Modeling of suspended sediment concentrations by artificial neural network and adaptive neuro fuzzy interference system method–study of five largest basins in Eastern Algeria

Abstract

Abstract Prediction of suspended sediment concentrations (SSC) in arid and semi-arid areas has aroused increasing interest in recent years because of its primary role in water resources planning and management. Today, given its simplicity and reliability, SSC modeling by artificial neural networks (ANNs) and adaptive neuro-fuzzy inference system (ANFIS) are the most developed and widely used methods. The main aim of this study is suspended sediment concentrations modeling using ANN and ANFIS methods at the five largest basins in eastern Algeria: the Constantinois Coastal, Highlands, Kébir-Rhumel, Seybouse, and Soummam basin, which are characterized by high water erosion and a lack of SSC measurements. An application was given for historical time series: liquid flows Ql and solid flows Qs as inputs, and daily SSC as outputs, for the 14 hydrometric stations controlling the entire area. The best models were achieved using a multi-layer perceptron (MLP) feed forward networks (FFN) trained with a Levenberg-Marquardt (LM) algorithm for ANN modeling and a first-order Takagi-Sugeno-Kang (TSK) FFN with a hybrid learning method for ANFIS modeling. The reliability of the created models was evaluated using five validation criteria: determination coefficient R2, Nash-Sutcliffe coefficient NSE, mean square error MSE, root-mean-square error RMSE, and the mean absolute error MAE. The ANN and ANFIS models showed high accuracy, confirmed by excellent R2 values ranging from 0.77 to 0.98. The NSE ranged from 0.67 to 0.97. The error values were very good, the MAE varies from 0.004 g/L to 0.028 g/L for both models. The comparison of the ANN and ANFIS models revealed that ANN models slightly outperformed the ANFISs; both of them had high accuracy in SSC prediction.