Professional Documents
Culture Documents
A Review of Solar Irradiance Prediction Techniques
A Review of Solar Irradiance Prediction Techniques
Contact Details:
Luis Martín Pomares
Departamento de Energía. CIEMAT / MEC. España.
Av. Complutense 22, 28040 Madrid, Spain.
Phone: +34 913 466 039. Fax: +34 913 466 037. e-mail: luis.martin@ciemat.es
Abstract:
Until now most of the efforts in the developing of solar energy has been made to
utilize solar energy efficiently, although only minimal resources have been
directed toward forecasting the incoming energy.
1. Introduction
The sun can be considered as a huge fusion nuclear reactor, which consists of
different gases which are retained inside by gravitational forces. The energy as
electromagnetic radiation is produced by the fusion reactions in the inner layers,
basically the kernel, and is transferred to the external layers to be radiated to the
space.
The solar energy reaches the earth in the form of electromagnetic radiation. It is
the main source of energy and produces all the main processes on earth:
dynamic in the atmosphere, the oceans and in general life on earth.
There are several human uses from this resource: solar water heating, water
detoxification, water desalinization, electric power energy generation from solar
thermal power and photovoltaic energy, agricultural applications, etc.
Solar thermal power and photovoltaic energy have a huge developing potential in
near future, supported on the Kyoto protocol agreements from different countries
and on the retribution system which different countries are promoting. In Spain,
Royal Decree 436/2004, which minimizes investment risk to promoters, opens
new perspectives to solar energy development. It is planned to have 500MW of
solar power plants by 2010, and as a remark in 2005 it is expected that more
than 50 power plants begin construction in the near future.
Although considerable effort has been made to utilize solar energy efficiently,
from industrial revolution expecting fossil fuels would run out in the future, only
minimal resources have been directed toward forecasting the incoming energy.
Even that the necessity to have forecasting models that allow optimize the
integration of solar thermal power and photovoltaic within different sources of
electric power generation will grow up as they gain recognition as a resource in
the next year.
Indeed royal decree 436/2005, says, “All facilities with power upper than 10MW
should communicate to the network distributor a prevision of electric energy in
each one of the scheduling periods of the market of electric energy generation.
The prevision should be communicated for each period of the day hourly, almost
30 hours the beginning of the day...”.
Summing up there is a basic need to characterize and predict incoming solar
radiation to be used as a energetic resource. Next sections describe prediction
techniques for different temporal horizons.
2. Forecasting techniques
The generic prediction can be divided in three categories: short term prediction
which prediction horizon is between one hour and one week in advance, medium
term prediction which goes from 1 week to one year and long term prediction
which is longer than one year.
Depending on the specific necessities from different users they will request a
different prediction horizon. For example, for designing a solar energy system or
back-up energy system it would be more useful a climatic prediction than a
prediction from day to day. Nevertheless, once the system is operational a daily
forecasting of incoming energy is more useful than climatic estimation (4).
In this forecasting horizon is where there are more works published. In the next
paragraph will be described most important ones following a chronological order
of appearance and ordering in different techniques.
In 1979, National Weather System from United States (NWS) (5) used the MOS
technique to develop a solar irradiance prediction system relating daily solar
irradiance observed to forecast output produced by one of the NWS numerical
weather prediction models. The equations tested with independent data
produced forecasts with mean absolute errors of 0.74, 0.80 y 0.89 kWh 2 for
m
advance periods of 1, 2 y 3 days, respectively. The mean absolute error for
forecasts based on the expected climatic value was 1.04 kWh 2 .
m
Prediction
Mean absolute error (MAE) kWh
m2
MOS 1 day 0.74
MOS 2 days 0.80
MOS 3 days 0.89
Prediction based on climatic value 1.04
Baker y Casper (6) developed and tested three types of regression equations for
objectively forecasting the percent of extraterrestrial radiation (I0) received in
three different locations: Bismark, ND, Madison, WI, y Chicago, IL. For the first
type of equation, the relationship between MOS forecasts of precipitation and the
observed insolation was determined. For the second type of equation, a MOS
approach was used to relate the observed insolation to relative humidity
forecasts from numerical weather prediction model. The third type of equation
was similar to the second except that the observed relative humidity was related
to the insolation rather than to predicted relative humidity. Verification results for
24 hours forecast indicated that third technique was best. In terms of the percent
of percent of extraterrestrial radiation root mean square error for this method was
18.0.
