Rainfall forecasting using ANNs has been the focus herein. Three types of ANNs suitable for
this task were identified, developed, and compared; these networks were
l multilayer feedforward neural network (MLFN),
l Elman partial recurrent neural network (Elman), and
l time delay neural network (TDNN).
All the above alternative networks could make reasonable forecast of rainfall one time step (15
minutes) ahead for 16 gauges concurrently.
In addition, the following points were observed.
For each type of network, there existed an optimal complexity, which was a function of
the number of hidden nodes and the lag of the network.
All three networks had comparable performance when they were developed and trained
to reach their optimal complexities.
Networks with lower lag tended to outperform the ones with higher lag. This indicates
that the 15min. rainfall time series have very short term memory characteristics.