Akhilesh KumarSingh, Praveen Vikram2017-05-222017-05-222014-09http://krishikosh.egranth.ac.in/handle/1/5810012053Artificial Neural Network (ANN) has been widely used in the field of rainfall-runoff modeling. ANN models based on Multilayer perceptron and single multiplicative neuron were developed and verified for Khunt watershed in Almora, India to predict daily runoff. The daily data of rainfall and runoff of Khunt watershed of active monsoon period (18th June to 30th September) during years 2005-2009 were used for the training of both the models, whereas the data of years 2010 were used for validation and 2011 for the verification. The MLP-ANN was developed for various preprocessing techniques and the effect of preprocessing techniques on the performance of the model was considered. It was found that standardization technique provided the best results with smallest network topology for multilayer perceptron based ANN while using tan sigmoid function in the hidden layer and linear activation function in the output layer. It was also observed during the development of the models the multiple random generations of weights and biases are essential to reach the global minima along with practically feasible values of runoff. Model MLP4 having network structure as 6-5-1 with daily input series exhibits best results. Single multiplicative neural network is different from multilayer perceptron neural network as SMN-ANN is composed of multiplicative single neuron instead of multiple additive neurons. Single neuron model using multiplicative function strengthens non-linearity characteristic of the model and involves less parameters than those employed in MLP networks. SMN-ANN model was calibrated for different learning rates for daily input series and performed best for 0.5 learning rate. Qualitative performance of the model was assessed by the visual observation, whereas, quantitative performance was verified by estimating the values of various statistical indices such as coefficient of efficiency (CE), coefficient of determination (R2), index of agreement (d), normalized root mean square error (NRMSE), Integral square error (ISE), standardized mean absolute percentage error (SMAPE), peak difference percentage error (PDPE) and volumetric error (VE). SMN based ANN model performed very satisfactory for runoff prediction as the value of coefficient of efficiency (CE) was found to be 0.9511 while in case of MLP based ANN model it was found to be 0.8185 respectively. The value integral square error (ISE) was found to be less than 0.03 and the volumetric error was found within 10% for the whole data series. SMN-ANN model performed either at par or better than MLP-ANN model when applied on different data sets. At the same time, the SMN-ANN model is more user friendly as in its development the tedious and cumbersome process involved for the selection of appropriate network architecture is completely eliminated. Hence the SMN-ANN model provides an effective alternative to MLP-ANN model for rainfall–runoff modeling.ennullThesis