Author(s): P. Varoonchotikul; M. J. Hall; A. W. Minns
Linked Author(s):
Keywords: Flood forecasting; Feed forward neural network; Jordan recurrent neural network; Recurrent scheme; Learning rate; Oscillation; Average mutual information
Abstract: There is increasing interest in the application of Feed Forward Neural Networks (FFNs) for flood forecasting, modelling of the relationship between rainfall and streamflow. The exogenous inputs of FFNs require some pre-processing and can consume large amounts of storage before outputs are produced. Particular advantage of FNNs are their numerical stability, which is not sensitive to learning rate, the simplicity of the basic network and the relative efficiency with which the training converges. A set of exogenous rainfall inputs can be derived by using the Average Mutual Information. Instead of employing the FNN, another alternative is to apply the recurrent scheme known as Jordan Recurrent Neural Network (JRN), which keeps a bank of previous outputs in its interanl memory and feeds back to its input layer. In this case, required storage can be reduced. Both forms of network were applied to forecast high flows of the Thrushel river, located in the south west of England, up to three time steps ahead. Results were found to be comparable. The FFN is straightforward scheme while the JRN is more difficult because of the need to select learning rate to damp out its oscillation. Lower learning rates are needed to stabilise the JRN-network. It is found that JRN requires about 40% less epochs than that of FFN.
Year: 2002