model marginally outperformed the LSTM model in prediction accuracy, the results highlight both models as feasible alternatives for longer lead time water level prediction.