slide 1: International Journal of Environmental Agriculture Research IJOEAR ISSN:2454-1850 Vol-3 Issue-10 October- 2017
Page | 15
Solar Irradiance Forecasting Using Intelligent Technology
Yi-Ju Chang
1
Ko-Ming Mu
2
Yih-Guang Leu
3
Li-Fen Chou
4
1.2.3
Department of Electrical Engineering National Taiwan Normal University Taiwan
4
Taiwan Power Research Institute Taiwan Power Company Taiwan
Email: leuykntnu.edu.tw corresponding author
Abstract —Because solar power is susceptible to clouds and substances in the air the solar photovoltaic cannot produce
stable power output. Solar irradiance is a measurement of the power output of photovoltaic module. Therefore this paper
uses some different combination inputs of the neural network to develop the solar irradiance forecasting with 24 hours
ahead. Their forecasting performances are evaluated and some comparison results in Taichung solar farm are given.
Keywords —Neural network Solar Irradiance Solar Power Systems.
I. INTRODUCTION
To solve the problems of limited fossil fuels and their impact on the environment renewable resources play an important
role. Solar energy is a very important renewable energy. Based on evaluated condition of solar power solar photovoltaic
becomes the most potential renewable energy in Taiwan. Solar irradiance is a measurement of the power output of
photovoltaic module. However because solar irradiance is influenced by substances in the air the solar photovoltaic cannot
produce stable power output. The power output of photovoltaic module is influenced immediately when the module is
sheltered from the clouds. Besides the material of solar cell air temperature module’s position and orientation also affect
the power output of the photovoltaic module. Therefore it is an important issue to forecast solar irradiance accurately.
Forecasting accuracy is not only influenced by the change of weather but also surroundings and the effectiveness of method
and data. Developing an excellent solar power systems not only can wield the change of solar power but also help power
company to allocate power. The accuracy of solar irradiance forecasting is the basis of solar power forecasting 1-6. There
are many intelligent approaches to forecast solar irradiance such as neural networks 7-13.
The paper uses neural network technology to develop the solar power forecasting with 1-24 hours ahead. Some different
features for solar forecasting are proposed and their forecasting performances are evaluated. Moreover comparison results in
Taichung solar farm in Taiwan 14 are given.
II. METHODS OF SOLAR IRRADIANCE FORECASTING
The backpropagation BP neural network is used in this paper and the main structure of which is input layer hidden layer
and output layer. Input layer transfer received external information into the network and which is used to represent the input
variable. Hidden layer process the input data which is used to represent the relationship between input variables. Input data
are transferred to the output layer after being converted by nonlinear transform function. Output layer receives and outputs
the information from upper layer and is used to represent the output variable of network. The layers connect to each other to
deliver information and this connection is called “weight”. The weight value is acquired from the iterative learning
calculation of the neuro. Using supervised learning method is to reduce the discrepancy between output and expected value
and to minimize the performance index of the network. The training method in this paper is Levenberg-Marquardt algorithm
which is a modified gradient descent method and possesses the ability of fast training.
Based on the assumption that there is no cloud and shelter solar irradiance can be calculated as follows 15.
A. The air mass is defined as follows.
1
where is zenith angle.
slide 2: International Journal of Environmental Agriculture Research IJOEAR ISSN:2454-1850 Vol-3 Issue-10 October- 2017
Page | 16
B. The intensity of the direct component of sunlight in units of kW/m
2
on the assumption that the location height
above sea level is zero is defined as follows.
2
where the value 1.353 is solar constant the value 0.7 means that solar irradiance irradiates to the earths surface through
atmosphere is 0.7 times of itself and the value 0.678 is empiric value.
C. The global solar irradiance is defined as follows.
3
where the diffuse radiation is about 10 of the direct component of sunlight.
