Estimation of Daily Suspended Sediment Concentration using Artificial Neural Networks and Data Clustering by Self-Organizing

Nowadays, the accurate estimation of rivers suspended sediment load (SSL), from various aspects, such as water resources engineering, environmental issues, water quality and so on is important. In this regard, because of various roles of fixed and dynamic variables of watersheds, the watershed hydrological models have not showen a proper efficiency in statimation of SSL. Also, the most SSL studies are based on only flow discharge variable whereas the results of the present study have proved that the efficiency of these modeles is very poor. On the other hands, the parameters such as rainfall type, year seasons and flow hydrograph shape have important role in watershed sediment yield that were ignored in the most SSL simulations. In the present study, multi layers perceptron neural network and hydro-meteorological data (daily flow discharge, suspended sediment concentration, daily rainfall and temperature) of Karaj dam watershed in a 30-year period (1981 to 2011) were used to estimate daily suspended sediment concentration of Sierra station. Due to the role of seasonal changes and flow conditions in sediment yield and sediment transport of the watershed, based on rainfall regime, hydrograph condition and runoff type, the data used in this study were first seperated into 5 groups and then for each group, a separate model was designed. In order to increase the generalization ability of the neural network models, self-organizing map (SOM) and Silhouette coefficient were used for data clustering and determination of the optimal number of clusters respectively. The research results showed that the use of daily precipitation and temperature variables along with flow discharge and data separating based on watershed time and hydro climatic conditions has had an important role in increasing the accurate estimation of the river sediment. In this regard, among the models, the maximum calculated error is when only a single model is used for all year seasons. The research results and its used structure can be used as a pattern to estimate and to forecast many environmental variables of watersheds.