Energy 81 (2015) 627e644

Contents lists available at ScienceDirect

Energy
journal homepage: www.elsevier.com/locate/energy

Forecasting solar radiation using an optimized hybrid model

by Cuckoo Search algorithm
Jianzhou Wang a, He Jiang b, *, Yujie Wu c, Yao Dong c
a
School of Statistics, Dongbei University of Finance and Economics, Dalian 116025, China
b
Department of Statistics, Florida State University, Tallahassee, FL 32306-4330, USA
c
School of Mathematics and Statistics, Lanzhou University, Lanzhou 730000, China

a r t i c l e i n f o a b s t r a c t

Article history: Due to energy crisis and environmental problems, it is very urgent to find alternative energy sources
Received 25 August 2014 nowadays. Solar energy, as one of the great potential clean energies, has widely attracted the attention of
Received in revised form researchers. In this paper, an optimized hybrid method by CS (Cuckoo Search) on the basis of the OP-ELM
8 December 2014
(Optimally Pruned Extreme Learning Machine), called CS-OP-ELM, is developed to forecast clear sky and
Accepted 1 January 2015
Available online 27 January 2015
real sky global horizontal radiation. First, MRSR (Multiresponse Sparse Regression) and LOO-CV (leave-
one-out cross-validation) can be applied to rank neurons and prune the possibly meaningless neurons of
the FFNN (Feed Forward Neural Network), respectively. Then, Direct strategy and Direct-Recursive
Keywords:
OP-ELM (Optimally Pruned Extreme
strategy based on OP-ELM are introduced to build a hybrid model. Furthermore, CS (Cuckoo Search)
Learning Machine) optimized algorithm is employed to determine the proper weight coefficients. In order to verify the
Direct strategy and Direct-Recursive effectiveness of the developed method, hourly solar radiation data from six sites of the United States has
strategy been collected, and methods like ARMA (Autoregression moving average), BP (Back Propagation) neural
Cuckoo Search algorithm network and OP-ELM can be compared with CS-OP-ELM. Experimental results show the optimized
Global horizontal radiation forecast hybrid method CS-OP-ELM has the best forecasting performance.
© 2015 Elsevier Ltd. All rights reserved.

1. Introduction Panel on Climate Change). Low carbon economy is aimed at mini-

Nowadays, due to a rise in price and a drop in reserves of fossil
fuels, renewable energy, such as solar, wind and biomass, has
caught more and more attention from both developed and devel-
oping countries, especially from the countries that are heavy en-
ergy consumers [1]. Besides, the use of fossil fuels (coal, oil and
natural gas) leads to environmental concern, such as greenhouse
gases emission. Carbon dioxide, as one of the major greenhouse
gases, can cause the rise of average surface temperature of the
earth. In 2011, global carbon dioxide emissions were 31.3 GtCO2.
Among this, carbon dioxide emissions from the combustion of coal
were13.7 GtCO2, from oil were 11.1 GtCO2, from gas were 6.3 GtCO2
and from others were 0.2 GtCO2 [2]. In order to avoid catastrophic
climate change, a sustainable development path based on a low
carbon economy has been canvassed by international development
agencies such as the UN IPCC (United Nations Intergovernmental
* Corresponding author. Tel.: þ1 850 644 3218; fax: þ1 850 644 5271.
E-mail address: hejiang@stat.fsu.edu (H. Jiang).
mizing the energy consumption of fossil fuels, reducing greenhouse
gas emissions, exploring renewable energy and achieving a
winewin situation between economic development and ecological
environment protection as far as possible. During Joint
Announcement on Climate Change of U.S.-China, United States
announced a new target cutting net greenhouse gas emission
26e28% below 2005 levels by 2025. At the same time, the Chinese
government also announced targets to peak carbon dioxide emis-
sions around 2030, with the intention to peak early, and to increase
the non-fossil fuel share of all energy to around 20% by 2030 [3].
All renewable energies, except for geothermal and tidal, stem
from the sun. Solar energy, as one of the great potential renewable
energies, refers mainly to the use of solar radiation in practical
application. Solar technologies are widely characterized as either
passive solar or active solar depending upon the way they capture,
convert and distribute solar energy. Passive solar techniques
contain orienting a building to the Sun, selecting materials with
favorable thermal mass or light dispersing natures, and designing
spaces that spontaneously circulate air. Active solar techniques
contain the use of solar photovoltaic modules and panels and solar

http://dx.doi.org/10.1016/j.energy.2015.01.006
0360-5442/© 2015 Elsevier Ltd. All rights reserved.
628 J. Wang et al. / Energy 81 (2015) 627e644
Nomenclature
Full name, abbreviation
Akaike Information Criterion AIC
an optimized hybrid
method by Cuckoo Search on the
basis of the Optimally Pruned
Extreme Learning Machine CS-OP-ELM
autoregression AR
autoregression integrated moving average ARIMA
autoregression moving average ARMA
a strategy that combines both Direct and Recursive
strategy Direct-Recursive strategy
back propagation BP
clear sky diffuse horizontal radiation CSKY Dif
clear sky direct normal radiation CSKY Dir
clear sky global horizontal radiation CSKY Glo
concentrated solar power CSP
Cuckoo Search algorithm CS
extraterrestrial radiation on a horizontal surface ETR
thermal collectors to utilize the energy. Furthermore, with rapid
economic development, many industrial sectors need more and
more electricity power supply. Solar energy can be converted into
electricity, either using PV (photovoltaic) directly, or using CSP
(concentrated solar power) indirectly. PV converts light into
electric current using the photoelectric effect and produces plenty
of waste heat, which can be recovered for thermal use by
attaching PV board with recuperating tubes filled with carrier
fluids. CSP systems use lenses or mirrors and tracking systems to
heat a fluid such as water in a boiler to produce steam used to
produce power in a steam turbine coupled to an electrical
generator [4].
Many researchers have developed several methods to forecast
the solar radiation. These approaches can be classified into three
different types [5]: (1) The first one allows forecasting the future
solar radiation at the time (t þ p) based on the past observed data at
the time (t þ p1, t þ p2, …, t). The aim is to find a relationship
between the inputs and the outputs data. So models like AR
(Autoregressive) [6], ARMA (Autoregressive moving average) [7],
ARIMA (autoregressive integrated moving average) [8] and Markov
chain [9], have been found to be very suitable; (2) The second one is
based on meteorological parameters at the time (t þ p1, t þ p2,
…, t) such as air temperature, relative humidity, wind speed, wind
direction, cloud, sunshine duration, clearness index, pressure, etc.
and geographical coordinates as latitude and longitude. The aim is
to find a relationship between these factors and solar radiation.
MLP (multilayer perceptron) network [10e13], RBF (Radial Basis
Function) network [14], and fuzzy logic [15] are suitable; (3) The
last one combines the two previous approaches. In this case, the
input data are the past observed solar radiation data and other
meteorological factors at the time (t þ p1, t þ p2, …, t). To sum
up, it is found that the first type only considers the past data and the
second one only applies relative parameters. Thus, it is more
reasonable to employ the third one.
In this paper, we use the third type of method-an optimized
hybrid model called CS-OP-ELM to forecast solar radiation. There
are four advantages of this optimized approach. To begin with,
high-dimension data can be taken into account as input datasets.
Specifically, different from traditional method like ARMA model,
the proposed model integrates some meteorological factors to do
Extreme Learning Machine ELM
Feed Forward Neural Network FFNN
Greenwich Mean Time GMT
Least Angle Regression LARS
leave-one-out cross-validation LOO-CV
mean azimuth angle Azimuth
mean relative error MRE
mean zenith angle Zenith
multilayer perceptron MLP
Multiresponse Sparse Regression MRSR
National Solar Radiation Database NSRDB
Optimally Pruned Extreme Learning Machine OP-ELM
Ordinary Least Squares OLS
Particle Swarm Optimization PSO
photovoltaic PV
Radial Basis Function RBF
real sky diffuse horizontal radiation RSKY Dif
real sky direct normal radiation RSKY Dir
real sky global horizontal radiation RSKY Glo
root mean square error RMSE
forecasting for solar radiation. Second, the optimization algorithm
called CS (Cuckoo Search) optimization algorithm can be used to
confirm the weight parameters. With proper selection of the pa-
rameters, redundant information can be eliminated or overlooked,
a more effective model can be built. Third, the model CS-OP-ELM
creates admirable improvements which are satisfactory for the
current research. In particular, OP-ELM model can build a neural
network with large number of neurons, rank neurons and deter-
mine the number of neurons. The optimized hybrid strategy by CS
algorithm, based on Direct strategy and Direct-Recursive strategy,
can lessen system loss. So there is no surprise that the hybrid
model can take full advantages of these models. Fourth, the pro-
posed CS-OP-ELM model is automatic in essence, and it does not
require to make complex decision about each case of the modeling
processes.
The structure of this work is presented. Section 2 introduces
the hybrid forecasting theory, CS (Cuckoo Search) optimization
algorithm and two kinds of forecasting strategies: Direct strategy
and Direct-Recursive strategy. Section 3 presents the methodol-
ogy including OP-ELM (Optimally Pruned Extreme Learning
Machine), ELM (Extreme Learning Machine), MRSR (Multi-
response Sparse Regression), and LOO-CV (leave-one-out cross-
validation) algorithm in detail. In Section 4, in order to illustrate
the effective of the optimized hybrid model, a case study about
solar radiation data forecast in four sites of the United States can
be simulated, BP (Back Propagation) and ARMA model can also be
applied to compare. Finally, Section 5 concludes the results of
this work.
2. The optimized hybrid forecasting method by CS (Cuckoo
Search) algorithm based on two kinds of strategies
This work proposes a new optimized hybrid method called CS-
OP-ELM. At first, two kinds of strategies based on OP-ELM (Opti-
mally Pruned Extreme Learning Machine) can be applied to forecast
solar radiation. Then a hybrid model is built, and CS (Cuckoo
Search) algorithm is employed to determine the weight coefficients
of the hybrid model. Finally, the optimized hybrid model can be
presented as the developed new method to forecast solar radiation.
The detailed process is presented in Fig. 1.
 J. Wang et al. / Energy 81 (2015) 627e644 629

Fig. 1. Flowchart of the optimized hybrid method CS-OP-ELM.

