Page 69 - Çevre Şehir İklim İngilizce - Sayı 4

Hacı Abdullah Uçan - Tayfun Dede - Sinan Nacar

Table 6. Performance statistics of the training, verification and test data sets of
ANN-based models developed for the number of flats variable
Training ANN_5 ANN_10 ANN_15 ANN_20 GWO_5 GWO_10 GWO_15 GWO_20
RMSE 74931 64988 64879 59652 81611 87071 78289 87232
(pcs)
MAE
(pcs) 44457 38363 37021 35852 52583 57431 50380 57639
NS 0,44 0,58 0,58 0,64 0,33 0,24 0,38 0,23
Doğru- ANN_5 ANN_10 ANN_15 ANN_20 GWO_5 GWO_10 GWO_15 GWO_20
lama
RMS 44271 44074 43927 34199 28646 28585 37752 28630
(pcs)
MAE 34159 30602 32474 24951 22800 23421 29689 22486
(pcs)
NS 0,53 0,54 0,54 0,72 0,80 0,80 0,66 0,80
Test ANN_5 ANN_10 ANN_15 ANN_20 GWO_5 GWO_10 GWO_15 GWO_20
RMSE 46223 47434 46344 55893 64557 65821 49846 66778
(pcs)

MAE
(pcs) 36249 37890 39994 41621 56257 58059 39870 59211
NS 0,55 0,53 0,55 0,35 0,13 0,09 0,48 0,07

In this table, as an example, the ANN_5 model represents the 5-neuron
ANN model, and GWO_5 represents the 5-neuron ANN model which was
used interactively with GWO technique. When Table 6 is examined, it is seen
that the performance statistics of the models established using the GWO
algorithm are low for training and test data sets compared to the ANN
models developed without using any optimization algorithm, and much
higher for the verification data set compared to the ANN. However, when
a general evaluation is made, it is seen that ANN_GWO does not improve
the performances of ANN models. Of the models developed using different
numbers of neurons, it was revealed that the one with the highest accuracy is
ANN_15 model established by using 15 neurons. Although the performance
values of the ANN_15 model are lower compared to the ANN_20 model
for training and test data sets, it can be said that it has a higher accuracy in
general when the test data set is also taken into account. When regression-
based models and ANN-based models are compared, it is seen that the model
developed using the TreeNet method shows higher performance compared
to the ANN model. In order to make this comparison clearer, time series of the
number of flats variable have been prepared (Figure 3).

58 The Journal of Environment, Urban and Climate