Predicting dissolved oxygen concentration in river using new advanced machines learning: Long-short term memory (LSTM) deep learning

Abstract

Accurate estimation of the dissolved oxygen concentration is critical and of significant importance for several environmental applications. Over the years, many types of models have been proposed to provide a more accurate estimation of dissolved oxygen at different time scales. Recently, the deep learning paradigm has been increasingly used in several environmental and engineering applications. This study presents the application of long short-term memory (LSTM) deep learning for dissolved oxygen (DO) prediction in rivers. The model was trained and calibrated using three predictors: (i) river water temperature (T_w), (ii) air temperature, and (iii) river discharge (Q). The variables were measured on an hourly time scale and collected from two USGS stations. The LSTM model was compared against genetic programming (GP), the group method of data handling neural network (GMDH), support vector regression (SVR), and Gaussian process regression (GPR) models. The proposed models were evaluated using well-known performance metrics, namely the coefficient of correlation (R), Nash-Sutcliffe efficiency (NSE), mean absolute error (MAE), and root mean square error (RMSE). This study demonstrates the utility and robustness of the proposed models for predicting dissolved oxygen, and the GPR was found to be slightly better than the SVR model, and significantly better than the GMDH, LSTM, GP, and MLR models. It was also demonstrated that the LSTM ranked third. Numerical results showed that using T_w, T_a, and Q as predictors combined with the periodicity (i.e., hour, day, and month number) leads to high accuracies with R, NSE, RMSE, and MAE of 0.991, 0.981, 0.085, and 0.062, respectively.