630 J. Wang et al. / Energy 81 (2015) 627e644

2.1. The hybrid forecasting theory

vy  u ¼ t l
Le (5)
If there are T types of methods to solve one forecasting problem
(in this paper, T ¼ 2), then the hybrid forecasting theory indicates that where xtþ1
i
and xti represent the solution i at times tþ1 and t,
these forecasting methods should be given proper weight co- respectively. a > 0 is the step size which should be related to the
efficients. Suppose that yi(i ¼ 1,2,…,N) is the actual data, N is the scales of the problem of interest. In general, it is taken as one. The
number of forecasting samples, 6t ðt ¼ 1; 2; …; TÞ is the weight co- product 4 means entry-wise multiplications.
efficient for the tth forecasting approach, and
yti(t ¼ 1,2,…,T; i ¼ 1,2,…,N) is the ith forecasting value for the tth 2.3. Two kinds of strategies based on OP-ELM (Optimally Pruned
forecasting approach, the mathematical formula can be shown as [16] Extreme Learning Machine)

X
T This work uses an optimized hybrid method including Direct
yi ¼ 6t ðyti þ εti Þ (1) strategy and DirRec strategy based on OP-ELM (Optimally Pruned
t¼1
Extreme Learning Machine) to forecast solar radiation. The detailed
So the hybrid forecasting method is expressed as process of strategies can be seen in Fig. 2.

X
T
2.3.1. Direct strategy
b
b
yi ¼ btb
6 y ti (2)
In the Direct strategy [20], the regressor size r is a hyper-
t¼1
parameter of the model. The aim is to forecast p steps ahead us-
where 6 b t is the estimated value of 6t , b
b
y i is the hybrid forecasting ing regressors yN,yN1,/,yNrþ1. The mathematical equation is
value, and t ¼ 1,2,…,T, i ¼ 1,2,…,N. written as:
It is quite crucial to determine the weight coefficients for each
b
y Nþ1 ¼ f1 ðyN ; yN1 ; /; yNrþ1 Þ
forecasting method. Currently, the optimization algorithm is a key
b
y Nþ2 ¼ f2 ðyN ; yN1 ; /; yNrþ1 Þ
tool. In this paper, an optimized approach called CS (Cuckoo Search) (6)
«
is applied to get the weight coefficients. The objective function (or
b
y Nþp ¼ fp ðyN ; yN1 ; /; yNrþ1 Þ
the fitness function) is given by Ref. [16]

X
N   X
T
f ¼ b
b
y i  yi st b t ¼ 1;
6 bi  1
06 (3)
i¼1 t¼1 2.3.2. Direct-Recursive strategy
The Direct-Recursive strategy combing both Direct [20] and
The aim is to minimize objective function f using CS (Cuckoo Recursive [21] strategy was proposed by Sorjamma and Lendasse
Search) algorithm. [22]. In the first step, it coincides with the Direct strategy, then all
forecasting values serve as new regressors and the order of model
grows. The mathematical form is written as:
2.2. CS (Cuckoo Search) algorithm
b
y Nþ1 ¼ f1 ðyN ; yN1 ; /; yNrþ1 Þ
CS (Cuckoo Search) is a nature inspired meta-heuristic algo- b
y Nþ2 ¼ f2 ð b
y Nþ1 ; yN ; yN1 ; /; yNrþ1 Þ
rithm based on the obligate brood parasitic behavior of some «  (7)

vy flight distribution of some

cuckoo spices in combination with Le b
y Nþp ¼ fp b y Nþp1 ; /; b y Nþ1 ; yN ; yN1 ; /; yNrþ1
birds and fruit flies [17]. The cuckoo bird lays its egg in the nests of
other host birds. The host bird takes care of this egg presuming it
as its own egg. If the host bird identifies this egg, then it either
destroys this egg or abandons the nest and builds a new nest at 3. OP-ELM (Optimally Pruned Extreme Learning Machine)
some new location [18]. For simplicity in describing CS algorithm,
we now use three idealized rules [18,19]: (1) Each cuckoo lays one The OP-ELM (Optimally Pruned Extreme Learning Machine) has
egg at a time, and dumps its egg in randomly chosen nest; (2) The three main steps presented as follows [22]:
best nests with high quality of eggs or solutions from each gen-
eration are carried over to the next generation; (3) The number of Step 1: Build a regular ELM model with initially large number of
available host nests is fixed in each generation. An egg is chosen neurons;
from a randomly selected nest, and the egg laid by a cuckoo is Step 2: Rank neurons using MRSR (Multiresponse Sparse
discovered by the host bird with a probability pa 2 [0,1]. Such Regression);
nests are discarded and removed from further calculations. The Step 3: Use LOO-CV (leave-one-out cross-validation) algorithm
last assumption can be approximated by a fraction pa of the n nests to determine the number of neurons to prune.
being replaced by new nests (with new random solutions at new
locations).
The generation of new solutions xtþ1 i
is done by using Le
vy 3.1. ELM (Extreme Learning Machine)
flights in Eq. (4). Le
vy flights essentially provide a random walk
while their random steps are drawn from a Le
vy distribution for ELM (Extreme Learning Machine) is a tuning-free learning al-
large steps that has an infinite variance with an infinite mean in Eq. gorithm for a unified single-hidden-layer FFNN (Feed Forward
(5). Correspondingly, the consecutive steps of a cuckoo essentially Neural Network) [23]. Its basic idea is that ELM has only one hidden
form a random walk process which obeys a power-law step-length layer, and the parameters of the hidden layer contain the input
distribution with a heavy tail [19]. weights and biases of the hidden nodes. All hidden node parame-
ters are determined randomly, which are independent on the target
xtþ1
i

vyðlÞ
¼ xti þ a4Le (4) function and the training data [24]. ELM method can be described
mathematically as follows [25,26]:
 J. Wang et al. / Energy 81 (2015) 627e644 631

Fig. 2. Flowchart of Direct strategy and Direct-Recursive strategy.

Given a training set ‫ ¼ א‬fðxi ; yi Þj i ¼ 1; 2; …; N; xi 2ℝn ; jth hidden node and the output nodes and the input weights
yi 2ℝm g, the output of ELM that the number of hidden nodes is L is wt ¼ [wt1,wt2,/,wtL]T (t ¼ 1,2,…,N). In general, Eq. (6) can be got as
presented: the following matrix format:

X
L   Hb ¼ Y (9)
f ðxi Þ ¼ bj g aj $xi þ bj (8)
j¼1
2 3 2 3
gða1 $x1 þ b1 Þ / gðaL $x1 þ bL Þ Hx1
D
where i ¼ 1,2,…,N,bj is the threshold of the jth hidden node, H¼4 « / « 5 ¼ 4« 5
aj ¼ [a1j,a2j,/,arj]T is the weight vector connecting the jth hidden gða1 $xN þ b1 Þ / gðaL $xN þ bL Þ NL HxN NL
node and the input nodes, g(x) is the activation function,
(10)
bj ¼ [bj1,bj2,/,bjm]T (j ¼ 1,2,…,L) is the weight vector connecting the
632 J. Wang et al. / Energy 81 (2015) 627e644

n o n o
b ¼ ½b1 ; b2 ; /; bL T ckmax ¼ max ckj ; A ¼ jjckj ¼ ckmax (13)
j

Y ¼ ½y1 ; y2 ; /; yN T Step 4: Collect the regressors that belong to A as an n  jAj

matrix XA ¼ ½/; xj ; /j2A and calculate an OLS (Ordinary Least
where H stands for the hidden layer output matrix.
Squares) estimate
Different from the classical learning approach, ELM tends to
obtain the least training error and the least norm of output weight
kþ1
 1
together for the better generalization performance by Y ¼ XA XTA XA XTA T (14)
b ¼ argminðkHb  Yk; kbkÞ, So both the standard optimization
method and the minimal norm least square method need to be Note that the OLS estimate involves kþ1 regressors at the kth
adopted [25,26]. In most real applications, the number of hidden step.
nodes is much less than the number of samples (L ≪ N). As long as
the hidden layer nodes are enough, FFNN will converge towards Step 5: Greedy forward selection adds regressors based on Eq.
any continuous function in ELM with the input weights and hidden (12) and the OLS estimate Eq. (14) is used. However, we define a
layer biases randomly determined. In fact, ELM process is a linear less greedy algorithm by moving from the MRSR estimate Yk
system and its basic steps are as follows [25,26]: kþ1
toward the OLS estimate Y , i.e. in the direction
k kþ1 k
Step 1: Randomly determine the input weight aj and the bias bj, U ¼Y  Y , but we will not reach it. The largest step
i ¼ 1,2…,N. possible is taken in the direction of Uk until some xj, where j;A,
Step 2: Calculate the hidden layer output matrix H. has as large cumulative correlation with the current residuals as
Step 3: Calculate the output weights the already added regressors. The MRSR estimate Yk is updated

 kþ1 
Ykþ1 ¼ Y k þ gk Y  Yk (15)
1
b¼H Y (11)
And the weight matrix can be updated
where H1 is the MooreePenrose generalized inverse of the matrix
  kþ1
H. Wkþ1 ¼ 1  gk Wk þ gk W (16)

In order to make the update, the correct step size gk needs to be

3.2. MRSR (Multiresponse Sparse Regression) calculated. The corresponding update is as follows:

LARS (Least Angle Regression) [27] is a new model selection ckþ1

j ðgÞ ¼ j1  gjckmax for all j2A (17)
mechanism that selects a parsimonious set of covariates from a
large dimension. LARS is computationally efficient to handle the  
 
large factor set and generally produces sparse solutions [28]. MRSR ckþ1
j
ðgÞ ¼ akþ1
j
 gbkj  for all j;A (18)
1
(Multiresponse Sparse Regression) is an extension of the LARS for
the multi-output case. Due to the ranking provided by the MRSR, it 8 9
< ck þ sT akj =
is used to rank the neurons of the model [22]. max
Gj ¼ (19)
Assume that an np matrix T ¼ [t1,t2,/,tp] denote the targets :ck þ sT bk ;
max j s2S
and an nm matrix X ¼ [x1,x2,/,xm] denote the regressors. MRSR
adds each column of the regressor matrix one by one to the model 
Yk ¼ XWk, where Yk ¼ [yk1,yk2,/,ykp] is the approximation of target T. gk ¼ min gjg  0 and g2Gj for some j;A (20)
The mp weight matrix Wk has k nonzero rows at the kth step of
the MRSR. When p ¼ 1 MRSR coincides with the LARS algorithm. where akj ¼ ðT  Y k ÞT xj ,bkj ¼ ðY
kþ1
 Y k ÞT xj , S is the set of all 2p sign
This makes MRSR rather an extension than an improvement of
vectors of size p1, i.e. the elements of s may be either 1 or 1, and
LARS. The detailed algorithm of MRSR is in the following way [29]:
A records all numbers of selected variables.
Step 1: Set k ¼ 0, initialize all elements of Y0 and W0 to zero, and
Step 6: Calculate g and update Yk based on the above method,
normalize both T and X to zero mean and one standard
repeat Step 2 to Step 5 until picking all the variables.
deviation.
Step 2: Calculate the cumulative correlation between the jth
regressor xj and the current residuals according to Eq. (12).
3.3. LOO-CV (leave-one-out cross-validation algorithm)
 T  p 
X T 
  LOO-CV (leave-one-out cross-validation) is a useful algorithm to
ckj ¼ 
 TY
k
xj 
 ¼
 ti  yk xj 
 i  (12)
1 estimate the generalization of model. So LOO-CV is widely
i¼1
employed to model selection [30], especially when data set size is
The criterion measures the sum of absolute correlations be- considered as small. Its basic thought is presented: Given n samples
tween the residuals and the regressor over all p target variables at available in a data set and n candidate models, each model is
the kth step of the MRSR. trained with n-1 samples and then is tested on the sample that was
left out. This process is repeated n times until every sample in the
Step 3: Let the maximum cumulative correlation be represented data set have been used once as cross-validation instance [31].
by ckmax and the group of regressors that meet the maximum by Finally, the errors of LOO-CV, a measure of the model generalization
A, capability, can be got with Eq. (21) for all candidate models [32].
 J. Wang et al. / Energy 81 (2015) 627e644 633

