Suspended sediment discharge modeling during flood events using two different artificial neural network algorithms

This paper presents modeling of artificial neural network (ANN) to forecast the suspended sediment discharges (SSD) during flood events in two different catchments in the Seybouse basin, northeastern Algeria. This study was carried out on hourly SSD and water discharge data during flood events from a period of 31 years in the Ressoul catchment and of 28 years in the Mellah catchment. The ANNs were trained according to two different algorithms: the Levenberg-Marquardt algorithm (LM) and the Quasi-Newton algorithm (BFGS). Seven input combinations were trained for the SSD prediction. The performance results indicated that both algorithms provided satisfactory simulations according to the determination coefficient (R-2) and root mean squared error (RMSE) performance criteria, with priority to the BFGS algorithm; the coefficient of determination using the LM algorithm varies between 51.0 and 90.2%, whereas using the BFGS algorithm it varies between 54.3 and 93.5% in both studied catchments, with calculated improvement for all seven developed networks with the best improvement in the Ressoul catchment presented in ANN06 with Delta(R2) 4.23% and Delta(RMSE) 1.74%, and with the best improvement presented in ANN05 with Delta(R2) 6.07% and Delta(RMSE) 0.71% in the Mellah catchment. The analysis showed that the use of Quasi-Newton method performed better than the Levenberg-Marquardt in both studied areas.