Geology Reference
In-Depth Information
Fig. 5.11 The BIC and AIC values corresponding to different input combinations (Each scenarios
are [1000], [0100], [0010], [0001], [0011], [1001], [1100], [0101], [1010], [0110], [0111], [1011],
[1101], [1110] and [1111] Mask indicates different combinations of the input effects (inclusion and
exclusion indicated by 1 or 0 in the mask). From left to right. The horizon extraterrestrial radiation
(ETR), air dry bulb temperature (DT), wet bulb temperature (WT), and atmospheric pressure (p).)
5.3.4 Modelling Results Using ANN and LLR on 6-Hourly
Records
To compare the performance of entropy and the GT, eight scenarios using LLR and
ANN models were developed. The scenarios 1 and 5 use four inputs (ETR, DT, WT
and p) with 1,010 data points in training process. Three inputs (ETR, DT and WT)
by 1,010 data points are used for training the Scenarios 2 and 6. The Scenarios 3
and 7 use four inputs and 550 data points. Three inputs with 550 data points are
generated the Scenarios 4 and 8. A total of 450 data points are considered for model
testing with all models. The performance of two input data selection methods is
compared by the LLR and ANN models using four global statistics (R
2
, root mean
squared error (RMSE), mean absolute error (MAE) and mean bias error (MBE))
(Figs.
5.12
,
5.13
,
5.14
and
5.15
).
In this study, the conjugate gradient training algorithms along with single layer
architecture is used. Various hidden layer neuron number combinations were tested
Fig. 5.12 Performance of LLR and Conjugate gradient ANN during Training Phase (MBE and
R2 Values)