! !
  X X
1 XX
1
yi  Hxi H Y 1 Hxi H1 ¼ uTi VV T
U ¼ uTi U
ri ¼ yi  fi ðxi Þ ¼   (21) i
1  Hxi H1 i
  i   i
X
L
ILL 0 I 0
¼ uTi U ¼ uTi LL ui ¼ u2li
Hxi H1 YðHxi H1 Þi yi
0 0 i
0 0
l¼1
where fi ðxi Þ ¼ , this proof is as follows.
1ðHxi H1 Þi
(27)
The following notations can be introduced:
PL
f0(x) ¼ Hxb0 is the training function of ELM in all the samples; where 2 ¼ 1. According to Lemma 2,
l¼1 uli
fi(x) ¼ Hxbi is the training function of ELM in one sample ‫א‬ðxi ; yi Þ;
  X
N
1  Hxi H1 ¼ u2li > 0 (28)
h iT i
l¼Lþ1
Hi ¼ HTx1 ; /; HTxi1 ; HTxiþ1 ; /; HTxN ;
ðN1ÞL
when N  2L, Eq. (28) holds with probability one.

Y ¼ ½y1 ; y2 ; /; yN T ; 4. Simulation experiment and results analysis

4.1. Data collection and discussion

Yi ¼ ½y1 ; /; yi1 ; yiþ1 ; /; yN T ;
This paper collects the hourly data from 2008 to 2010 in six sites
of the United States, provided by NSRDB (National Solar Radiation
Y0i ¼ ½y1 ; /; yi1 ; fi ðxi Þyiþ1 ; /; yN T : Database) [33]. Prior to the experiments, the collected data can be
classified in terms of the same months. That is, the same months of
Hxi H1 YðHxi H1 Þi yi
In order to prove fi ðxi Þ ¼ , two lemmas are 2008e2010 are put in one group. Then data can be split into two
1ðHxi H1 Þi
presented. parts: training data (in 2008 and 2009) and test data (in 2010). The
Lemma 1. Two linear equations Eq. (22) and Eq. (23) have the former is applied to build the hybrid model, while the latter is
same minimum two-norm least squares solution. employed to calculate the test errors.
In this study, in order to verify the effectiveness of the proposed
model called CS-OP-ELM adequately, four models, i.e. ARMA, BP,
Hi b ¼ Y i (22) OP-ELM by Direct strategy and OP-ELM by Direct-Recursive strat-
egy are going to be compared. Two experiments including clear sky
condition and real sky condition have been done in the experi-
Hb ¼ Y0i (23) ments. For forecasting hourly clear sky global horizontal radiation
(CSKY Glo, unit: Wh/m2), these eight groups of data including
hourly mean zenith angle (Zenith, unit: deg), mean azimuth angle
P (Azimuth, unit: deg), extraterrestrial radiation on a horizontal
Lemma 2. H is defined according to Eq. (10) (N  2L), H ¼ UT V, surface (ETR, unit: Wh/m2), real sky direct normal radiation (RSKY
D D
and U,V is orthogonal matrix. If U¼ ½u1 ; u2 ; /; uN ¼ ðuij ÞNN , then Dir, unit: Wh/m2), real sky diffuse horizontal radiation (RSKY Dif,
PN 2
cj2f1; 2; …; Ng; i¼Lþ1 uij > 0 with probability one. unit: Wh/m2), clear sky global horizontal radiation (CSKY Glo, unit:
Based on Lemma 1, Wh/m2, clear sky means when cloud cover set to zero), clear sky
direct normal radiation (CSKY Dir, unit: Wh/m2) and clear sky
diffuse horizontal radiation (CSKY Dif, unit: Wh/m2) can be applied
fi ðxi Þ ¼ Hxi bi ¼0Hxi H1 0
2 Yi 31 as input variables. For forecasting real sky global horizontal radia-
0ði1Þ1 tion (RSKY Glo, unit: Wh/m2), these eight sets of data including
1 @
¼ Hxi H Y  yi  fi ðxi Þ 5A
4
(24) hourly mean zenith angle (Zenith, unit: deg), mean azimuth angle
0ðNiÞ1 (Azimuth, unit: deg), extraterrestrial radiation on a horizontal
   
¼ Hxi H Y  Hxi H1 yi þ Hxi H1 fi ðxi Þ
1 surface (ETR, unit: Wh/m2), real sky global horizontal radiation
i i

After some simple formulations,

Table 1
The locations of six sites.
    
1  Hxi H1 fi ðxi Þ ¼ Hxi H1 Y  Hxi H1 yi Site 1 Site 2
i i ADAK NAS, AK FAYETTEVILLE DRAKE FIELD, AR
  Latitude 51.883 36
1 1 (25)
Hxi H Y  Hxi H yi Longitude 176.65 94.167
⇔fi ðxi Þ ¼   i Time zone 10 6
1
1  Hxi H Site 3 Site 4
i
TROY AF, AL TUSCALOOSA MUNICIPAL AP, AL
31.867 33.217
where Eq. (25) holds if and only if 1(HxiH1)i s 0.
Latitude
P Longitude 86.017 87.617
According to H ¼ UT V, it can be got Time zone 6 6
Site 5 Site 6
TAMPA INTERNATIONAL AP, FL LONDON-CORBIN AP, KY
X X
1 Latitude 27.967 33.217
Hxi ¼ uTi V; H1 ¼ VT U (26) Longitude 82.533 87.617
Time zone 5 6
thus,
 634
Table 2
Results comparison by five kinds of forecasting models in Site 1.

Feb 9 Model (clear sky) Model (real sky) Nov 6 Model (clear sky) Model (real sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

10:00 65.11% 41.10% 12.50% 43.13% 5.35% 13.58% 21.30% 5.88% 3.99% 0.27% 10:00 44.09% 53.76% 34.15% 35.41% 11.01% 17.77% 39.17% 25.58% 19.71% 15.95%
11:00 34.08% 15.59% 0.00% 37.62% 4.43% 2.58% 33.18% 2.59% 1.50% 0.47% 11:00 50.89% 42.15% 49.15% 35.24% 51.05% 8.69% 0.19% 12.75% 8.13% 5.98%
12:00 81.46% 34.11% 42.65% 38.87% 40.68% 1.16% 0.34% 1.73% 0.94% 1.08% 12:00 58.82% 27.73% 39.47% 8.39% 31.62% 5.36% 13.84% 8.40% 4.57% 3.64%
13:00 16.66% 80.04% 1.66% 49.73% 7.06% 0.29% 14.31% 1.86% 0.48% 1.69% 13:00 30.03% 35.21% 50.00% 27.27% 45.56% 4.68% 2.64% 6.94% 3.62% 3.58%
14:00 63.42% 19.29% 0.92% 53.30% 7.46% 0.03% 1.48% 1.88% 1.04% 0.98% 14:00 25.98% 41.17% 45.66% 30.72% 47.64% 5.42% 4.51% 6.76% 3.71% 4.15%
15:00 33.11% 53.70% 0.92% 48.74% 6.80% 0.05% 6.00% 1.88% 0.94% 0.50% 15:00 30.90% 37.19% 12.50% 14.59% 32.01% 5.16% 3.38% 7.28% 3.40% 2.32%

