IEEE Computational Intelligence Magazine - August 2021 - 63

training samples does not always enhance
the prediction performance and that a partition
ratio of 4:1 may be a suitable choice.
In addition, for a partition ratio of 1:1,
SVR-r and SVR-l yielded superior results
on both the long-term and short-term
wind speed data compared to those
obtained using the other algorithms. However,
when the partition ratio was 4:1 or 9:1, DNR-SMS
achieved the highest prediction performance among the tested
algorithms. These findings imply that the training processes
for SVR-r and SVR-l require less data than those for
other algorithms. Thus, when the length of the wind speed
time series is insufficient, SVR-r and SVR-l can provide
more satisfactory results. However, if an adequate amount of
wind speed time series data is available, DNR can demonstrate
highly competitive performance.
To identify the significant differences between the
compared algorithms, the p-values of the Wilcoxon rank
sum test are also provided in Tables IV and V. Although
To fairly evaluate the performance of each algorithm,
several evaluation methods were adopted in this
study, such as the evaluation indicators, nonparametric
statistical test, and relative graphs.
SVR-r and SVR-l outperform DNR-SMS for the ratio of
1:1, the corresponding p-values are larger than 5%, indicating
that the performance is not significantly higher than
that of DNR-SMS. However, for ratios of 4:1 and 9:1,
DNR-SMS can produce significantly superior results compared
to the other algorithms, such as the ENN, the MLP,
the various SVR methods and the LSTM network, based
on the corresponding p-values. Moreover, compared with
the prediction performance of the original DNM-BP algorithm
and EDNM, that of the proposed DNR model is
significantly higher, which implies that considering the
strength of each dendritic branch individually can enhance
10
8
6
4
2
Actual Value
Predictive Value
Training and Prediction (Ten-Min 1:1)
Training
Prediction
1 234567
Data
Fit
Y = T
1
2
3
4
5
6
7
200
400
600
(a)
10
8
6
4
2
Actual Value
Predictive Value
800
Training and Prediction (Ten-Min 4:1)
Training
1,000
1,200
Target
123456
Prediction
Data
Fit
Y = T
Data
Fit
Y = T
1
2
3
4
5
6
200
400
600
(b)
10
8
6
4
2
Actual Value
Predictive Value
Training
800
Training and Prediction (Ten-Min 9:1)
Prediction
1,000
1,200
23 45 6
Target
Data
Fit
Y = T
Data
Fit
Y = T
2
3
4
5
6
200
400
600
(c)
800
1,000
1,200
Target
FIGURE 6 Training and prediction results and predicted correlation coefficients of DNR-SMS for the wind speed time series with an interval of
ten minutes and partition ratios of (a) 1:1, (b) 4:1, and (c) 9:1.
AUGUST 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 63
The Values of InFlow (M/S)
The Values of InFlow (M/S)
The Values of InFlow (M/S)
Output ~= 0.84 ∗ Target + 0.66
Output ~= 0.85 ∗ Target + 0.58
Output ~= 0.74 ∗ Target + 1
IEEE Computational Intelligence Magazine - August 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - August 2021
