1) The study uses neural network models to predict half-daily solar radiation values in Spain, aiming to improve predictions for solar energy feed-in tariffs.
2) The best neural network model reduced the mean root squared error of predictions compared to a persistence model by over 9%.
3) While the first neural network model had errors limited by nonlinear signal behavior, a second model further improved predictions, achieving the best error level possible with the presented methodology.
1 of 1
Download to read offline
More Related Content
Solar radiation forecasting with non lineal statistical techniques and qualitative predictions from spanish national weather service
1. Solar radiation forecasting with non-lineal statistical techniques and 20 08
UN 008
qualitative predictions from Spanish National Weather Service OS r 2
EUR be
to BON
Oc IS
Martín L., Zarzalejo L.F., Polo J., Navarro A., Marchante R. L
1. INTRODUCTION 3. RESULTS
Solar energy feed-in tariff is regulated by (RD 436/2004, Errors of the models essayed are measured in terms of
661/2007) in Spain. Predictions must be given for next 72 mean root mean squared deviation (RMSD). The best
hours and deviations are strongly penalized. A new method NN(z) model is compared to persistence model (PER) in
to predict half daily values of solar radiation is presented. terms of improvement of RMDS.
2. METHODOLOGY 1 N 45.0 ° N
∑ ( xi − xi )
2
RMSD = ˆ 42.5 ° N
Solar radiation is transformed to a new gaussian and N i=1 40.0 ° N
• Madrid RRN AEMet
stationary variable. “Lost component” (LC) is the difference
i − ierrorm 37.5 ° N
betwen extratrrestrial and ground measured solar radiation. improvement = 1 − ÷
i − ierror ÷
35.0 ° N
°
15.0 ° W12 ° ° 0.0 E
.5 W10.0 ° W ° E 5.0 ° E 7.5 E 1
p
7.5 ° W 5.0 ° W 2.5 ° W 0.0 ° 2.5
6000
38
5000
Lost Component L
36
NN(1)
4000 C NN(2)
34
Halfday)
NN(3)
P NN(4)
3000 R NN(5)
32
2
E NN(6)
2000 D NN(7)
% RM SE Prediction G (W /m
I 30 NN(8)
C NN(9)
T NN(10)
1000 28 P ersistence
I
O
0
0
N 26
100 200 300 400 500 600 700 S
Half Day
24
Synoptic predictions of sky conditions (SYN) are used as 22
1 2 3 4 5 6
input to the neual network to test the improvement of the 40
Pre diction horizon (Halfdaily )
predictions. AEMET offers this predicitons in its web page W
I
for each location of Spain and 7 days in adavance. T
35
H NN(1)
Halfday)
NN(2)
30
NN(3)
S NN(4)
Y
2
NN(5)
%RMSE Prediction G (W/m
N 25
NN(6)
NN(7)
C 20
NN(8)
NN(9)
O
NN(10)
N Persistence
D 15
I
T
10
I
O
N 5
S 1 2 3 4 5 6
Prediction horizon (Halfdaily)
4. CONCLUSIONS
The error of the first model is limited by an upper level
which is due to deterministic nonlinear behaviour of the
signal which can’t be followed correctly by neural
network models. The second model improves
Neural Network (NN) is used to predict future values from considerably the prediction. The error has a lower level
observations. NN(z) índica el tamaño del vector patrón de of nine percent which is the best prediction error that can
entrada empleado z=1…10. be achieved with the methology presented.
División de Energías Renovables (Departamento de Energía), CIEMAT, Av. Complutense
nº22, Madrid, 28040, (Madrid) España, +34 913466048, luis.martin@ciemat.es