Simulation of Rainfall-Runoff Using Artificial Neural Networks(ANNs) (Case Study: Faridan Watershed)


University of Isfahan


Extended Abstract
1- Introduction
Due to the various flood risk and risks arising from the event that human society and its structures are threatened, The process of rainfall - runoff and hydrograph flow from rivers and basins is of special importance. On the other hand, simulated rainfall - runoff is one of the basic needs of flood warning systems. Because , experimental models such as natural or synthetic unit hydrograph are unable to simulation the nonlinear flood. Therefore, the use of models like artificial neural network for nonlinear behavior of the basin finds.
2- Methodology
In this research, seven stations in the daily rainfall data synoptic - hydrometer has been used in Faridan watershed. The input parameters to the three-part was divided: training, holdout and test . 70% data for training, 20 percent for holdout and 10 percent of the remaining data for test was used inthe artificial neural network. For the calculation and construction of artificial neural networks used MATLAB software is in branch of neural network.
3– Discussion
To achieve the optimal structure of neural networks, six scenarios were designed and was evaluated number 2 to 40 neurons in the hidden layer . Ultimately determined that Scenario 6 with 32 neurons in the hidden layer, has the highest correlation and lowest RMSE error, that shown high correlation and significance between observed and predicted value simulated rainfall - runoff in the Faridan watershed.
4– Conclusion
Results showed that a multi-layer perceptron artificial neural network with error back propagation algorithm and 32 neurons in the hidden layer can the process of rainfall - runoff simulation in Faridan watershed and reducing or increasing the neurons in this layer, the simulation will be reduced.
- Based on analyzes conducted to compare the mean square error of the model was determined the Neurons in the hidden layer structure(1-32-10) in various stages of training, test and holdout is respectively : 0/23, 0/19 and 0/21.
- Correlation coefficient in the best scenario in the training holdout and test is respectively : 98%, 97% and 96%, that show high correlation between observed and predicted values.