Geoscience Reference
In-Depth Information
Fig. 2.
Architecture of three-layer FFNN.
This calibration process is generally referred to as “training”. The global
error function most commonly used is the quadratic (mean-squared error)
function.
The connection weights are then adjusted using a form of the generalized
delta-learning rule in an attempt to reduce the error function. The amount
bywhicheachconnectionweightisadjusteddependsonthelearningrate
(
η
), the momentum value (
µ
), the epoch size (
∈
), the derivative of the
transfer function and the node output. The weight update equation for the
connection weight between nodes
i
and
j
is given in Eq. (3).
s
=1
{
e
∆
w
ji
(
t
)=
η
(
d
j
−
y
j
)
f
(
·
)
y
i
}
+
µ
∆
w
ji
(
t
−
1)
,
(3)
where
w
ji
is the connection weight between nodes
i
and
j
,(
d
j
−
y
j
)isthe
difference between actual and predicted values (error),
f
(
) is the derivative
of the transfer function with respect to its input,
y
i
is the current output of
processing element
i
,and
s
is the training sample presented to the network.
The output from linear programming model; inflow, storage, release and
actual demand are given as input into ANN model of the basin. Monthly
values of inflow, initial storage, demand and time period are the input into
a three-layer neural network and output from this network are monthly
release. The training set consisted of data from 1969 to 1994. The same
·