Page 72 - Çevre Şehir İklim İngilizce - Sayı 4
P. 72

Urban Transformation Forecasts with Artificial Intelligence
                                        Based Algorithms

               Table 7. Performance statistics of the training, verification and test data sets of
                  regression-based models developed for the number of licenses variable
             Training     KRA_Lin  KRA_Üs   KRA_Eks   KRA_Kua    MARS    TreeNet
             RMSE(pcs)     4293     4110     4294      3035      2621     1708
             MAE (pcs)     3207     3102     3208      2374      1964      649
             NS            0.69     0.72      0.69     0.84      0.88      0.95
             Verification  KRA_Lin  KRA_Üs  KRA_Eks   KRA_Kua    MARS    TreeNet
             RMSE(pcs)     8762     9057     8761      7741      7129     7677
             MAE (pcs)     6265     6739     6263      5569      4375     5481
             NS            0.21     0.16      0.21     0.39      0.48      0.40
             Test         KRA_Lin  KRA_Üs   KRA_Eks   KRA_Kua    MARS    TreeNet
             RMSE(pcs)     4657     4564     4658      3919      3980     3472
             MAE (pcs)     3812     3871     3812      3177      2922     2864
             NS            0.56     0.58      0.56     0.69      0.68      0.76
               When Table 7 is examined, it is seen that the lowest RMSE, MAE values
            and the highest NS values for the training and test data sets are obtained
            from the TreeNet model. When looking at the verification data sets, it is seen
            that the lowest RMSE, MAE values and the highest NS values are achieved
            from the MARS model. In other words, the highest performance values for
            the training and test data were obtained from the TreeNet model and the
            highest performance values for the verification dataset were obtained from
            the MARS method. KRA analyses have not given good results compared to
            other forecasting models for the number of licenses. As a general evaluation,
            it can be asserted that the TreeNet method provides high-accuracy results for
            all three data sets, including training, verification and


             Table 9. The equation used for achieving the number of licenses variable and basic
                             functions obtained from the MARS method
            Basic functions
            BF1 = max( 0, N - 0,808965); BF2 = max( 0, 0,808965 -
            N); BF3 = max( 0, Y - 0,633333);
            BF5 = max( 0, G - 0,644563) * BF2; BF11 = max( 0, Y -
            0,366667) * BF2; BF13 = max( 0, G - 0,1) * BF11;
            F_ruhsat_sayısı = 0,524018 - 2,71245 * BF1 - 0,533989 * BF2 + 0,594553 * BF3 + 2,71546 * BF5 -
            1,00556 * BF13;







                                                                 Year 2 / Issue 4 / July 2023  61
   67   68   69   70   71   72   73   74   75   76   77