Application of Attention Mechanism Combined with Long Short-Term Memory for Forecasting Dissolved Oxygen in Ganga River

Abstract

Accurate forecasting of water quality parameters is a significant part of the process of water resource management. In this paper we demonstrate the applicability of Long Short-Term Memory (LSTM) combined with attention mechanism for the long-term forecasting (after 24 h) of Dissolved Oxygen content at various stations of Ganga River flowing through the state of Uttar Pradesh, India. In the given model, the hidden states of the LSTM units are passed to the attention layer. The attention layer then gives different weights to the hidden states based on their relevance. The performance of the models is evaluated using root mean square error, mean absolute error and coefficient of determination. The experimental results indicate that combining attention mechanism with LSTM significantly improves the forecasted values of Dissolved Oxygen when compared with state-of-the-art models like Recurrent Neural Network, LSTM, and bidirectional LSTM. The demonstrated model is particularly useful during the availability of only univariate datasets.