Although the MOS approach worked well for the 34 stations used to develop
equations, there was no way to determine how well the equations would work in
locations where insolation data were unavailable. This way NWS (8) developed
and tested a MOS approach bases on second order equations. The relation
between global radiation from 30 stations from NOAA Solar Radiation Network
and MOS cloud cover and dew point predictions to develop an unique equation
for all stations. To develop the model data from 1977 January to 1978 December
was used. As can be seen in Table 2 errors for MOS predictions are considerably
lower than predictions based on persistence (prediction is the same as observed
data the day before) or expected climate value. The correlation between
observed and predicted data is much higher for MOS predictions.
GHI
= g ( SK )
GHI clear − sky
where GHIclear-sky is global irradiance under clear sky conditions and g(SK) is a
function from sky cover predicted.
GHI
Figure 2.1 shows relation between SK index and index. It can be
GHI clear − sky
observed that relation isn’t totally lineal.
Figure 2. Relation between SK prediction for 4-8 hours and global irradiance
index observed (10).
For this reason it is used two new types of expressions to model the relationship.
First one was proposed by (11) to relate sky cover index and GHI index from the
following expression:
GHI
GHI clear − sky
(
= 1 − 0, 75 ( N 8 )
3,4
)
where N is the sky cover observed over surface defined in octas.
Resides, Perez establish a new empiric adjust between SK and GHI index,
obtaining new coefficients to define the relation in the following way:
GHI
GHI clear − sky
(
= 1 − 0, 75 ( N 8 )
3,4
)
Verification of model is tested with surface data and with GHI values estimated
from satellite images for different temporal forecasting periods. Table 3
synthesize the estimation results from three models.
Table 3. Comparison of the models Lineal approximation to sky cover index (A),
Kasten & Czeplak (B) and Best Fit Formula (C) (12)
Prediction horizon Relative mean bias error Root mean square error
A B C A B C
Satellite
Results show from equation 3 that parameters surface observed sky cover index
and satellite sky cover index are different.
Satellite data are a high quality source for irradiance information because of
excellent temporal and spatial resolution within short term forecasting horizon. Its
utility is based basically on the strong impact of cloudiness on surface irradiance,
so the description of temporal development of the cloud situation is essential for
irradiance forecasting.
There are several publication related to solar energy forecasting from satellite
images (13) (14) (15) (16) (17), which belongs to the energy meteorology group
from Oldenburg University in Germany.
Cloud index images from Heliosat methods (18) derives a cloud measure from
satellite images and obtain solar surface irradiance incoming.
To predict the future clear sky index two main approaches are used: motion
vector field (19) and neural networks.
Figure 3. Mean RMSE of píxel intensities (14).
Figure 4. Prediction estimation for a location and for a region 35kmx45km (13).
Estimation results from predictions are shown in Figure 4. For high values of
solar irradiance the error is within 10% for a forecasting horizon of 30 meanwhile
if the horizon is of 6 hours the error grows to 25%.
Medium and long term prediction is based in past climatic data and in day by day
climatic data. Past climatic data, for example, can be used to determine the
probability of 3, 4 or 5 consecutives days of cloudy weather. In general,
predictions based on past climatic data assume that the meteorological patterns
are somehow regular. This kind of assumption will not be useful for resource
evaluation seasonally or annually because climatic condition can fluctuate
considerably, instead they will be useful to estimate meteorological conditions
averaged over the lifetime of a solar energy system that will usually deviate only
slightly from the mean value (21).
3. Future works
In a first phase the work to develop will be centred on daily solar irradiance. Pre-
processing of signals with wavelet analyses (22-24) to obtain the signals
expressed in different basic frequencies is the technique with best results to date.
From this irradiance signals it is possible to apply different techniques (temporal
series or neural networks) to predict signals. Finally, solar daily irradiance
prediction will be obtained reconstructing the signal with inverse wavelet
transform.
