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

Dissolved oxygen (DO) is an essential resource of aquatic ecosystems. The DO concentration plays a critical role in regulating various biogeochemical processes and biological communities in rivers. DO concentrations can fluctuate over the day and night in response to climate changes and the respiratory requirements of aquatic plants (Heddam 2014a). Aquatic organisms are sensitive to fluctuations of DO levels in water bodies, especially for DO reductions. Severe oxygen depletion can lead to fish kills (Meding & Jackson 2003; Robarts et al. 2005), and changes in community composition and trophic state (Wetzel 2001; Ruuhijärvi et al. 2010; Branco et al. 2016). The overall DO concentrations in a river are balanced by re-aeration at the water surface, primary production by photosynthesis and consumptions by biochemical oxygen demands in the water column or sediment oxygen demand (Poulson & Sullivan 2010).

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 2003). 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 2003; Kannel et al. 2010). 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 2006) and Putzu River (Yang et al. 2011), and DO simulations with the QUAL2 K model (Du et al. 2008; Cho & Ha 2010). 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 al. 2003; Schmid & Koskiaho 2006; Diamantopoulou et al. 2007; Singh et al. 2009; Chen et al. 2010; Ay & Kisi 2012; Wen et al. 2013; Antanasijević et al. 2013; Heddam 2014a; Keshtegar & Heddam 2017; Csábrági et al. 2017), fuzzy logic models (Altunkaynak et al. 2005; Giusti & Marsili-Libelli 2009; Zounemat-Kermani & Scholz 2014), neurofuzzy models (Heddam 2014b; Najah et al. 2014; Ay & Kisi 2017), support vector machine models (Li et al. 2013; Liu et al. 2013; Ji et al. 2017; Heddam & Kisi 2018) and extreme learning machine (ELM) models (Heddam 2016; Heddam & Kisi 2017). These approaches use available water quality parameters, such as water temperature, pH, electrical conductivity (EC) as inputs. Various ANN models have been developed, and the most reported models are the multilayer perceptron neural networks (MLPNN) (Schmid & Koskiaho 2006; Ay & Kisi 2012; Wen et al. 2013; Keshtegar & Heddam 2017). Recently, Heddam (2016) and Heddam & Kisi (2017) proposed a new approach (ELM) to model DO concentrations in water bodies. The proposed ELM models were applied in eight rivers in the US for estimating DO using four water quality variables as inputs (water temperature, turbidity, pH and EC). In this study, the ELM model was applied in four urban rivers in the Three Gorges Reservoir (TGR) region, China, and model performance was compared with the MLPNN approach. The main objective of this study is to develop models which can be used to inform water quality management for one of the largest reservoirs in the world (TGR).

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 km², a storage capacity of 39.3 billion m³ and a watershed area larger than 1 million km² (Wang et al. 2005). 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 quality conditions are good generally (DO: 6.1–10.0). The Huaxi river basin is mainly an urban watershed where large and medium-sized enterprises and agricultural crop areas are densely distributed. In recent years, the population in the Huaxi basin has increased rapidly. Excessive household and municipal sewage, industrial wastewater and agricultural fertilizers contribute greatly to water pollution in the Huaxi River (DO: 2.9–10.0, COD: 10.0–39.0, TN: 0.69–8.17). The water quality conditions of the tributary of the Huaxi River are the poorest among all the rivers due to excessive household and municipal sewage (DO: 0.4–7.8, COD: 13.1–156.0, TN: 5.03–41.4). The Yipin River was also impacted by anthropogenic activities in recent years, and its water quality conditions are poor as well (DO: 5.2–10.5). Water quality assessment results indicated that Tributary of Huaxi River > Huaxi River > Yipin River > Wubu River for pollution level (Zhu et al. 2018). Water quality data sets used were in the period from 2013 to 2015, with one sampling per month. Water quality parameters include water temperature (TE), pH, DO, permanganate index (PI), NH₃-N, EC, chemical oxygen demand (COD), total nitrogen (TN) and total phosphorus (TP). The statistical summary of the used data sets for all rivers are summarized in 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 Table 2 that the Yangtze River and its tributaries result in a nonhomogeneous data set, especially for EC, TN, TP and PI. The biological variables (DO and COD) are the same, as shown by the mean, max and min values in Table 2. The high values of COD along the Tributary of Huaxi River indicate that the river receives highly non-biodegradable organic matter. The findings suggest the potential and substantial variability in water quality data across the four rivers, especially between the Tributary of Huaxi River and the other three rivers. The data set for the four rivers was divided into two sub-data sets: (i) training subset (70%) and (ii) validation subset (30%).