J. Wang et al. / Energy 81 (2015) 627e644

16:00 36.11% 27.86% 14.74% 35.79% 8.30% 0.47% 12.82% 1.54% 0.81% 1.17% 16:00 85.04% 97.85% 57.43% 70.66% 56.71% 7.63% 6.47% 9.27% 6.35% 7.18%
17:00 73.96% 57.19% 1.43% 20.12% 3.11% 0.89% 7.69% 1.69% 0.66% 0.12% 17:00 97.12% 95.69% 191.30% 71.91% 41.74% 11.15% 1.00% 13.51% 10.42% 11.17%
18:00 67.28% 38.66% 2.70% 51.38% 8.70% 1.58% 14.50% 2.48% 0.93% 1.88% 18:00 5.11% 60.10% 23.53% 51.83% 15.08% 21.40% 60.34% 28.57% 20.64% 13.19%
19:00 5.30% 104.2% 33.33% 73.86% 18.54% 11.31% 32.49% 5.26% 4.72% 5.72%
Daily MRE 47.65% 47.18% 11.08% 45.26% 11.04% 3.20% 14.41% 2.68% 1.60% 1.39% Daily MRE 47.55% 54.54% 55.91% 38.45% 36.94% 9.69% 14.62% 13.23% 8.95% 7.46%
Daily RMSE 37.49 35.81 10.26 33.29 10.07 1.93 25.61 4.59 2.22 2.72 Daily RMSE 57.62 60.09 65.34 37.27 59.78 13.32 16.87 18.34 10.94 10.02
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (real sky) Aug 8 Model (clear sky) Model (real sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 19.99% 82.99% 26.37% 55.81% 9.18% 1.40% 1.29% 0.00% 1.54% 6.02%
8:00 29.02% 26.55% 47.50% 27.67% 1.15% 7.65% 1.23% 14.29% 4.24% 0.40% 8:00 101.6% 113.5% 18.81% 132.5% 48.84% 2.90% 23.05% 1.02% 2.59% 2.21%
9:00 28.27% 103.08% 32.50% 79.60% 20.63% 2.61% 1.10% 9.66% 5.45% 1.90% 9:00 43.62% 128.5% 24.39% 126.3% 39.06% 2.17% 0.92% 1.64% 3.23% 1.33%
10:00 25.29% 7.08% 20.47% 26.37% 41.64% 2.05% 1.62% 6.33% 3.81% 0.52% 10:00 34.69% 227.2% 47.50% 193.7% 63.44% 1.90% 0.80% 1.72% 3.34% 2.81%
11:00 7.78% 29.82% 24.85% 27.26% 35.21% 2.13% 0.61% 4.74% 2.24% 1.09% 11:00 92.69% 171.8% 90.91% 75.7% 84.13% 1.73% 3.14% 1.80% 3.38% 1.01%
12:00 57.89% 13.30% 23.15% 31.14% 39.52% 0.36% 1.31% 3.67% 3.69% 0.85% 12:00 211.2% 805.9% 150% 1788.8% 769.2% 1.81% 3.50% 1.91% 3.48% 0.31%
13:00 33.11% 8.28% 16.92% 28.93% 35.81% 0.28% 3.97% 3.06% 3.11% 0.27% 13:00 1133% 130.4% 420% 571.7% 470.1% 1.85% 1.93% 1.96% 3.55% 1.45%
14:00 32.31% 41.56% 17.45% 37.95% 28.25% 0.23% 0.35% 2.82% 3.15% 0.36% 14:00 78.71% 51.4% 67.86% 90.06% 71.5% 2.01% 0.21% 2.10% 3.70% 2.60%
15:00 76.69% 93.61% 133.8% 23.87% 17.29% 0.73% 1.39% 2.86% 3.39% 0.64% 15:00 84.99% 64.50% 384.8% 193.1% 275.3% 1.93% 3.86% 1.87% 3.46% 1.43%
16:00 64.65% 44.00% 39.58% 49.84% 27.44% 1.39% 0.17% 3.09% 3.67% 0.94% 16:00 49.87% 249.5% 501.3% 184.3% 328.3% 1.97% 1.80% 1.74% 3.34% 1.96%
17:00 33.16% 76.58% 7.57% 74.07% 20.88% 0.72% 0.59% 3.88% 3.60% 0.71% 17:00 51.37% 17.96% 466% 147.8% 293.9% 2.05% 3.05% 1.53% 3.11% 0.00%
18:00 88.71% 107.7% 10.27% 19.36% 42.56% 2.14% 1.84% 5.29% 3.70% 0.57% 18:00 56.14% 59.38% 57.31% 43.4% 52.53% 2.08% 7.52% 1.15% 2.76% 1.19%
19:00 20.90% 63.10% 21.00% 40.07% 36.30% 3.78% 1.32% 7.67% 2.57% 1.22% 19:00 33.72% 69.89% 27.89% 105.7% 59.19% 2.10% 1.22% 0.55% 2.15% 0.27%
20:00 55.40% 61.41% 70.18% 50.62% 76.70% 9.26% 0.09% 11.56% 3.76% 0.49% 20:00 71.75% 140.9% 221.4% 80.85% 246.8% 2.41% 3.71% 0.00% 1.70% 5.27%
21:00 16.95% 31.46% 7.14% 15.29% 44.57% 4.57% 12.15% 14.29% 0.38% 5.07% 21:00 65.41% 32.75% 57.66% 25.74% 44.41% 1.12% 14.16% 3.08% 1.34% 1.47%
Daily MRE 40.72% 50.54% 33.74% 38.00% 33.42% 2.71% 1.98% 6.66% 3.34% 1.07% Daily MRE 141.9% 156.4% 170.8% 254.4% 190.4% 1.96% 4.68% 1.47% 2.85% 1.96%
Daily RMSE 87.62 98.44 92.42 62.91 59.54 9.14 10.62 24.22 18.55 4.06 Daily RMSE 112.15 130.85 195.48 107.17 139.45 10.36 17.93 9.67 17.98 8.70
(Wh/m2) (Wh/m2)
Table 3
Results comparison by five kinds of forecasting models in Site 2.

Feb 9 Model (clear sky) Model (real sky) Nov 6 Model (clear sky) Model (real sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

8:00 44.26% 48.95% 65.79% 41.82% 72.17% 14.14% 32.47% 23.68% 24.96% 8.96% 7:00 32.38% 51.29% 0.00% 46.31% 11.27% 8.76% 1.98% 0.00% 11.20% 16.58%
9:00 54.74% 30.08% 74.77% 54.14% 79.09% 11.20% 5.84% 15.14% 16.94% 9.18% 8:00 3.59% 23.10% 11.49% 4.47% 12.36% 14.13% 32.66% 11.49% 5.84% 1.74%
10:00 42.55% 13.41% 67.27% 38.01% 72.54% 6.94% 2.38% 10.19% 12.22% 7.26% 9:00 14.89% 37.15% 6.37% 12.02% 10.89% 7.43% 11.59% 6.37% 4.84% 1.21%
11:00 25.89% 8.01% 67.01% 7.13% 40.62% 5.53% 1.95% 7.94% 9.91% 2.14% 10:00 12.20% 12.99% 3.64% 18.55% 9.75% 4.84% 3.99% 3.64% 4.85% 1.56%
12:00 41.02% 0.69% 66.74% 32.93% 16.85% 5.13% 4.74% 6.62% 8.74% 3.99% 11:00 15.53% 17.71% 2.30% 7.88% 6.97% 3.53% 4.44% 2.30% 5.36% 2.31%
13:00 67.71% 111.5% 33.99% 98.12% 14.35% 4.30% 4.17% 6.29% 8.42% 3.50% 12:00 12.75% 13.28% 2.06% 16.41% 8.14% 2.84% 3.59% 2.06% 5.08% 2.01%

J. Wang et al. / Energy 81 (2015) 627e644

14:00 115.1% 90.40% 34.87% 101% 35.06% 4.43% 4.87% 6.32% 8.43% 3.22% 13:00 15.19% 18.28% 2.06% 16.38% 8.13% 2.89% 12.47% 2.06% 2.76% 0.68%
15:00 87.13% 218.9% 35.71% 136.5% 30.83% 4.48% 7.09% 7.07% 9.17% 4.00% 14:00 22.37% 18.93% 2.83% 18.12% 9.03% 3.62% 5.24% 2.83% 6.71% 3.80%
16:00 2.12% 59.66% 63.38% 14.41% 67.12% 5.33% 4.40% 8.74% 10.60% 0.31% 15:00 14.18% 26.51% 4.32% 12.80% 9.37% 4.81% 0.28% 4.32% 3.81% 0.27%
17:00 40.57% 7.36% 71.43% 33.97% 74.45% 8.41% 1.12% 13.12% 15.00% 7.95% 16:00 17.61% 29.52% 6.37% 18.76% 11.97% 6.03% 0.03% 6.37% 1.52% 2.63%
18:00 29.37% 31.33% 58.62% 23.55% 31.05% 14.02% 55.27% 23.81% 25.20% 12.50% 17:00 22.66% 36.69% 11.49% 8.67% 14.46% 10.09% 19.39% 11.49% 3.16% 8.67%
Daily MRE 50.04% 56.39% 58.14% 52.87% 48.56% 7.63% 11.30% 11.72% 13.60% 5.73% Daily MRE 16.67% 25.95% 4.81% 16.40% 10.21% 6.27% 8.70% 4.81% 5.01% 3.77%
Daily RMSE 125.71 148.16 159.40 130.70 120.30 24.99 23.18 36.22 45.68 19.22 Daily RMSE 67.68 83.23 14.01 65.65 37.08 17.20 31.03 14.01 20.55 9.44
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (real sky) Aug 8 Model (clear sky) Model (real sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

6:00 29.34% 1.27% 18.18% 38.78% 3.43% 6.29% 1.78% 14.29% 14.81% 4.15% 6:00 19.52% 13.74% 12.50% 22.95% 21.96% 11.33% 18.11% 12.50% 2.76% 0.62%
7:00 131.4% 188.5% 40.48% 118.2% 50.52% 7.22% 4.49% 3.59% 4.50% 3.77% 7:00 20.66% 22.43% 50.39% 20.82% 20.05% 10.61% 26.32% 11.63% 9.66% 8.24%
8:00 153.1% 142.1% 13.83% 111.2% 39.83% 2.98% 12.32% 1.06% 2.07% 4.07% 8:00 10.96% 15.16% 6.36% 20.00% 9.40% 2.89% 9.71% 6.36% 5.04% 3.62%
9:00 244.7% 66.4% 147.6% 145.5% 93.23% 0.00% 3.32% 0.17% 1.17% 2.59% 9:00 15.90% 53.49% 15.15% 19.72% 19.75% 2.30% 1.05% 3.55% 3.96% 2.66%
10:00 202.1% 224.1% 253.2% 156.7% 127.2% 0.88% 0.47% 0.39% 0.53% 0.96% 10:00 15.37% 24.18% 9.38% 16.86% 13.06% 1.19% 0.76% 2.24% 1.61% 0.19%
11:00 208.6% 254.9% 242% 221.7% 157.9% 0.80% 1.73% 0.45% 0.49% 0.43% 11:00 22.48% 2.45% 32.00% 14.88% 6.86% 1.07% 3.54% 1.53% 1.94% 0.62%
12:00 245.4% 259.6% 214% 188.4% 133.2% 0.40% 0.57% 0.52% 0.38% 1.96% 12:00 2.76% 23.28% 6.75% 12.10% 12.07% 0.22% 0.12% 1.29% 1.76% 0.43%
13:00 232.3% 167.8% 164% 192% 121.7% 0.63% 1.15% 0.51% 0.43% 0.54% 13:00 23.68% 40.56% 1.26% 8.73% 8.38% 0.94% 0.05% 1.26% 1.40% 0.04%
14:00 124.6% 107.3% 30.88% 131.3% 54.76% 0.69% 0.11% 0.43% 0.51% 0.48% 14:00 3.29% 4.39% 24.73% 3.82% 8.32% 1.11% 0.78% 1.42% 1.93% 0.61%
15:00 84.60% 60.4% 24.38% 159.4% 67.60% 0.21% 0.00% 0.24% 0.72% 0.79% 15:00 1.22% 3.46% 9.12% 6.66% 4.92% 1.57% 0.29% 1.83% 1.82% 0.46%
16:00 114.1% 203.6% 1.52% 112.5% 36.41% 0.72% 0.18% 0.60% 1.51% 0.12% 16:00 12.81% 18.87% 2.68% 8.83% 8.70% 2.39% 3.05% 2.68% 1.13% 0.37%
17:00 127.6% 137.3% 3.57% 140.1% 50.16% 2.31% 0.90% 1.91% 2.79% 0.84% 17:00 18.92% 4.92% 25.62% 1.22% 6.33% 3.87% 1.96% 4.34% 2.95% 1.49%
18:00 92.8% 54.28% 5.26% 141.4% 50.40% 6.23% 5.44% 4.25% 5.09% 2.46% 18:00 13.10% 1.09% 7.27% 14.18% 13.87% 7.39% 4.97% 7.27% 7.69% 6.45%
19:00 36.14% 1.74% 61.11% 24.38% 50.44% 7.96% 15.37% 7.14% 7.87% 2.25% 19:00 14.13% 3.09% 13.10% 22.03% 21.30% 14.63% 0.29% 13.10% 14.73% 13.68%
Daily MRE 144.8% 133.5% 87.14% 134.4% 74.05% 2.67% 3.42% 2.54% 3.06% 1.82% Daily MRE 13.91% 16.51% 15.45% 13.77% 12.50% 4.39% 5.07% 5.07% 4.17% 2.82%
Daily RMSE 315.22 312.23 263.89 288.39 177.70 8.02 15.40 5.35 7.45 8.81 Daily RMSE 95.50 153.87 104.16 73.21 63.05 11.86 16.13 15.10 14.45 8.38
(Wh/m2) (Wh/m2)

635
 636
Table 4
Results comparison by five kinds of forecasting models in Site 3.

