Rainfall is illustrative of a nonlinear process and its
forecasting with relatively limited data makes it very complex
phenomenon. To overcome part of this complexity in
modeling the rainfall, an ANN procedure (MFNN) integrated
with an evolutionary optimization method such as
GA was applied. The model input(s), the most effective
neighborhood stations and their rainfall lag times, will be
derived from optimized ANNs by GA to achieve an efficient
input configuration for minimum error. Basically,
by detecting effective inputs, the best input combination(s)
for intelligence prediction will appear. Delineation of optimum
lag time(s) at a particular location, investigation of
the extend to which inclusion of spatial information might
improve network performance, and assessment of performance
indicator with regard to application of cumulative
versus discrete data were addressed in this paper as well.
Seven cases of simulation models have been selected to
illustrate the performance of the proposed technique. These
runs differ from one another mainly on the genetic algorithm
parameters (i.e. Pm and Pc), surrounding rain gauges
used, number of time lags associated with each rain gauge,
temporal resolution of rainfall events, and data type
(cumulative versus discrete). The probability of mutation
(Pm) varied, traditionally, between 1.2% and 1.4%, while
crossover probability (Pc) varies between 92% and 96%.
Evaluation of network performance was achieved through
computation of three good fitness criteria, namely, mean
square error (MSE), normal mean square error (NMSE)
and coefficient of determination (R2). Table 1 summarizes
various experimental runs based on GA parameters. For
the first five models (data in discrete form), two previous
lags of all surrounding recording rain gauges are used to
forecast current rainfall depth at the rain gauge of interest
(Station No. 7261). In models 6 and 7, only rainfall information
from the target station, with several subsequent
lags, were used as input data. According to Bowden et al.
(2005a), mathematical sensitivity analysis is implemented
to validate selected input parameters with GA.
Table 2 presents the result of sensitivity analysis for the
first five models using rainfall data from all rain gauges
during the training phase as input parameter. Zeros in
Table 2 indicate gauges in which the GA algorithm disqualified
them either by generating an optimized topology or
left them out based on error minimization as sources of
input data because of their ineffectiveness to predict current
rainfall of the target station. The sensitivity coefficients
indicated by ‘‘-’’ in the 5th model are those eliminated manually
based on the results obtained in previous model
(model four) and they include those gauges with coefficients
of small values (indication of ineffectiveness) in order
to reduce the execution time of the model.
Table 2 also reveals that while the rainfall data from the
first lag appeared to be relatively sensitive to the target rain
gauge, the data from the second lag is almost very ineffective
in predicting the target rainfall. Another important
result drawn from Table 2 is that when the time resolution
gradually increased in the simulation model, the most effective
stations in predicting the target rainfall would appear.
Besides the target station itself, stations 7259, 7265, 7285
and 7299 have the major contribution on forecasting and
they are not necessarily the closest ones to target station,