Evaluating the Efficiency of Skew Correction Factors in Sediment Rating Curve and Comparison with Intelligent Models (Case Study: Jelogir Station, Khuzestan - Karkheh Catchment)

Estimation of the sediment load in rivers is one of the important issues in studies related to water quality and transport of pollutants, construction and operation of hydraulic structures, maintenance of reservoirs, water transmission networks, and water resources management. An accurate understanding of the sedimentation of a watershed can provide a correct understanding of soil erosion and its consequences. Since sediment changes in the river are often a function of flow discharge changes; therefore, methods of measuring suspended sediment load based on the suspended sediment concentration and flow discharge will be useful in estimating the amount of sediment load. The sediment rating curve is one of the methods that is based on flow discharge and sediment discharge and expresses the relationship between these two parameters in the form of power regression (Eq 1). (1) where Qs  is the suspended sediment discharge (in tons per day), Qw is the flow discharge (in cubic meters per second), and a and b are the coefficients of the equation . Rating curves can be drawn in different ways according to the way of data separation. Among these methods, we can refer to one-line, multi-line, mean of categories, seasonal, monthly, annual models, etc. The presence of bias in the sediment discharge relationship makes this relationship unable to show the exact sediment concentration in different flow discharges. This bias causes the amount of sediment to be underestimated. Various researchers have proposed some statistical correction factors to achieve the minimum error, which are applied in the sediment rating equation. In this research, in order to increase the accuracy of sediment estimation by using a sediment rating curve, at first, different types of rating curves were drawn for the station and, finally, correction factors consisting of QMLE, Smearing, MVUE, and (Beta) β were applied for the selected curve. Also, an attempt was made to separate the data into three categories of dry, normal and wet by using the percentage of normal precipitation and to draw the sediment rating curve for each. At the end, the results obtained from the statistical model (SRC) were compared with artificial intelligence models including two models of multilayer perceptron (MLP) and radial basis set (RBF) neural networks.

In this research, the flow and sediment discharge data from 1350 to 1397 for the Jelogir station in Khuzestan province located on the main Karkhe River were prepared from the Khuzestan Regional Water Organization. Sediment rating curve models, including common linear curve (USBR), mean of categories, monthly, seasonal, bilinear, trilinear, dry, normal and wet models were drawn for the station. Then, for the drawn curves, evaluation criteria including RMSE, ME and P were checked and, finally, by ranking these criteria, the curve with the least error was selected. In determining the rank of each model, the values of the evaluation indices were compared with each other. In this way, the closest P and ME index value to 1 and the closest RMSE index value to zero, which indicates the least difference between the estimated and observed sediment values, was assigned the first rank. In order to investigate the effect of skew correction coefficients on the accuracy of sediment rating curves, coefficients including MVUE, FAO, QMLE and Smearing were applied on the rating curve which was selected as the optimal model in the previous step. The data were processed using neural network models. For this purpose, different structures of neural networks with different layers, neurons and functions were investigated through trial and error.

According to the obtained results, the mean categories method has the highest correlation coefficient (0.85). The RMSE in rainy and flooding months (April and March) and also in high flow discharge rates (in bilinear, and trilinear models, at flow discharge greater than 201 and 114 cubic meters per second, respectively), has allocated the largest amount. The lowest value of RMSE is related to the months of August and September, which is reasonable due to the lack of rainfall and flooding in these months and as a result of low erosion of sediments. According to the ranking values, the periods of low rainfall, including summer and July, August and September are in the first ranks, and as a result, the sediment rating curve has more accuracy in estimating sediments. Finally, the rating curve of August, which has the lowest total ranking value, was chosen as the optimal curve. According to the ranking of the correction coefficients, it can be seen that the sediment rating curve without applying the correction coefficients (the highest rank) has the highest amount of error and by applying the coefficients, the error of sediment flow estimation can be reduced. Finally, MVUE with the lowest total ranking was chosen as the optimal correction coefficient, and by applying it, the accuracy of the model in estimating the sediment discharge increases. In the neural network model, Lunberg-Marquardt optimization algorithm was used and the number of hidden layer neurons in the best MLP and RBF structure was obtained as 5 and 6, respectively. Also, the activator function in the hidden layer in MLP was selected as sigmoid tangent and Gaussian function in RBF. The results show that by using neural networks of multilayer perceptron, it is possible to predict the amount of suspended sediment with higher accuracy, and the accuracy of the results obtained from the artificial neural network method is far higher than the accuracy of the rating curve method with and without data classification. According to the results, the MLP model has shown a lower error value than the RBF radial base model.

In this article, in order to estimate the suspended sediment in the Jelogir station, the data were separated into different forms and the sediment rating curves were drawn into linear curve (USBR), mean of categories, monthly, seasonal, dry, normal, wet, bilinear, and trilinear types. The obtained results showed that the accuracy of the relationship obtained for the classification of data based on August (R2= 0.785) and the total rating of 9 (the lowest value) was more than the other models. And at high flow discharge, the accuracy of the models decreases. It was found that the correction coefficients are effective in increasing the accuracy of the models, and the lowest amount of error for the optimal model is obtained by using MVUE. Comparing the results of statistical methods and neural networks showed that neural network models are more accurate in estimating daily sediment. The better performance of artificial neural networks compared to statistical methods can be expressed in the nonlinear approximation capability of neural networks.