ANN and their several algorithms have played a critical role in modelling, forecasting and classifying water quality variables. The availability of large data sets and the increase of monitoring stations worldwide will continue to encourage the use of ANN models in several areas of environmental science. One of the principal advantages that have potentially increased the applications of ANN models is that they do not make any assumption about the structure of the data set, and the models are developed based on training algorithm. ANN is a special kind of non-linear model, proposed and inspired by the function of the human brain. The term network refers to a system of interconnected nodes or neurons, similar to biological neurons (Haykin 1999). Generally, the ANN models have three kinds of layers: (i) input layer that contains the water quality variables selected for developing the DO model (reported as x_i), (ii) one or more hidden layers and (iii) the output layer, having only the dependent variable (the DO concentration reported as y). If the interconnection between the neurons is unidirectional, such an ANN model is called feed-forward neural network (FFNN).

As stated earlier in the description of the MLPNN model, during the training process, the matrices of weights and biases must be updated, and consequently, the complexity of the model increases with the increase of the number of neurons in the hidden and input layers. One of the most important algorithms proposed during the last decade for training the MLPNN model is the ELM model introduced by Huang et al. (2006a, 2006b), for single layer feed-forward neural network (SLFN). Contrary to the SLFN in which the weights between the input and the hidden layer (W_ij, Figure 1) are determined iteratively using the BP algorithm, they are randomly initialized and fixed using the ELM. Regarding the weights between the hidden and output layers (W_jk, Figure 1), they are optimized by solving the Moore–Penrose generalized inverse of matrix (Huang et al. 2006a, 2006b).

In the following, we assess the capability and usefulness of the MLPNN and ELM models for predicting DO concentrations at four rivers in China, using eight water quality variables as predictors (TE, pH, PI, EC, TP, NH₃-N, TN and COD). To prevent overfitting, cross-validation is conducted for both models. The estimated values of the performance indices in the training and validation phases are shown in Tables 4–7, respectively. As a preliminary analysis, results obtained show that the ELM and MLPNN models perform well for the Wubu River, acceptably for the Yipin River, moderately for the Huaxi River, and poorly for the Tributary of Huaxi River in DO prediction. This can be explained by considering that: (i) model performance is negatively correlated with pollution level in each river, (ii) the 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. According to the obtained results, it is clearly shown that the MLPNN models performed best at two rivers (Wubu and Tributary of Huaxi River) and the ELM model performed best at the two other rivers (Huaxi and Yipin). Additionally, it was observed that the best accuracy obtained using MLPNN and ELM models differs widely from river to river, and it is sometimes difficult to select the best architecture among the nine input combinations. For example, at Wubu River the best accuracy using the MLPNN model was achieved using MLPNN4, while ELM1 performed better. At Yipin River, MLPNN6 performed better and ELM6 provided the best accuracy. Similarly, at the Huaxi River, the best accuracy was achieved using ELM2 and MLPNN2, respectively. Finally, at the tributary of the Huaxi River, MLPNN8 yielded higher accuracy, while the best accuracy was obtained using the ELM6 model. These statements reveal that, although the models used the same input variables at the four rivers, the effect of each independent water quality variable on DO concentrations differs from one river to another.

Estimated DO concentrations at the Wubu River using the ELM and MLPNN models are shown in Table 4. In the following, a more detailed evaluation for each of the different models is provided and several main points are highlighted. First, in the training phase, the MLPNN models worked very well and provided high accuracy for all input combinations compared to the ELM models. DO concentrations were better fitted to the measured values using MLPNN with R and d ranging from 0.897 to 0.998 and 0.940 to 0.999, respectively, compared to the ELM models which supplied values of R and d ranging from 0.609 to 0.859 and 0.733 to 0.920, respectively. Second, the best accuracy in the training phase was obtained using MLPNN1, while the best accuracy for ELM models was obtained using the ELM9 model. This statement leads to the conclusion that the influence of the different water quality variables on the estimation of DO in rivers during model calibration is not as similar, and the MLPNN models that benefit from training times higher than the ELM models are capable of capturing the non-linear relationships between water quality variables and DO concentrations. By comparing the performances of the MLPNN and ELM models during the validation phase, it is clear that for the MLPNN models, the best results were achieved using the MLPNN4 model (R= 0.937, d= 0.968) with five water quality variables, TE, pH, EC, NH₃-N and TP, as inputs. Significant variability was observed between the nine developed models. R and d ranged from 0.542 to 0.937 and 0.860 to 0.968, with an average of 0.837 and 0.906, respectively. Although MLPNN4 is the best model, MLPNN2, MLPNN4 and MLPNN8 showed relatively similar results and MLPNN4 is slightly better than MLPNN2 and MLPNN8 when focusing only on the R and d values. However, when comparing the three models based on the error indices (RMSE and MAE), MLPNN4 performed much better than the other two models. Regarding the remaining set of models, it is clear from Table 4 that MLPNN1, MLPNN3, MLPNN6 and MLPNN7 showed relatively similar results, with average R values of 0.872 and average d values of 0.923, respectively, and performed better when compared to the MLPNN5 and MLPNN9 models. Finally, the MLPNN5 model performs worst compared to all the other developed models (RMSE= 1.137 mg/L, MAE= 0.942 mg/L, R= 0.542, d= 0.729). However, the MLPNN5 model would have the worst performance for DO estimation because the TE variable is removed from the input variables. MLPNN9 that uses fewer input variables (three inputs), may have the same problem, but suffers less than MLPNN5, because TE is included accompanied by the PI and EC variables.

