The same structure with one hidden layer of four neurons has been taken. Due to its
convergence capabilities, the Levenberg-Marquart training algorithm was used. For the
activation function of the neurons, the tangent hyperbolic was chosen due to its non-
linearity, continuity and derivability. The training data were relative values because they
were divided by the theoretical maximum solar irradiance, i.e. the solar irradiance with
an atmospheric transmission factor of 1.0.
The Artificial Neural Network (ANN) used for the solar radiation predictor has four
normalized inputs:
1
Grel(k): Relative solar irradiance at current time k
2
Grel(k - 1): Relative solar irradiance at time k - 1 (one hour ago)
3
Grel(k + 6 - 24): Relative solar irradiance 24 hours before the time of prediction
4
Gmax(k + 6): Computed maximum solar irradiance at the time of prediction
and one normalized output:
1
Grel(k + 6): Relative solar radiation at the time of prediction
The newly developed predictor (called "new ANN") is compared with the one used in
the NEUROBAT project, with a reference model that uses the current measurement of
the relative solar irradiance as the prediction value and with a more recent meteorological
physical model (MRM) developed by Muneer et al. [Muneer et al., 1998]. Weather data
used for the comparison are synthetic values generated by the METEONORM program
[MeteoTest, 1996] (except the results of the Muneer model that have been obtained with
real weather data). Training is performed on the six first months of the year, and
evaluation is performed on the last six months. Results are given in table of figure 2.13.
Both ANN models give better results than the reference one, which shows that it is worth
using ANN for prediction. The accuracy of the new ANN model is confirmed by its
results quite similar to the NEUROBAT ones. Moreover, results of ANN models are
even better than the ones of MRM. But it should be mentioned that the latter come from
real weather data, which is maybe detrimental.
Figure 2.13: Mean values and standard deviations of the 6-hours prediction error of the
horizontal global solar radiation for different models
Even with ANN models, standard deviation is quite large, which attests to the diffi- culty
of solar radiation prediction. Qualitative results of the prediction with the new ANN
model are depicted on Figure 2.14. They are sufficiently accurate to provide valuable
information to the heating system about future solar gains.
Summary :
The Artificial Neural Network (ANN) used for the solar radiation predictor has four normalized inputs: 1 Grel(k): Relative solar irradiance at current time k 2 Grel(k - 1): Relative solar irradiance at time k - 1 (one hour ago) 3 Grel(k + 6 - 24): Relative solar irradiance 24 hours before the time of prediction 4 Gmax(k + 6): Computed maximum solar irradiance at the time of prediction and one normalized output: 1 Grel(k + 6): Relative solar radiation at the time of prediction The newly developed predictor (called "new ANN") is compared with the one used in the NEUROBAT project, with a reference model that uses the current measurement of the relative solar irradiance as the prediction value and with a more recent meteorological physical model (MRM) developed by Muneer et al.
Tags :
ann,prediction,irradiance,results,relatie,time,model,used,radiation,models,data,grelk,which