Feb 9 Model (clear sky) Model (real sky) Nov 6 Model (clear sky) Model (real sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 51.9% 25.2% 28.6% 22.7% 10.7% 9.16% 2.05% 0.00% 4.16% 8.48% 7:00 48.35% 5.28% 31.25% 38.72% 26.59% 35.37% 44.37% 31.25% 23.33% 17.53%
8:00 56.5% 67.3% 101.4% 87.9% 56.0% 7.58% 2.89% 5.07% 4.31% 2.34% 8:00 57.14% 17.37% 22.98% 26.15% 10.88% 17.32% 44.36% 22.98% 12.03% 5.15%
9:00 37.3% 23.1% 70.4% 55.9% 20.1% 3.18% 9.60% 3.85% 4.74% 4.90% 9:00 45.40% 45.97% 52.42% 36.47% 21.22% 13.17% 3.46% 14.78% 8.36% 1.83%
10:00 31.9% 8.6% 55.2% 41.6% 7.4% 3.28% 7.24% 3.09% 2.56% 0.92% 10:00 46.43% 76.57% 13.54% 31.77% 19.39% 9.93% 0.96% 10.49% 8.43% 2.45%
11:00 73.7% 57.0% 116.6% 96.8% 47.1% 2.65% 6.87% 2.31% 2.57% 1.94% 11:00 33.50% 33.23% 30.49% 32.86% 18.92% 7.83% 10.47% 8.82% 9.37% 3.78%
12:00 91.9% 45.1% 142.7% 133.3% 115.3% 2.39% 5.41% 4.73% 2.92% 2.20% 12:00 42.75% 19.88% 69.20% 38.08% 21.45% 7.93% 13.31% 8.15% 8.76% 3.14%

J. Wang et al. / Energy 81 (2015) 627e644

13:00 244.3% 246.5% 311.9% 280.0% 202.5% 1.94% 1.82% 5.56% 1.96% 1.73% 13:00 24.20% 20.14% 30.17% 16.40% 2.85% 7.08% 7.87% 8.75% 7.13% 1.13%
14:00 300.6% 253.2% 390.3% 348.0% 242.7% 1.87% 2.42% 2.77% 1.83% 1.63% 14:00 34.61% 6.63% 58.10% 34.82% 18.39% 10.14% 15.70% 10.69% 6.79% 0.48%
15:00 209.8% 223.7% 307.0% 271.2% 181.7% 1.48% 2.35% 0.77% 0.81% 0.10% 15:00 37.71% 16.09% 64.99% 43.45% 29.04% 13.09% 11.60% 14.87% 9.31% 2.94%
16:00 163.6% 133.0% 238.0% 194.4% 79.5% 1.66% 1.01% 1.20% 1.91% 1.86% 16:00 30.20% 36.54% 23.85% 25.23% 9.56% 22.62% 12.79% 23.85% 12.61% 5.73%
17:00 305.8% 253.3% 291.2% 267.4% 212.9% 0.83% 13.70% 1.53% 3.33% 4.66% 17:00 33.11% 1.33% 31.58% 32.80% 18.71% 33.01% 10.88% 31.58% 22.35% 16.33%
Daily MRE 142.5% 121.5% 186.7% 163.6% 106.9% 3.27% 5.03% 2.81% 2.83% 2.80% Daily MRE 39.40% 25.37% 38.96% 32.43% 17.91% 16.14% 15.98% 16.93% 11.68% 5.50%
Daily RMSE 234.85 206.08 326.90 288.84 201.36 11.83 25.01 18.62 12.08 9.82 Daily RMSE 185.65 178.66 231.46 159.97 90.73 48.86 60.64 54.29 41.80 13.82
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (real sky) Aug 8 Model (clear sky) Model (real sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

6:00 11.22% 6.85% 7.14% 4.31% 5.37% 14.94% 18.12% 7.14% 15.53% 8.07% 6:00 27.89% 32.52% 17.14% 17.78% 15.90% 7.55% 3.10% 2.70% 5.52% 7.91%
7:00 20.29% 15.60% 7.76% 14.13% 17.32% 14.77% 13.77% 7.76% 9.48% 8.24% 7:00 23.94% 6.09% 1.46% 3.80% 26.17% 10.23% 1.26% 1.46% 3.29% 5.41%
8:00 26.44% 32.37% 5.07% 15.20% 0.68% 6.86% 15.11% 5.07% 5.87% 4.32% 8:00 18.44% 23.78% 0.47% 3.63% 19.69% 5.98% 5.28% 0.47% 1.23% 0.49%
9:00 24.95% 5.40% 3.81% 6.34% 7.16% 4.14% 2.81% 3.78% 4.32% 2.68% 9:00 11.41% 13.11% 0.32% 3.11% 15.51% 3.92% 3.03% 0.32% 0.65% 1.09%
10:00 13.11% 27.68% 3.14% 16.84% 19.69% 2.62% 4.48% 3.14% 2.29% 0.11% 10:00 12.63% 13.80% 18.81% 14.38% 2.27% 2.84% 2.50% 0.25% 0.00% 1.78%
11:00 18.98% 4.17% 2.99% 13.77% 15.96% 2.09% 3.51% 2.99% 3.21% 1.45% 11:00 0.23% 1.59% 0.22% 4.20% 11.05% 2.10% 1.72% 0.22% 0.96% 0.75%
12:00 28.79% 3.58% 2.96% 24.17% 15.22% 1.91% 2.45% 13.16% 3.19% 1.44% 12:00 5.46% 21.34% 1.56% 4.34% 2.11% 1.73% 0.95% 0.10% 0.70% 1.03%
13:00 0.96% 26.18% 28.09% 13.43% 11.97% 2.01% 2.23% 2.93% 3.33% 1.66% 13:00 25.55% 23.83% 40.68% 35.65% 17.07% 1.72% 0.12% 0.11% 1.54% 0.12%
14:00 11.31% 2.00% 4.56% 2.28% 2.42% 2.30% 3.25% 2.99% 3.61% 2.02% 14:00 4.69% 14.05% 19.10% 15.97% 6.55% 1.78% 0.61% 0.11% 0.49% 1.25%
15:00 28.39% 20.80% 3.47% 11.73% 0.27% 3.29% 4.24% 3.47% 4.36% 2.87% 15:00 6.82% 10.09% 0.27% 1.52% 4.85% 1.86% 6.33% 0.27% 0.64% 2.46%
16:00 37.64% 34.87% 85.33% 53.38% 15.41% 5.41% 2.37% 4.36% 5.68% 4.34% 16:00 5.95% 32.14% 0.36% 2.87% 14.14% 2.55% 5.98% 0.36% 1.20% 3.07%
17:00 4.81% 9.07% 6.19% 2.38% 2.88% 10.32% 6.53% 17.26% 8.59% 7.66% 17:00 8.83% 17.12% 0.58% 2.33% 11.79% 7.23% 7.92% 0.58% 1.64% 3.57%
18:00 5.87% 15.74% 8.41% 2.19% 2.22% 20.34% 2.54% 13.08% 13.68% 13.87% 18:00 5.87% 22.14% 2.19% 1.43% 14.76% 3.27% 0.41% 2.19% 3.29% 5.48%
19:00 16.85% 16.81% 0.00% 20.87% 26.35% 4.28% 19.23% 0.00% 22.66% 30.01% 19:00 11.93% 25.36% 22.22% 17.79% 1.34% 6.06% 8.77% 0.00% 1.68% 0.04%
Daily MRE 17.83% 15.79% 12.06% 14.36% 10.21% 6.81% 7.19% 6.70% 7.56% 6.34% Daily MRE 12.12% 18.35% 8.95% 9.20% 11.66% 4.20% 3.43% 0.65% 1.63% 2.46%
Daily RMSE 128.41 105.03 91.52 101.11 78.39 23.58 29.23 42.44 25.52 16.39 Daily RMSE 64.82 100.11 88.21 77.23 62.65 17.71 20.17 1.89 6.37 10.31
(Wh/m2) (Wh/m2)
Table 5
Results comparison by five kinds of forecasting models in Site 4.

Feb 9 Model (clear sky) Model (real sky) Nov 6 Model (clear sky) Model (real sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 28.81% 3.73% 15.38% 5.84% 2.34% 0.68% 14.27% 15.40% 1.35% 6.93%
8:00 87.1% 33.0% 8.8% 98.3% 85.3% 4.89% 10.61% 4.30% 6.02% 0.86% 8:00 37.96% 22.34% 5.38% 19.51% 38.75% 4.09% 0.45% 5.39% 3.34% 0.76%
9:00 130.7% 105.9% 14.3% 146.8% 152.0% 2.96% 11.34% 1.90% 1.23% 0.05% 9:00 36.67% 7.17% 2.63% 16.24% 20.00% 2.14% 0.62% 2.64% 1.24% 0.24%
10:00 82.8% 109.4% 16.4% 92.6% 191.6% 2.47% 18.42% 1.40% 1.03% 0.11% 10:00 22.19% 17.96% 4.62% 12.84% 3.87% 3.25% 6.98% 1.67% 2.37% 0.13%
11:00 156.0% 162.9% 17.6% 174.9% 185.1% 2.47% 5.99% 1.41% 0.67% 0.21% 11:00 41.28% 17.32% 34.78% 12.67% 37.44% 2.18% 7.08% 1.25% 1.25% 0.23%
12:00 170.0% 192.2% 95.5% 181.6% 137.7% 2.58% 1.22% 1.54% 0.45% 0.40% 12:00 57.79% 53.29% 27.26% 45.58% 0.14% 2.88% 12.42% 1.18% 1.98% 0.25%