Estimated DO at Yipin River using the ELM and MLPNN models is shown in Table 5. At first glance, the two models MLPNN and ELM provided low accuracy compared to the performances obtained at the Wubu River. For the MLPNN models, R values range from 0.274 to 0.824 and d values range from 0.587 to 0.906. Similarly, for the ELM models, R values range from 0.270 to 0.828 and d values range from 0.579 to 0.913. In the validation phase, the average R and d values using the MLPNN models were 0.646 and 0.821, respectively, while the average values of the same indices for the ELM models were 0.642 and 0.799, respectively. MLPNN3 gives the worst performances with the highest RMSE and MAE, and lowest R and d, among the nine models. Regarding the ELM models, ELM5 gives the worst performances among the nine ELM models. For overall comparison, ELM6 was the most predictive model and performed slightly better than the MLPNN6 model. For numerical comparison between the best two models, ELM6 yielded a high and best improvement of the MLPNN6, improving its accuracy by increasing the values of the R and d by 0.4% and 0.7%, respectively, and decreasing the values of the RMSE and MAE by 8.53%, and 10.26%, respectively.

The summary of statistical indices of the training and validation data in prediction of DO using ELM and MLPNN models at the Huaxi River are presented in Table 6. According to the obtained results, it is clear that the two models performed less well than for the two previous rivers (Wubu and Yipin), and this leads to an informed judgement: increases in pollution level associated with low level of DO concentrations in the river results in the models being unable to correctly capture the relationship between DO and water quality variables. The R and d values of MLPNN models ranged between 0.107 and 0.687 and 0.580 and 0.855, with average values of 0.532 and 0.771, respectively. Similarly, the R and d values of ELM models ranged between 0.466 and 0.757 and 0.713 and 0.857, with average values of 0.654 and 0.806, respectively. Overall, the ELM models performed better compared to the MLPNN models. At Huaxi River, the ELM2 model provides better DO estimates than the MLPNN2, and the results indicate that DO can be predicted with R and d values equal to 0.757 and 0.857 (RMSE= 0.815 mg/L, MAE= 0.605 mg/L), respectively, which are higher than the values obtained using the MLPNN2 model. According to Legates & McCabe (1999) and Moriasi et al. (2007), values of R greater than 0.70 are considered acceptable; consequently, none of the nine MLPNN models' results are acceptable.

Training and validation results of the ELM and MLPNN models at the tributary of Huaxi River are presented in Table 7. Models' performances were generally unsatisfactory for all nine input combinations according to the guidelines from Legates & McCabe (1999) and Moriasi et al. (2007). 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, the average values of R and d for the MLPNN models during the validation phase were 0.475 and 0.648, respectively, while for the ELM models the values of the two indices dropped to 0.266 and 0.547, with 20% and 10% reduction, respectively. There are several potential factors that could impact the accuracy of the models at this specific site. First, the low correlation coefficient between DO and TE certainly has a great influence. Second, the accuracy of the models decreased dramatically with the increase of the pollution level in the studied rivers, and it is difficult for the models to consider the impact of anthropogenic influences in highly polluted rivers. Measured and predicted DO concentrations with MLPNN and ELM models in the four rivers in the validation phase are presented in Figures 2 and 3.

In this study, ELM and MLPNN models were implemented to predict the DO concentrations using the daily observed river temperature, pH, PI, NH₃-N, EC, COD, TN and TP for four urban rivers in the backwater zone of the TGR. Model results showed that the ELM and MLPNN models perform well for the Wubu River, acceptably for the Yipin River, moderately for the Huaxi River, and poorly for the Tributary of Huaxi River in DO prediction, and model performance is negatively correlated with pollution level in each river. 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 since it is difficult for these models to consider the impact of anthropogenic influences.

This work was jointly funded by the National Key R&D Program of China (2018YFC0407200), China Postdoctoral Science Foundation (2018M640499), and the research project from Nanjing Hydraulic Research Institute (Y118009). The authors acknowledge Chongqing Environment Protection Bureau for providing the water quality data used in this manuscript.