Root Mean Square Error RMSE can show the difference between the forecasting values and observed values judge the
quality of the neural network and judge the level of convergence in the process. RMSE is closely related to the number of
training data. The larger the RMSE the greater the difference of the forecasting values and observed values may be. The
closer the difference between RMSE and 0 the closer the difference between the forecasting values and observed values may
be. RMSE is defined as follows.
4
where Y
j
is forecasting values is observed values.
Based on some different combination inputs of the neural network this paper proposes six methods to develop the solar
irradiance forecasting with 24 hours ahead. The method 1 is to use the historical solar irradiance values in the daytime. The
method 2 is to use the historical solar irradiance values in the daytime and the temperature. The method 3 is to use the
historical solar irradiance values in the daytime and the solar altitude. The method 4 is to use the historical solar irradiance
values in the daytime the temperature and the solar altitude. The method 5 is to use the difference value of historical solar
irradiance values and the solar altitude. The method 6 is to use the historical solar irradiance values and the future solar
irradiance value that is calculated on the assumption that there is no cloud and shelter 15. The training data in Taichung
solar farm in Taiwan 14 were in the period from March 2014 to September 2015. In the methods 1-6 the values of solar
irradiance next hour is the only output. The data of the network training only apply the daytime data. Defining the day or
night is depending on the calculated values of solar altitude 15. If the solar altitude is larger than zero it is in the daytime.
Conversely it is at night.
In the method 1 the architecture of back propagation neural network was made from 3 inputs 7 hidden layer neurons and 1
output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 3 inputs are the observed
value of solar irradiance at the time t the time t-1 the time t-2.
In the method 2 the architecture of back propagation neural network was made from 4 inputs 9 hidden layer neurons and 1
output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 4 inputs are the observed
value of solar irradiance at the time t the time t-1 the time t-2 and the temperature.
In the method 3 the architecture of back propagation neural network was made from 4 inputs 11 hidden layer neurons and
1 output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 4 inputs are the solar
altitude the calculated value of solar irradiance the observed value of solar irradiance at the time t and the time t-1.
In the method 4 the architecture of back propagation neural network was made from 5 inputs 10 hidden layer neurons and
1 output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 5 inputs are the solar
altitude the temperature the calculated value of solar irradiance the observed value of solar irradiance at the time t and the
time t-1.
In the method 5 the architecture of back propagation neural network was made from 3 inputs 8 hidden layer neurons and 1
output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 3 inputs are the solar
slide 3: International Journal of Environmental Agriculture Research IJOEAR ISSN:2454-1850 Vol-3 Issue-10 October- 2017
Page | 17
altitude the difference value of calculated value and observed value of solar irradiance and the difference values of solar
irradiance at the time t and the time t-1.
In the method 6 the architecture of back propagation neural network was made from 4 inputs 7 hidden layer neurons and 1
output. The transition function of hidden layer neurons is on-linear hyperbolic tangent function. 4 inputs are the calculated
value of solar irradiance next hour the observed value of solar irradiance at the time t the time t-1 and the time t-2.
In order to judge the quality of the forecasting the RMSEs of the methods 1-6 are calculated by equation 4. The RMSEs of
the methods 1-6 are shown in Figs.1-5. The results of RMSE comparison for the solar irradiance forecasting with 1 3 6 12
and 24 hours ahead are shown in Figs.1-5 respectively.
According to comparison for the values of RMSE neural networks combined with future solar irradiance have the lowest
value of RMSE.
FIG.1. RMSE of each forecasting methods with 1 hour ahead.
FIG.2. RMSE of each forecasting methods with 3 hours ahead.
slide 4: International Journal of Environmental Agriculture Research IJOEAR ISSN:2454-1850 Vol-3 Issue-10 October- 2017
Page | 18
FIG.3. RMSE of each forecasting methods with 6 hours ahead.
FIG.4. RMSE of each forecasting methods with 12 hours ahead.