J. Wang et al. / Energy 81 (2015) 627e644

13:00 77.6% 17.0% 232.1% 61.7% 136.6% 3.00% 6.81% 1.94% 1.30% 1.05% 13:00 55.49% 31.58% 48.85% 34.35% 36.16% 2.08% 7.38% 1.38% 1.15% 0.05%
14:00 162.4% 147.6% 228.2% 157.8% 153.9% 3.49% 12.24% 2.43% 1.85% 1.57% 14:00 55.66% 55.30% 80.25% 24.50% 38.59% 1.99% 0.08% 1.81% 1.07% 0.47%
15:00 166.3% 181.6% 221.3% 162.9% 257.9% 4.08% 5.36% 3.01% 2.90% 2.09% 15:00 59.88% 1.03% 56.97% 39.83% 40.02% 3.66% 5.02% 3.00% 2.83% 1.68%
16:00 167.6% 190.7% 207.3% 166.0% 138.5% 5.45% 76.48% 4.51% 1.90% 2.95% 16:00 53.13% 41.14% 67.77% 23.88% 6.04% 8.63% 10.47% 5.22% 8.04% 1.11%
17:00 32.1% 37.5% 69.1% 27.1% 48.4% 8.15% 7.36% 7.23% 4.86% 5.19% 17:00 34.00% 85.44% 2.56% 9.63% 11.33% 0.03% 3.63% 12.83% 0.68% 8.74%
Daily MRE 123.3% 117.8% 111.0% 127.0% 148.7% 3.95% 15.58% 2.97% 2.22% 1.45% Daily MRE 43.90% 30.57% 31.50% 22.26% 21.33% 2.87% 6.22% 4.71% 2.30% 1.87%
Daily RMSE 202.40 206.15 256.52 206.06 244.01 16.84 97.18 11.77 8.29 6.84 Daily RMSE 226.90 166.36 204.84 138.79 134.32 13.07 35.72 9.07 9.47 2.72
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (real sky) Aug 8 Model (clear sky) Model (real sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

6:00 1.7% 22.7% 1.7% 21.3% 14.7% 4.91% 6.09% 1.71% 7.20% 7.01% 6:00 12.74% 37.85% 28.00% 10.84% 20.64% 19.72% 15.40% 28.01% 15.07% 19.67%
7:00 1.7% 25.7% 1.6% 23.8% 21.9% 2.33% 7.32% 1.64% 4.86% 3.45% 7:00 6.00% 15.53% 19.75% 10.39% 7.87% 12.47% 16.18% 18.30% 8.58% 9.77%
8:00 242.9% 203.4% 242.8% 164.8% 104.6% 0.48% 4.88% 1.73% 1.42% 1.27% 8:00 14.70% 1.16% 9.41% 18.84% 16.96% 7.20% 7.83% 9.42% 4.31% 4.89%
9:00 246.3% 89.8% 246.1% 188.0% 219.1% 0.51% 2.48% 1.96% 1.44% 1.06% 9:00 3.83% 54.19% 5.35% 4.63% 25.21% 4.83% 0.23% 5.36% 2.42% 2.68%
10:00 247.3% 185.5% 247.1% 193.6% 160.8% 1.50% 0.25% 1.81% 0.19% 1.00% 10:00 12.34% 14.33% 3.60% 14.69% 11.58% 4.10% 3.78% 3.61% 1.95% 2.40%
11:00 248.7% 200.1% 248.6% 150.0% 162.3% 1.84% 0.33% 1.81% 0.23% 0.74% 11:00 16.72% 44.47% 7.00% 17.79% 11.80% 3.44% 4.35% 2.76% 1.35% 1.64%
12:00 249.0% 198.7% 248.8% 210.8% 169.7% 1.82% 0.54% 1.72% 0.23% 0.46% 12:00 25.82% 0.52% 55.73% 18.47% 25.99% 3.39% 1.54% 2.37% 1.39% 1.11%
13:00 191.6% 214.0% 191.5% 147.8% 116.9% 1.36% 1.92% 1.44% 0.27% 0.74% 13:00 26.66% 2.25% 50.65% 20.60% 27.78% 4.01% 2.51% 2.49% 2.12% 1.01%
14:00 83.4% 55.3% 83.3% 67.7% 43.6% 1.23% 1.17% 1.12% 0.35% 1.21% 14:00 187.29% 171.28% 145.76% 194.55% 138.01% 3.57% 2.21% 2.92% 1.47% 1.23%
15:00 243.9% 119.4% 243.8% 156.6% 155.0% 0.97% 1.81% 0.93% 0.63% 1.54% 15:00 13.96% 17.19% 9.04% 13.97% 27.36% 3.12% 4.44% 3.98% 0.66% 1.70%
16:00 164.1% 132.0% 164.0% 123.0% 78.6% 0.40% 3.03% 0.87% 1.31% 2.35% 16:00 0.32% 14.75% 17.35% 2.90% 0.04% 4.71% 1.28% 6.18% 2.07% 3.18%
17:00 137.8% 112.7% 137.7% 97.1% 93.9% 0.45% 1.63% 1.37% 2.48% 3.16% 17:00 12.93% 6.10% 21.55% 12.46% 4.63% 8.25% 14.40% 10.54% 5.33% 5.40%
18:00 0.7% 54.0% 0.6% 15.3% 13.2% 4.56% 3.31% 0.66% 7.36% 6.36% 18:00 10.89% 16.41% 36.59% 6.73% 17.00% 4.40% 15.36% 20.01% 19.71% 4.01%
19:00 0.1% 5.5% 0.0% 3.2% 0.4% 1.64% 1.41% 0.01% 0.43% 5.47% 19:00 61.37% 91.98% 81.82% 58.95% 70.12% 25.87% 0.47% 31.26% 21.80% 23.00%
Daily MRE 147.1% 115.6% 147.0% 111.6% 96.8% 1.71% 2.58% 1.34% 2.03% 2.56% Daily MRE 28.97% 34.86% 35.11% 28.99% 28.93% 7.79% 6.43% 10.52% 6.30% 5.83%
Daily RMSE 444.90 353.70 444.53 332.02 290.50 9.44 12.50 10.24 6.64 8.61 Daily RMSE 160.38 178.64 212.99 153.82 150.13 25.58 25.23 27.02 14.62 13.45
(Wh/m2) (Wh/m2)

637
 638
Table 6
Results comparison by five kinds of forecasting models in Site 5.

Feb 9 Model (clear sky) Model (real sky) Nov 6 Model (clear sky) Model (real sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

8:00 280.7% 253.6% 50.0% 87.8% 103.6% 74.12% 88.54% 16.68% 25.52% 12.84% 8:00 49.46% 63.3% 11.76% 70.80% 53.17% 2.91% 11.28% 3.14% 3.37% 8.45%
9:00 240.5% 104.8% 184.2% 38.2% 11.1% 19.62% 16.72% 12.86% 13.95% 9.96% 9:00 242.12% 285.0% 20.93% 57.82% 262.00% 0.27% 46.98% 0.56% 0.48% 4.56%
10:00 175.3% 376.2% 76.3% 127.0% 64.0% 12.44% 15.40% 8.08% 7.26% 6.70% 10:00 147.61% 219.5% 0.00% 65.63% 163.70% 0.90% 3.29% 0.55% 1.24% 1.05%
11:00 170.3% 238.4% 269.5% 182.5% 155.9% 8.59% 3.38% 5.59% 5.49% 7.66% 11:00 16.08% 45.9% 60.17% 86.19% 13.75% 1.10% 2.21% 0.59% 1.47% 0.27%
12:00 192.3% 14.6% 263.6% 172.8% 144.7% 6.71% 3.80% 4.26% 3.15% 1.38% 12:00 19.85% 11.8% 68.45% 89.08% 11.24% 1.02% 2.08% 0.50% 1.87% 1.55%
13:00 184.9% 117.3% 80.5% 107.4% 119.6% 5.61% 15.75% 3.39% 2.84% 3.23% 13:00 142.16% 218.5% 20.42% 60.07% 114.61% 0.98% 2.28% 0.48% 1.78% 1.23%

J. Wang et al. / Energy 81 (2015) 627e644

14:00 168.1% 101.7% 116.0% 126.6% 133.5% 3.05% 2.21% 3.28% 3.07% 4.76% 14:00 128.46% 82.3% 0.63% 65.39% 166.70% 1.00% 1.73% 0.52% 1.34% 0.59%
15:00 300.8% 377.8% 278.7% 216.2% 198.6% 3.10% 2.03% 3.73% 3.53% 5.26% 15:00 1.57% 1.8% 48.59% 82.60% 16.34% 0.72% 1.12% 0.43% 1.08% 1.32%
16:00 215.8% 82.3% 56.3% 178.3% 224.9% 3.80% 32.97% 4.71% 4.66% 6.98% 16:00 27.66% 70.1% 33.86% 77.35% 68.45% 0.15% 1.91% 0.01% 0.31% 3.07%
17:00 149.8% 228.5% 36.4% 106.8% 134.4% 5.37% 70.04% 6.91% 6.71% 8.53% 17:00 55.49% 118.8% 26.42% 74.96% 76.25% 1.47% 5.22% 0.82% 1.31% 7.04%
18:00 63.5% 16.6% 4.2% 68.1% 95.8% 5.83% 47.68% 11.12% 10.66% 11.56% 18:00 3.30% 30.2% 14.29% 72.14% 13.03% 10.56% 26.54% 11.12% 8.30% 4.95%
Daily MRE 194.7% 173.8% 128.7% 128.4% 126.0% 13.48% 27.14% 7.33% 7.89% 7.17% Daily MRE 75.80% 104.3% 27.77% 72.91% 87.20% 1.92% 9.51% 1.70% 2.05% 3.10%
Daily RMSE 217.59 214.14 212.22 170.33 161.46 25.04 72.56 18.71 17.19 21.34 Daily RMSE 109.04 140.96 131.82 187.49 117.59 3.81 27.17 2.05 6.00 6.31
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (real sky) Aug 8 Model (clear sky) Model (real sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 1.52% 46.71% 10.28% 38.73% 22.58% 0.12% 3.65% 10.27% 2.01% 3.12% 7:00 17.46% 7.89% 4.84% 3.01% 1.08% 11.04% 12.20% 4.83% 2.56% 0.83%
8:00 18.66% 3.78% 5.54% 25.39% 5.25% 1.72% 2.00% 5.53% 4.38% 1.48% 8:00 27.68% 13.23% 4.70% 17.31% 20.48% 4.87% 5.57% 4.69% 2.33% 1.83%
9:00 17.02% 40.84% 2.87% 42.69% 28.01% 1.61% 0.96% 2.86% 3.01% 1.66% 9:00 21.06% 11.92% 2.49% 18.85% 23.31% 2.32% 0.25% 2.48% 0.41% 0.88%
10:00 29.01% 24.37% 40.00% 41.64% 25.44% 1.42% 5.25% 1.67% 4.22% 1.63% 10:00 17.25% 10.42% 1.10% 16.30% 20.33% 1.63% 1.40% 1.09% 0.81% 0.42%
11:00 16.64% 41.17% 42.25% 42.16% 26.05% 1.10% 2.90% 0.92% 1.83% 0.99% 11:00 17.92% 12.99% 0.28% 10.67% 13.00% 1.14% 1.04% 0.50% 1.23% 0.79%
12:00 28.20% 46.05% 45.93% 47.89% 33.44% 0.83% 1.03% 0.41% 1.53% 0.84% 12:00 23.23% 9.20% 0.13% 5.05% 5.44% 1.23% 0.19% 0.21% 1.52% 0.49%
13:00 16.89% 32.47% 37.03% 40.42% 13.84% 0.68% 2.06% 0.19% 1.15% 0.34% 13:00 13.71% 1.36% 0.00% 1.93% 1.26% 1.02% 0.89% 0.01% 1.80% 0.25%
14:00 23.73% 50.90% 70.17% 46.93% 31.40% 0.52% 2.46% 0.01% 1.20% 0.56% 14:00 22.22% 6.49% 41.83% 22.66% 15.09% 1.09% 1.53% 0.10% 1.67% 0.03%
15:00 12.10% 2.25% 7.27% 34.70% 17.39% 0.10% 0.86% 0.11% 0.83% 0.15% 15:00 34.03% 9.91% 43.61% 23.86% 16.11% 1.22% 0.12% 0.23% 1.57% 0.12%
16:00 42.54% 2.22% 60.38% 42.06% 11.44% 0.33% 2.11% 0.27% 1.18% 0.33% 16:00 13.48% 2.47% 0.43% 5.81% 6.37% 1.56% 0.81% 0.56% 1.30% 0.46%
17:00 12.06% 17.05% 48.56% 32.17% 12.75% 0.63% 6.80% 1.31% 1.43% 0.03% 17:00 14.17% 38.38% 1.73% 7.75% 8.55% 2.67% 4.92% 1.72% 0.27% 0.71%
18:00 5.53% 6.28% 20.16% 26.82% 7.96% 1.37% 4.88% 3.42% 2.18% 0.16% 18:00 23.00% 4.17% 44.59% 15.76% 4.82% 4.63% 11.12% 3.49% 1.30% 1.23%
19:00 190.85% 168.39% 187.18% 111.48% 167.94% 3.23% 1.08% 7.43% 2.77% 1.99% 19:00 12.19% 17.79% 5.83% 3.50% 1.40% 8.08% 10.47% 5.82% 3.24% 2.98%
Daily MRE 31.90% 37.11% 44.43% 44.08% 31.04% 1.05% 2.77% 2.64% 2.13% 1.02% Daily MRE 19.80% 11.25% 11.66% 11.73% 10.56% 3.27% 3.89% 1.98% 1.54% 0.85%
Daily RMSE 162.27 240.96 311.89 286.96 165.69 6.06 19.66 9.24 13.34 5.78 Daily RMSE 137.91 76.87 152.09 95.15 80.52 10.75 14.43 6.71 9.41 3.48
(Wh/m2) (Wh/m2)
Table 7
Results comparison by five kinds of forecasting models in Site 6.

