Prediction of dissolved oxygen in urban rivers at the Three Gorges Reservoir, China: extreme learning machines (ELM) versus artificial neural network (ANN)

In the present study, two non-linear mathematical modelling approaches, namely, extreme learning machine (ELM) and multilayer perceptron neural network (MLPNN) were developed to predict daily dissolved oxygen (DO) concentrations. Water quality data from four urban rivers in the backwater zone of the Three Gorges Reservoir, China were used. The water quality data selected consisted of daily observed water temperature, pH, permanganate index, ammonia nitrogen, electrical conductivity, chemical oxygen demand, total nitrogen, total phosphorus and DO. The accuracy of the ELM model was compared with the standard MLPNN using several error statistics such as root mean squared error, mean absolute error, the coefficient of correlation and the Willmott index of agreement. Results showed that the ELM and MLPNN models perform well for the Wubu River, acceptably for the Yipin River and moderately for the Huaxi River, while poor model performance was obtained at the Tributary of Huaxi River. Model performance is negatively correlated with pollution level in each river. The MLPNN model slightly outperforms the ELM model in DO prediction. Overall, it can be concluded that MLPNN and ELM models can be applied for DO prediction in low-impacted rivers, while they may not be appropriate for DO modelling for highly polluted rivers. This is an Open Access article distributed under the terms of the Creative Commons Attribution Licence (CC BY 4.0), which permits copying, adaptation and redistribution, provided the original work is properly cited (http://creativecommons.org/licenses/by/4.0/). doi: 10.2166/wqrj.2019.053 om http://iwaponline.com/wqrj/article-pdf/55/1/106/709271/wqrjc0550106.pdf 2020 Senlin Zhu (corresponding author) State Key Laboratory of Hydrology-Water Resources and Hydraulic Engineering, Nanjing Hydraulic Research Institute, Nanjing 210029, China E-mail: slzhu@nhri.cn Salim Heddam Faculty of Science, Agronomy Department, Hydraulics Division, Laboratory of Research in Biodiversity Interaction Ecosystem and Biotechnology, University 20 Août 1955, Route El Hadaik, BP 26, Skikda, Algeria This article has been made Open Access thanks to the kind support of CAWQ/ACQE (https://www. cawq.ca).


INTRODUCTION
Due to the complexity of factors impacting DO levels, it is important to understand how these factors determine the level of oxygen available for living organisms, and prediction of DO concentrations is crucial for aquatic managers responsible for the maintenance of ecosystem health (Meding & Jackson ). Mathematical models provide useful tools to predict the spatio-temporal dynamics of DO in water bodies. Many sophisticated deterministic models have been developed in the past years to predict DO levels in rivers, such as QUAL2E, QUAL2 K and WASP (Cox ; Kannel et al. ). These mechanistic computer softwares can simulate processes which impact DO levels, such as hydrodynamics, dispersion and pollutant kinetics in the natural environment. These models have been widely used in different river systems, such as applications of the QUAL2E model in the Corumbataí River (Palmieri & Carvalho ) and Putzu River (Yang et al. ), and DO simulations with the QUAL2 K model (Du et al. ; Cho & Ha ). Generally, many input data are needed to run these models, such as topography, flow discharge and water level, water quality concentrations and meteorological data. The highly intensive data need sometimes limit the applications of these mechanistic models.
Except for the mechanistic models, there has been a widespread interest in the application of artificial intelligence techniques for DO modelling in water bodies, such as the artificial neural network (ANN)-based approach (Soyupak et

Study area and data set
The Yangtze River is the largest river in China and the third largest in the world. The TGR is located at the end of the upper Yangtze River. It is one of the largest man-made reservoirs in the world with a surface area of 1,084 km 2 , a storage capacity of 39.3 billion m 3 and a watershed area larger than 1 million km 2 (Wang et al. ). Four urban tributaries in Chongqing City, located in the terminal of the backwater zone of the TGR, were studied in this paper. Observed data from ten monitoring stations in these four rivers were used in the water quality analysis (Table 1).
The Wubu River is listed as a water source protection area for centralized drinking water supply, thus its water   Table 2. According to the statistical indices reported in Table 2, the data are not homogenous and there is a large variability trend among the water quality variables. Except for water temperature and pH, for which the variability is not noticeable, it is clear from Generally, DO in all rivers negatively correlated with TE, and with the increase of pollution level, the coefficient of correlation (R) decreased (Table 2). Additionally, DO presented poor correlations with other water quality parameters ( Table 2). All the input water quality variables and DO were standardized using the Z-score method (Olden & Jackson ): where Z n is the normalized value of the observation n, x n is the measured value of the observation n, x m and σ x are the mean value and standard deviation of the variable x. In the present study, we evaluated several combinations of the water quality variables based on the correlations between water quality variables and DO and, in total, nine scenarios were compared (Table 3).
where x i is the input variable, w ij is the weight between the input i and the hidden neuron j, δ j is the bias of the hidden neuron j, f 1 the activation sigmoid function, represented by Equation (3), w jk is the weight of connection of neuron j in the hidden layer to unique neuron k in the output layer; δ 0 is the bias of the output neuron k, and finally f 2 is a linear activation function for the neuron in the output layer: Extreme learning machines ( Let us consider two set of variables, dependent y i and independent x i which comprises a training data set {x i , y i }, i ¼ 1, …, N, in which, x i ε R d and y i ε R c , the ELM with L hidden neuron can be expressed as: where β j is a weight vector connecting the jth hidden neuron and the output neurons, and h( where H þ is the Moore-Penrose generalized inverse of matrix H (Huang et al. a, b). For applying the ELM models, we used the Matlab codes available at http://www.ntu.edu.sg/ home/egbhuang/elm_codes.html.

Performance assessment of the models
In this study, model performance was evaluated using the following statistical indices metrics: the coefficient of correlation (R), the Willmott index of agreement (d ), the root mean squared error (RMSE) and the mean absolute error (MAE): (P i À P m ) 2 s 2 6 6 6 6 4 3 7 7 7 7 5 (6) where N is the number of data points, O i is the measured and P i is the corresponding model prediction of dissolved   Although the performances of the MLPNN models in the training phase were very satisfactory, the models performed poorly during the validation phase. As is shown in Table 7