Geology Reference
In-Depth Information
Fig. 6.17 The modelling result of ARMAX model during training phase
6.5.2 In
uence of Data Interval on Data Based Real Time
Flood Forecasting
This analysis was served by implementing a feed forward back propagation (FFBP)
ANN with multiple-layer perceptron (MLP) architecture (one input layer, one
hidden layer and one output layer). This topology has proved its ability in mod-
elling many real-world functional problems. The FFBP was adopted because it is
the most popular ANN training method in water resources literature. In this study,
the FFBPs were trained using the Levenberg
Marquardt (LM) optimization tech-
nique as we found it better than other training algorithms in our study area and data.
In this study, multi-input single-output (MISO) neural network architecture strategy
was adopted since it has been popularly used as a neural network architecture
strategy for multi-step-ahead forecasting [
5
]. We constructed three independent
networks to forecast y
-
(t + 6), respectively. Even though
we require n networks for n-step-ahead forecasting, MISO networks were consid-
ered better than multi-input multi-output (MIMO) scheme, with reduced training
time and increased accuracy [
8
]. For a MISO network, the number of parameters
and weights between hidden and output layers is much less than that of MIMO thus
the complexity is less. The selection of hidden neurons is another tricky part in
ANN modelling as it relates to the complexity of the system. However, in this
study, the Hecht-Nielsen approach (twice the input layer dimension plus one) has
been adopted according to our past experience. Evaluation of the model
(t + 2), y
(t + 4) and y
ˆ
ˆ
ˆ
s predictive
abilities for different data time intervals was compared with statistical terms such as
the root-mean-square error (RMSE) between the observed and forecasted values,
'