FIG.5. RMSE of each forecasting methods with 24 hours ahead.
slide 5: International Journal of Environmental Agriculture Research IJOEAR ISSN:2454-1850 Vol-3 Issue-10 October- 2017
Page | 19
III. CONCLUSION
This paper has proposed some different features for solar irradiance forecasting and presented some comparison results of
solar irradiance forecasting in Taichung solar farm. According to the RMSE comparison figures the method depend on
historical and future solar irradiance values is better than other methods in forecasting with 1-24 hours ahead. Moreover the
results are likely to be affected by the historical data in the forecasting with short-time ahead.
ACKNOWLEDGEMENTS
This work was supported by the Taiwan Power Company R.O.C. and the National Science Council of Taiwan R.O.C.
under Grants NSC 102-2221-E-003-011.
REFERENCES
1 R. Perez M. Beauharnois E. Lorenz S. Pelland and J. Schlemmer “Evaluation of numerical weather prediction solar irradiance
forecasts in the US” in Proc. ASES Annual Conference Raleigh NC USA May. 2011.
2 P. Mathiesen and J. Kleissl “Evaluation of numerical weather prediction for intra-day solar forecasting in the continental United
States” Solar Energy vol. 85 pp.967-977 2011.
3 E. Lorenz J. Hurka D. Heinemann and HG. Hans“Irradiance Forecasting for the Power Prediction of Grid-Connected Photovoltaic
Systems” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing vol. 2 no. 1 Mar. 2009.
4 P. Bacher H. Madsen and H.A. Nielsen “Online short-term solar power forecasting” Solar Energy vol. 83 pp.1772-1783 2009.
5 A. Hammer D. Heinemann E. Lorenz and B. Luckehe “Short-term forecasting of solar radiation: a statistical approach using
satellite data” Solar Energy vol. 67 No.1-3 pp.139-150 1999.
6 C.W. Chow B. Urquhart M. Lave A. Dominguez J. Kleissl J. Shields and B. Washom “Intra-hour forecasting with a total sky
imager at the UC San Diego solar energy testbed” Solar Energy vol. 85 pp.2881-2893 2011.
7 F. Wang Z. Mi S. Su and H. Zhao“Short-Term Solar Irradiance Forecasting Model Based on Artificial Neural Network Using
Statistical Feature Parameters” Energies vol.5 pp.1355-1370 2012.
8 A. Mellit and A. M.Pavan “A 24-h forecast of solar irradiance using artificial neural network: Application for performance prediction
of a grid-connected PV plant at Trieste Italy” Solar Energy vol. 84 pp.807-821 2010.
9 A. Yona T. Senjyu A.Y. Saber F. Toshihisa S. Hideomi and K. Chul-Hwan“Application of Neural Network to One-Day-Ahead 24
hours Generating Power Forecasting for Photovoltaic System” IEEE Intelligent Systems Applications to Power Systems Niigata
pp.1-6 2007.
10 J. Cao and X. Lin“Application of the diagonal recurrent wavelet neural network to solar irradiation forecast assisted with fuzzy
technique” Engineering Applications of Artificial Intelligence vol. 21 pp.1255-1263 2008.
11 A. Mellit M. Benghanem and S.A. Kalogirou “An adaptive wavelet-network model for forecasting daily total solar-radiation”
Applied Energy vol. 83 pp.705-722 2006.
12 C. Paoli C. Voyant M. Muselli and M.-L. Nivet “Forecasting of Preprocessed Daily Solar Radiation Time Series Using Neural
Networks” Solar Energy vol. 84 no. 12 pp. 2146–2160 Dec. 2010.
13 F. Gutierrez-Coreaab M. Manso-Callejoa M. Moreno-Regidora and M. Manrique-Sanchoa “Forecasting short-term solar irradiance
based on artificial neural networks and data from neighboring meteorological stations” Solar Energy vol.134 pp. 119-131 2016.
14 http://www.taipower.com.tw/content/new_info/new_info-b31.aspxLinkID8 2017.
15 http://www.pveducation.org/ 2017.