Search

Rivers are an integral part of the hydrological cycle and are the major geological agents which erode the continents and transport water and sediments to the oceans. Thus rivers act an important link between continents and oceans for the mass balance. Estimating the suspended sediment yield is one of the crucial aims in the field of managing, designing and planning of any river system or reservoir. To determine the suspended sediment yield in a river basin demands more labour or it is more costly when automatic sampling devices are used. The number of variables and the inter-relationship existing among them influence the suspended sediment yield and the nature of these inter-relationships is neither linear nor simple. Unfortunately, it is a difficult task to determine the suspended sediment yield when traditional mathematical models are used as complex variables and processes are involved. The major key factors, such as basin geology (rock type), relief, rainfall, temperature, water discharge and catchment area that affect sediment yield, are used as inputs to develop the model for predicting the suspended sediment yield in the Mahanadi River. In this paper, a multi-objective genetic algorithm for artificial neural network (MOGA-ANN)-based approach is used for predicting the suspended sediment yield. The MOGA assists ANN to minimize the two competing objectives, i.e. mean error and variance simultaneously. Thus in this study, a hybrid artificial intelligence-based method, MOGA-ANN model, is developed using the hydro-geological-climatic factors where all parameters associated with the ANN models are optimized simultaneously using MOGAs to estimate the suspended sediment yield in the Mahanadi River basin. The ANN's parameters are optimized globally by the MOGA to accurate estimation. The study has been carried out to develop MOGA-ANN for estimating the suspended sediment load using 20-year data at the Tikarapara gauging station which is the last downstream station in the Mahanadi River. The MOGA-ANN model provided the root mean square error 0.0281, correlation coefficient (r) 0.966 and efficiency factor 0.919 during the testing phase. The results suggested that the hybrid MOGA-ANN model exhibited satisfactory performance with an accuracy of 96.41%. [ABSTRACT FROM AUTHOR]

Accurate simulation of extreme events of runoff is very important for public safety and hydrological engineering. The simulation of extreme events mainly employs the rainfall-runoff models. This paper tests the potential of soft computing technique-based rainfall-runoff models to simulate the peak runoff events in order to develop an alarming system. This study mainly focusses on hybrid Artificial Neural Networks (ANNs) and Multivariate Adaptive Regression Splines (MARS). The six different regression models used are conventional Artificial Neural Network (ANN), Wavelet Artificial Neural Network (WANN), Bootstrapped Artificial Neural Network (BANN), Wavelet Bootstrapped Artificial Neural Network (WBANN), Multivariate Adaptive Regression Splines (MARS) and Wavelet Multivariate Adaptive Regression Splines (WMARS). The potential inputs were selected based on Auto Correlation Function (ACF) and Cross Correlation Functions (CCF). To implement the methodology, the Jhelum basin in the northern part of India was selected. Based on the results, it was found that all the models except BANN showed overall good performance. The WANN model (NSE = 0.95, RMSE = 1943.15 cusecs, MAPE = 25.5, R = 0.96, DA = 46.8) shows slightly better performance than ANN, WBANN, MARS and WMARS and far better than BANN. The accuracy of the forecasts was checked between WANN and ANN, WBANN and BANN, WMARS and MARS. The results show that decomposition method improves the forecasting accuracy of time series data. Again, the simulation of the peak events was done using the above six models. The efficiency of the models was evaluated on the basis of Normalized Root Mean Square Error (NRMSE). It was found that WANN outperforms the other five models with NRMSE (0.37) for all peak events. All the other models except BANN showed fair results with NRMSE for ANN = 0.66, MARS = 0.68, WBANN = 0.68 and WMARS = 0.53. Thus, it is recommended from the present study that WANN model is more promising for the simulation of peak events in time series data. [ABSTRACT FROM AUTHOR]