Geoscience Reference
In-Depth Information
data were used to test model performance as learning progressed. This study
examines the effect of sample size and network architecture on the accuracy
of neural network estimates for the known posterior probabilities. Neural
network toolbox in MATLAB 6.1
8
Release 12, software is used to solve the
developed model with four input variables and one output variables. The
length of data for both input and output values is from 1 to 300. A better
prediction is given by the three layers ANN model with 25 neurons in each
hidden layer. This network is used to compute water release from Vaigai
reservoir. It is also found that BPLM is quicker in the search for a good
result.
Third, the special type of neural network called recurrent neural network
(RNN) shown in Fig. 3 is also used to fit the flow data for reservoir operation
problem. The release produced by RNN is compared with that of other
neural network models for reservoir operation. The results are compared
in terms of meeting the demand of the basin. The results are shown in
Fig. 4. Comparison of performance by different methods of NN is exhibited
in the plot.
Validation of the model
: Once the training (optimization of weights)
phase has been completed the performance of the trained network needs
to be validated on an independent data set using the criteria chosen. The
data of water year 1994-1997 are used for validating the developed ANN
Fig. 3.
Architecture of recurrent neural network.
