Artificial Neural Networks for Estimating Daily Total Suspended Sediment in Natural Streams

Abstract

Estimates of sediment loads in natural streams are required for a wide spectrum of water resources engineering problems from optimal reservoir design to water quality in lakes. Suspended sediment constitutes 75-95% of the total load. The nonlinear problem of suspended sediment estimation requires a nonlinear model. An artificial neural network (ANN) model has been developed to predict daily total suspended sediment (TSS) in rivers. The model is constructed as a three-layer feedforward network using the back-propagation algorithm as a training tool. The model predicts TSS rates using precipitation (P) data as input. For network training and testing 240 sets of data sets were used. The model successfully predicted daily TSS loads using the present and past 4 days precipitation data in the input vector with R2 = 0.91 and MAE = 34.22 mg/L. The performance of the model was also tested against the most recently developed non-linear black box model based upon two-dimensional unit sediment graph theory (2D-USGT). The comparison of results revealed that the ANN has a significantly better performance than the 2D-USGT. Investigation results revealed that the ANN model requires a period of more than 75 d of measured P-TSS data for training the model for satisfactory TSS estimation. The statistical parameter range (xmin - xmax) plays a major role for optimal partitioning of data into training and testing sets. Both sets should have comparable values for the range parameter.