Modeling total suspended solids using artificial intelligence methods based on pre processing algorithms

Water quality is one of the most important factors in healthy living and human life. In this sense, some of the water quality parameters should be controlled in maintaining the human health and welfare. In today’s industrial world, most of the global natural water sources, including those in Iran, contain impurities such as the TDS. Numerous factors that are includes include cations such as sodium ion (Na+), potassium ion (K+), calcium ion (Ca+2), and magnesium ion (Mg+2) and anions such as chloride ion (Cl-) and bicarbonate ion (HCO3- ) with sulphate ion (SO42-) affect the concentration of these parameters in natural water systems. The total dissolved solids (TDS) is one of its most important factors; Many water resources development programs will be implemented to identify these factors. Accurate prediction of water quality parameters is a basic need for water quality management, human health, public consumption and household consumption. In the last decades, artificial intelligence (AI) techniques have become viable and popular due to their advantages, and have been widely developed in solving a variety of environmental engineering and water quality engineering problems.

For the estimation of water quality parameters (WQPs), Singh et al. (2011) utilized the clustering method, or support vector clustering (SVC), to optimize surface water quality monitoring in the city of Lucknow, India. The overall view of the water quality index of their study area revealed that most of the study area come under highly to very highly polluted zones. Tan et al. (2012) predicted phosphorus values in China with the least square support vector regression (LSSVR) method. They compared the efficiency of the LSSVR method with neural networks of the radial basis function (RBF) and back-propagation (BP). Experimental results showed that the small sample case with noise, LSSVM method was better than multi-layer BP and RBF neural network and is able to better meet the requirements of water quality prediction. Liu et al. (2013) addressed WQPs prediction in aquaculture employing the GP and real-value genetic algorithm-SVM (RGA-SVR). They used the GA to modify the coefficients of the SVR method. The results showed the superiority of the RGA-SVR algorithm over other methods based on the root mean square error (RMSE) and mean absolute percentage error (MAPE). Ghavidel and Montaseri (2014) employed ANN, GEP, and ANFIS with grid partition as well as ANFIS with subtractive clustering (ANFIS-SC) to predict TDS values of the Zarinehroud basin, Iran. A comparison was made between the above AI approaches, and the results demonstrated the superiority of GEP over the other intelligent models. Abyane (2014) compared artificial neural network (ANN) with multivariate linear regression (MLR) for prediction of BOD and COD in the wastewater treatment plant. In their study, ANN could predict BOD and COD parameters with higher precision than MLR.

Due to complex characteristics of time series WQPs, a standalone model can hardly satisfy the estimation accuracy requirements. Therefore, the hybrid models combined with different single models will be an effective way to improve the WQPs estimation accuracy. This study proposes a new and accurate hybrid model for predicting WQP (i.e., TDS) using ions at Varand and Garmrood, two hydrometric stations of Tajan basin, Iran. The proposed WQP estimating framework was developed based on the combination of a data pre-processing algorithms (i.e., EEMD) with two AI-based models that was not addressed by the literature related to the WQPs modelling. Acceptance and reliability of proposed hybridized and standalone models (e.g., artificial neural networks (ANN), EEMD-ANN, support vector machine (SVM) and EEMD-SVM) using five performance criteria and visual diagrams were evaluated. Comparison of results between independent and hybrid models showed that EEMD data pre-processing algorithm can increase the performance of the hybrid SVM model for estimating the TDS quality parameter in both training and testing stages at both considered hydrometric stations. For example, the EEMD-SVM model with RMSE = 20.23 for the training phase and RMSE = 27.29 for the test phase at Varand station and RMSE = 45.26 for the training phase and RMSE = 40.06 for the test phase at Garmrood station has performed better than other hybrid and standalone models. In general, the proposed hybridized model of support vector machines based on EEMD data pre-processing algorithm can be proposed as a superior model to decision makers for planning and management in the field of river water quality detection and determination.