A way to model the sequential and stationary properties of solar irradiance signal
mapped into several time-frequency domains is using stochastic process
analysis. It is based on analyzing past data to determine, from this information,
the statistical properties that can be used to predict future signal.
Other way to model and forecast signal is using neural networks as it is done in
(22-24). However several improvements can be applied like the use of other kind
of neural networks like self-organising networks (Kohonen maps) that uses rules
of competitive learning. The main characteristics of this kind of networks:
For long term solar irradiance prediction it would be interesting used statistical
techniques like EOF analysis to relate different atmospheric oscillation patterns,
NAO (North Atlantic Oscillation), ENSO (El Niño-Southern Oscillation),… with
expected solar irradiance.
4. References
(1) José Luis García Ortega, Alicia Cantero. Renovables 2050: Un informe sobre el
potencial de las energías renovables en la España peninsular. 2004.
(2) Jose Manuel Gutiérrez, Rafael Cano, Antonio S.Cofiño, Carmen M.Sordo. Redes
probabilísticas y neuronales en las ciencias atmosféricas. Universidad de
cantabria; INM; 2006.
(3) Daniel S.Wilks. Statistical methods in the atmospheric sciences. Academic Press;
1995.
(4) Hulstrom RL. Solar user needs for forecasts and satellite mapping of insolation.
Proc First Workshop on Terrestrial Solar Resource Forecasting and on the Use of
Satellites for Terrestrial Solar Resource Assessment 1981;131-3.
(5) Jensenius JS, Jr., G.M.Carter. Specialized Agricultural Weather Guidance for
South Carolina. TDL Office Note 1979;79(15):16.
(6) Baker DG, M.A.Casper. Subjective forecasting of received solar radiation. Proc
First Workshop on Terrestrial Solar Resource Forecasting and on the Use of
Satellites for Terrestrial Solar Resource Assessment 1981;8-11.
(7) Falconer P. A methodology for Forecasting Morning, Afternoon, and Daily Totals
of Normal Incidence Solar Radiation at Ground Level under Clear or Partly
Cloudy Skies. ASRC 1981;Publication 805:31.
(8) Jensenius JS, Jr. Automated forecasts of daily global solar energy. Progress in
Solar Energy 1983 Jun 1;859-64.
(9) Jensenius JS, Jr. Automated forecasts of daily global solar energy. Progress in
Solar Energy 1983 Jun 1;859-64.
(10) Richard Perez, Kathleen Moore, Steve Wilcox & David Renné, Antoine Zelenka.
Forecasting Solar Radiation - Preliminary evaluation of an approach based upon
the national forecast data base. 2005.
(11) Kasten F., G.Czeplak. Solar and Terrestrial Radiation dependent on the Amount
and type of Cloud. Solar Energy 1979;24:177-89.
(12) Richard Perez, Kathleen Moore, Steve Wilcox & David Renné, Antoine Zelenka.
Forecasting Solar Radiation - Preliminary evaluation of an approach based upon
the national forecast data base. 2006.
(13) Elke Lorenz, Annette Hammer, Detlev Heinemann. Short term forecasting of
solar radiation based on satellite data. 2000.
(14) Annette Hammer, Detlev Heinemann, Carster Hoyer, Elke Lorenz. Satellite based
short-term forecasting of solar irradiance - comparison of methods and error
analysis. 2000.
(17) Annette Hammer, Detlev Heinemann, Elke Lorenz, Bertram Lückehe. Short-term
forecasting of solar radiation based on image analysis of meteosat data. 2000.
(18) Cano D., Monget J.M., Albussion M., Guillard H., Regas N., Wald L. A method
for the determination of global solar radiation from meteorological satellite data.
Solar Energy 1986;37:31-9.
(20) Shuanghua Cao, Jiacong Cao. Forecast of solar irradiance using recurrent neural
networks combined with wavelet analysis. Applied Thermal Engeneering 2005.
(22) Cao JC, Cao SH. Study of forecasting solar irradiance using neural networks with
preprocessing sample data by wavelet analysis. Energy 2006;In Press, Corrected
Proof.
(23) Cao S, Cao J. Forecast of solar irradiance using recurrent neural networks
combined with wavelet analysis. Applied Thermal Engineering 2005 Feb;25(2-
3):161-72.