An artificial intelligence predictive model for turbidity current concentration in reservoirs

Generally, reservoirs are designed to achieve some goals such as providing urban water, irrigation, power generation, flood control... Because of settlement of all or a part of sediment in reservoirs, they lose a considerable percentage of their effective capacity. Turbidity currents, which are the gravity currents containing suspended sediment, are the main processes in transportation and sedimentation in reservoir. Owing to importance of water management in the world, specifically in dam reservoir, water engineers get to know turbidity current structure and behavior recently using different methods. The aim of this research is utilization of Artificial Neural Networks (ANN) and experimental method to predict concentration profiles of turbidity current in reservoirs.

The experiments were done in a rectangular channel with abrupt bed slope change in the hydraulic laboratory at Tarbiat Modares University-Tehran. A Vectrino II was used to obtain a signal amplitude by which the sediment concentration profiles were determined based on acoustic back scattering techniques. These experimental data were considered for training of the ANN.

An ANN consists of a multi-processor computer system with simple processing elements, simple scalar messages, a high degree of inter-connection and adoptive interaction between elements. In the present study the Multi-Layer Feed Forward (MLFF), the most common type of a neural network, was considered, which uses the back-error propagation technique for training. Generally the ANN consists of an input layer which receives the input data, hidden layers as intermediary layers, an output layer for providing the output and connection weights to store knowledge. The inputs were depth, distance from the inlet and the target output was corresponding concentration. The ANN was trained using part of experimental data obtained from our laboratory model.  Then the performance of ANN was evaluated by using rest of data which had not been used in the training process. Results showed that the average error between the experimental values of concentration and neural network outputs is about 5.92%. It can be concluded that the outputs of designed ANN are appropriately close to the experimental data which means that the ANN is trained, well.