Feb 9 Model (clear sky) Model (clear sky) Nov 6 Model (clear sky) Model (clear sky)

1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 8.18% 41.92% 0.00% 4.55% 5.30% 80.36% 23.55% 100.0% 14.59% 15.69%
8:00 122.9% 70.3% 135.7% 125.2% 99.3% 16.12% 11.43% 19.35% 21.95% 9.35% 8:00 37.45% 27.33% 51.81% 40.72% 22.52% 0.21% 5.09% 3.12% 2.32% 5.86%
9:00 132.3% 116.4% 164.0% 150.3% 116.2% 9.81% 3.08% 10.73% 10.58% 5.15% 9:00 33.58% 39.60% 43.02% 38.49% 35.41% 0.78% 7.41% 1.33% 2.70% 3.92%
10:00 113.1% 4.6% 71.5% 82.3% 112.1% 7.16% 7.38% 7.00% 7.05% 3.47% 10:00 19.22% 2.51% 14.29% 15.62% 15.00% 1.15% 0.75% 0.80% 1.06% 3.70%
11:00 99.1% 14.6% 60.0% 70.1% 98.2% 6.29% 1.34% 5.77% 6.11% 2.94% 11:00 40.90% 74.79% 56.40% 46.43% 40.88% 1.42% 7.38% 0.46% 1.21% 3.69%
12:00 156.5% 170.9% 60.2% 85.6% 154.1% 6.21% 0.47% 5.37% 5.71% 3.12% 12:00 60.01% 57.67% 12.37% 24.03% 28.45% 1.07% 4.54% 0.27% 2.05% 1.34%
13:00 195.9% 133.9% 65.2% 88.5% 151.4% 6.10% 0.55% 5.28% 5.60% 2.99% 13:00 59.12% 38.66% 19.45% 26.50% 28.76% 1.46% 0.11% 0.26% 2.07% 3.19%

J. Wang et al. / Energy 81 (2015) 627e644

14:00 145.8% 172.7% 64.2% 86.9% 148.4% 6.22% 3.86% 5.41% 5.71% 3.10% 14:00 46.66% 47.89% 33.19% 33.79% 33.08% 1.31% 2.26% 0.28% 2.25% 2.55%
15:00 219.7% 362.8% 222.1% 206.8% 169.1% 6.27% 8.69% 5.68% 5.99% 2.97% 15:00 43.10% 7.51% 26.12% 33.14% 35.49% 1.14% 5.79% 0.52% 2.53% 2.62%
16:00 200.4% 149.2% 228.6% 197.2% 117.0% 6.99% 6.08% 6.80% 7.06% 3.34% 16:00 26.04% 70.03% 1.42% 19.46% 26.85% 0.76% 5.00% 1.26% 1.36% 3.52%
17:00 213.4% 301.0% 253.3% 241.0% 211.3% 8.53% 3.52% 9.18% 9.12% 4.14% 17:00 34.71% 4.21% 12.90% 0.52% 3.88% 0.42% 1.85% 1.60% 2.96% 0.82%
18:00 174.5% 170.6% 61.3% 86.1% 152.9% 9.99% 43.18% 13.86% 14.13% 2.91% 18:00 13.52% 9.43% 17.65% 7.30% 0.99% 4.89% 13.16% 9.09% 2.74% 5.78%
Daily MRE 161.2% 151.5% 126.0% 129.1% 139.1% 8.15% 8.14% 8.59% 9.00% 3.95% Daily MRE 35.21% 35.13% 24.05% 24.21% 23.05% 7.91% 6.41% 9.92% 3.15% 4.39%
Daily RMSE 263.54 291.57 196.88 197.09 224.79 33.74 27.01 31.13 32.49 16.45 Daily RMSE 226.04 230.37 142.04 145.95 146.70 6.14 21.99 3.02 9.87 14.55
(Wh/m2) (Wh/m2)
May 10 Model (clear sky) Model (clear sky) Aug 8 Model (clear sky) Model (clear sky)
1 2 3 4 5 1 2 3 4 5 1 2 3 4 5 1 2 3 4 5

7:00 51.48% 52.57% 55.93% 16.21% 9.30% 0.29% 3.43% 2.25% 5.16% 1.92% 7:00 25.13% 27.67% 27.78% 25.13% 26.11% 15.50% 1.93% 13.73% 5.40% 0.30%
8:00 27.65% 5.05% 4.18% 14.72% 17.63% 0.67% 2.16% 2.36% 0.97% 2.32% 8:00 30.62% 35.68% 24.41% 30.60% 33.46% 17.47% 4.48% 15.75% 7.01% 0.48%
9:00 28.74% 29.05% 4.35% 18.08% 19.96% 0.62% 2.24% 1.53% 1.40% 0.93% 9:00 134.9% 69.48% 126.49% 116.02% 116.51% 11.46% 2.93% 9.61% 3.57% 0.12%
10:00 14.92% 26.16% 7.64% 7.46% 9.49% 0.98% 0.14% 1.10% 1.57% 0.53% 10:00 129.2% 27.58% 147.01% 108.12% 102.69% 7.95% 1.59% 6.03% 2.19% 0.50%
11:00 25.49% 36.53% 56.12% 20.61% 14.64% 1.84% 2.95% 1.02% 2.19% 0.37% 11:00 6.78% 12.04% 14.73% 13.67% 12.41% 6.08% 0.07% 4.12% 1.70% 0.43%
12:00 6.74% 2.12% 24.71% 7.61% 5.35% 1.28% 1.08% 1.02% 0.87% 0.21% 12:00 29.13% 13.93% 0.78% 21.35% 27.03% 5.31% 0.80% 3.33% 1.38% 0.25%
13:00 45.49% 87.32% 87.88% 57.66% 53.26% 1.89% 0.40% 1.19% 0.68% 0.41% 13:00 11.48% 8.21% 27.92% 9.23% 6.73% 5.08% 0.21% 3.09% 1.14% 0.15%
14:00 0.33% 19.57% 52.49% 11.96% 4.84% 1.65% 1.09% 1.34% 0.55% 0.15% 14:00 2.64% 15.19% 40.84% 1.50% 5.32% 5.17% 1.50% 3.18% 1.35% 0.07%
15:00 14.12% 23.99% 34.31% 6.22% 1.63% 1.60% 1.09% 1.50% 0.46% 0.09% 15:00 131.5% 71.88% 142.21% 100.35% 94.22% 5.85% 0.13% 3.88% 1.36% 0.03%
16:00 5.05% 7.42% 29.82% 0.36% 4.59% 1.50% 0.44% 1.85% 0.79% 0.03% 16:00 10.32% 67.07% 8.05% 2.77% 2.94% 7.38% 0.79% 5.44% 2.78% 0.25%
17:00 15.84% 19.04% 12.28% 11.71% 15.65% 1.23% 2.15% 2.21% 2.81% 0.34% 17:00 6.23% 42.21% 23.33% 10.61% 9.36% 10.48% 1.50% 8.60% 3.14% 0.60%
18:00 6.91% 20.62% 34.15% 0.79% 4.97% 0.36% 3.02% 2.96% 4.71% 1.28% 18:00 9.63% 3.61% 7.05% 12.35% 14.80% 16.20% 1.61% 14.44% 6.69% 0.47%
19:00 19.12% 29.28% 4.29% 15.91% 19.11% 0.85% 12.72% 4.28% 1.32% 0.85% 19:00 13.94% 24.08% 32.65% 24.12% 23.89% 17.12% 0.69% 15.39% 13.74% 0.79%
20:00 20.14% 27.60% 31.40% 14.56% 13.88% 1.14% 2.53% 1.89% 1.81% 0.73% 20:00 4.03% 339.87% 0.00% 3.03% 2.46% 2.05% 11.58% 0.00% 5.28% 2.60%
Daily MRE 51.48% 52.57% 55.93% 16.21% 9.30% 0.29% 3.43% 2.25% 5.16% 1.92% Daily MRE 38.97% 54.18% 44.52% 34.20% 34.14% 9.51% 2.13% 7.61% 4.05% 0.50%
Daily RMSE 99.63 164.40 204.43 99.47 90.87 10.55 10.61 9.63 9.36 3.27 Daily RMSE 138.68 106.42 162.93 113.60 111.11 44.04 7.42 32.73 13.68 1.84
(Wh/m2) (Wh/m2)

