Surface water electrical conductivity and bicarbonate ion determination using a smart hybridization of optimal Boruta package with Elman recurrent neural network

Surface water electrical conductivity and bicarbonate ion determination using a smart hybridization of optimal Boruta package with Elman recurrent neural network

Water quality (WQ) monitoring in the surface water resources is a crucial concern as it has an impact on human health and ecosystem equilibrium. An accurate simulation of river WQ indicators as a function of available variables with data mining techniques is not much explored by the researchers. In this study, two smart dual-preprocessing hybridized with Elman recurrent neural network (ERNN) were developed for an accurate simulation of two surface WQ indices including electrical conductivity (EC) and bicarbonate (HCO3-) in the Northern zones of Karun River, Iran. Nine input features including sodium adsorption ratio (SAR), magnesium (Mg2+), Ca+2 (calcium) sum of the anions (Sum.A), SO42-(sulphate) Cl-(chloride), pH, discharge (Q), and Na+ (sodium) were employed to simulate the EC and HCO3- in surface water Two Boruta-data filtering strategies including Boruta-XGBoost (BXGB) and Boruta-Extra Tree (BET) were utilized to extract the most important WQ indicators in available features. The best subset regression (BSR) scheme was utilized to optimize the input combinations of available features to predict EC and HCO3-. Three superior scenarios (C1, C2, and C3) were considered for each target variable to feed the machine learning (ML) models. The dual pre-processing was hybridized with ERNN to compare the results of advanced ML approaches i.e., long short-term memory (LSTM), Kernel ridge repression (KRR), and Elastic net regression (ELNET). Eight hybrid modeling paradigms (BXGB-ERNN, BET-ERNN, BXGB-KRR, BET-KRR, BXGB-ELNET, BET-ELNET, BXGB-LSTM, and BET-LSTM) were evaluated in terms of performance using statistical criterions such as correlation coefficient (R), root mean square error (RMSE), and Kling-Gupta efficiency (KGE). Results revealed that the BET-ERNN-C1 hybrid model outperformed the other models for HCO3- (R = 0.9847, RMSE = 0.0793 mEq/L, and KGE = 0.9782) and EC simulation (R = 0.9543, RMSE = 51.0260 μS/cm, and KGE = 0.9406) in terms of performance efficiency. Results indicated that the BET-ERNN-C1 model resulted in minimum values of IQR for EC (0. 7.455) and HCO3- (0.1819). Overall, the result of modeling showed that the ERNN-C1 model had a superior performance in simulating the WQ indices followed by BXGB-ERNN-C3 and BXGB-KRR approaches. Results of this study suggested that the predictions of WQ indicators using hybrid models can be used to assess the acceptable levels of EC and HCO3- in surface water using appropriate input variables efficiently and reliably.

Files associated with this item cannot be displayed due to copyright restrictions.