639
640 J. Wang et al. / Energy 81 (2015) 627e644
(RSKY Glo, unit: Wh/m2), real sky direct normal radiation (RSKY Dir,
unit: Wh/m2), real sky diffuse horizontal radiation (RSKY Dif, unit:
Wh/m2), clear sky direct normal radiation (CSKY Dir, unit: Wh/m2)
and clear sky diffuse horizontal radiation (CSKY Dif, unit: Wh/m2)
can be used as input variables. For instance, for five models in site 1,
we use eight groups of data at 10:00e19:00 on Feb 9 in 2008 and
2009 as input variables, CSKY Glo or RSKY Glo at 10:00e19:00 on
Feb 9 in 2008 and 2009 as output variables to build models, which
is training set. After building models, we apply these eight groups
of data at 10:00e19:00 on Feb 9 in 2010 as input variables, the
output variable would be the forecasting values of CSKY Glo or
RSKY Glo at 10:00e19:00 on Feb 9 in 2010, which is test set.
Table 1 shows the location of six sites. The latitude and the
longitude are in decimal degrees and negative indicates west of the
prime meridian, the time zone is the hourly offset from GMT
(Greenwich Mean Time) (Eastern Standard Time is designated as
5). Because the locations of each site are different, the length of
day in the same month is not the same. Meanwhile, only can solar
radiation in the daytime be predicted since solar radiation is pretty
weak at night.
4.2. Evaluation criterion
Two loss functions serve as the criteria to evaluate the predic-
tion performance relative to solar radiation value, namely, RMSE
(root mean square error) and MRE (mean relative error) [34,35].
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u BP neural network [36] and ARMA model [37] are competitive
u1 X N
RMSE ¼ t
2
ðy  b yiÞ (29)
N i¼1 i
N   ARMA is one of the most significant time series methods, and it is
1 X  y i 
yi  b
MRE ¼ (30)
N i¼1  yi 
where yi is the actual value, b
y i is the forecasting value and N is the
number of the forecasting samples. The forecasting performance is
better when the loss function value is smaller [34,35].
4.3. Experimental analysis and discussion
4.3.1. Experiments
This work proposes an optimized hybrid model called CS-OP-
ELM model and adopts multiple variables dataset as input vari-
ables to forecast clear sky global horizontal radiation (CSKY Glo)
and real sky global horizontal radiation (RSKY Glo). The proposed
CS-OP-ELM algorithm can be built on the basis of the OP-ELM. The
first step of the OP-ELM is to construct FFNN applying the ELM
method with a lot of neurons. The second and third steps are meant
to prune the possibly meaningless neurons of the FFNN: MRSR can
get a ranking of the neurons and the actual pruning can be per-
formed by evaluating the LOO error versus the number of neurons
used. The OP-ELM method employs a combination of three
different types of kernels including linear, sigmoid and Gaussian
kernels for robustness and more generality, while the original ELM
only uses sigmoid kernels. Based on the theory of the OP-ELM al-
gorithm, two kinds of strategies, Direct strategy and Direct-
Recursive strategy, can be used to establish a hybrid model. Then
CS (Cuckoo Search) algorithm is employed to determine the weight
parameters. The detailed optimized process of the hybrid model is
given as follows:
Step 1: Initialize the cuckoo flock and randomly generate the
initial locations xi (i ¼ 1,2,…,30) of the cuckoo's nests. In this
experiment, the number of the cuckoo's nests is 30, the number
of the cuckoo's eggs (the number of components about objective
function corresponding variables) is 2, the egg laid by a cuckoo is
discovered by the host bird with a probability pa(pa ¼ 0.2), the
maximum iteration is 500, and the expected error is 0.001.
Step 2: Calculate the objective function f(x) (x ¼ (x1,x2,/,x30)T)
(or fitness function) of each nest and record the current optimal
solution. The objective function is mean absolute error between
actual value and forecasting value.
Step3: Retain the previous generation of the optimal nest's
location, update other nest's location according to the location
update formula Eq. (4) and Eq. (5) and get the new generation.
Step 4: Compare the previous generation of the optimal nest's
location with the new generation of the optimal nest's location,
and keep the better one of the location (the value of the
objective function is the least).
Step 5: Random number R is the probability that the host bird
discovers the egg laid by a cuckoo. If R > pa, then randomly
change the location of the nest and get a new set of the nest.
Step 6: if the value of objective function is less than the expected
value 0.001, the iteration will stop, otherwise go to Step 2.
Step 7: Output the global optimal location.
Fig. 1 also indicates the whole process of the optimized hybrid
model.
4.3.2. Parameters determination of BP neural network and ARMA
model
techniques that have been widely applied for nonlinear modeling.
Specifically, BP has strong nonlinearity mapping ability that can
summarize and get the rule from acquired data automatically.
applied to analyze information for a long-term research. Conse-
quently, it is interesting to compare them with our proposed hybrid
method. However, it is crucial to determine the parameters for good
models. In BP neural network, the number of hidden layers and the
number of neuron nodes need to be adjusted. In order to get the
best results, PSO (Particle Swam Optimization) algorithm can be
used. Through experimental simulation, the hidden layer is 2, the
number of neuron nodes is 8 and 16, and the maximum iteration is
500. In ARMA (p,q) model, it is also necessary to determine the
parameters p and q. In this study, the dynamic method is employed.
Different forecasting points have different ARMA models. First we
set the maximum values of p and q (pmax ¼ 4, qmax ¼ 4), then let p
and q do cycle test from 1 to the maximum value. Finally, p and q
corresponding with the minimum AIC (Akaike Information Crite-
rion) will be considered as the optimal parameter values.
4.3.3. Results comparisons of five kinds of models
The Forecasting performances of hourly and daily MRE and daily
RMSE of five kinds of models in six sites are present in Tables 2e7.
Table 8
The percentage of the lowest MRE and RMSE in six sites.
Clear sky Real sky
Models The lowest The lowest The lowest The lowest
MRE RMSE MRE RMSE
ARMA 4.17% 8.33% 4.17% 4.17%
BP 0 8.33% 0 0
OP-ELM by 20.83% 12.5% 16.67% 16.67%
Direct strategy
OP-ELM by 0 8.3% 4.17% 8.33%
Direct-Recursive
strategy
CS-OP-ELM 75% 62.5% 75% 70.83%
 J. Wang et al. / Energy 81 (2015) 627e644 641

Fig. 3. Comparisons of MRE about five kinds of models in six sites.

642 J. Wang et al. / Energy 81 (2015) 627e644

Fig. 4. Comparisons of RMSE about five kinds of models in six sites.

 J. Wang et al. / Energy 81 (2015) 627e644
In order to illustrate the sufficiency of our experiment, four days
(Feb 9, May 10, Aug 8 and Nov 6) are selected from different seasons
for the purpose of forecasting. Due to limited space, in Tables 2e7,
Model 1 represents ARMA, Model 2 represents BP, Model 3 repre-
sents OP-ELM by Direct strategy, Model 4 represents OP-ELM by
Direct-Recursive strategy, and Model 5 represents CS-OP-ELM.
CS-OP-ELM model has the lowest MRE except for Nov 6 what-
ever it is clear sky or real sky, in site 1; in site 2, CS-OP-ELM shows
the best forecasting results in four days; in site 3, except for Aug 8,
CS-OP-ELM is the best model; in site 5, except for Nov 6, CS-OP-ELM
still has the lowest MRE in other three days. For site 4 and site 6,
clear sky forecasting and real sky forecasting reveal different
forecasting effects. Table 8 describes the percentage of the lowest
daily MRE and daily RMSE in six sites. In terms of forecasting clear
sky global horizontal radiation (CSKY Glo), CS-OP-ELM takes up 75%
of the lowest MRE and 62.5% of the lowest RMSE. For forecasting
real sky global horizontal radiation (RSKY Glo), CS-OP-ELM ac-
counts for 75% of the lowest MRE and 70.83% of the lowest RMSE.
From an overall perspective, forecasting accuracies of real sky
condition are better than that of clear sky condition. Further notice
that CS-OP-ELM model has the lowest MRE, but its RMSE is not the
lowest. This illustrates that the deviations of individual forecasting
values and real data are relatively larger. Figs. 3 and 4 draw up
histogram about MRE and RMSE of all models in Nov 6. It's easy to
see CS-OP-ELM model has the optimal forecasting effects. Taking
real sky global horizontal radiation forecasting (RSKY Glo) in site 2
as an example, the proposed model CS-OP-ELM can be compared
with other four models (see Table 2).
(1) ARMA vs. CS-OP-ELM: In terms of significant evaluation
criteria daily RMSE and daily MRE, CS-OP-ELM shows desired
higher forecasting precision than ARMA in four days. Spe-
cifically, it is observed that CS-OP-ELM leads to 5,77, 0.79,
3.54 and 7.76 reductions in RMSE (unit: Wh/m2), 1.9%, 0.85%,
1.57% and 2.5% reductions in MRE (unitless) respectively in
comparison with ARMA.
(2) BP vs. CS-OP-ELM: Evaluation criteria daily RMSE and daily
MRE from CS-OP-ELM are obviously lower than BP in four
days. Concretely, it is clear that CS-OP-ELM gives rise to 3.96,
6.59, 7.75 and 21.59 reductions in RMSE, 5.57%, 1.6%, 2.25%
and 4.93% reductions in MRE respectively in comparison
with BP.
(3) OP-ELM by Direct strategy vs. CS-OP-ELM: Based on impor-
tant evaluation criteria daily RMSE and daily MRE, CS-OP-
ELM has better forecasting performance than OP-ELM by
Direct strategy in four days. In detail, it can be seen that CS-
OP-ELM results in 17, -3.46, 6.72 and 4.57 reductions in
RMSE, 5.99%, 0.72%, 2.25% and 1.04% reductions in MRE
respectively in comparison with OP-ELM by Direct strategy.
(4) OP-ELM by Direct-Recursive strategy vs. CS-OP-ELM: Ac-
cording to crucial evaluation criteria daily RMSE and daily
MRE, CS-OP-ELM presents the better forecasting effect than
OP-ELM by Direct-Recursive strategy. In fact, it is found that
CS-OP-ELM causes 26.46, 1.36, 6.07 and 11.11 reductions in
RMSE, 7.87%, 1.24%, 1.35% and 1.24% reductions in MRE
respectively in comparison with OP-ELM by Direct-Recursive
strategy.
To sum up, for the data in different sites, five models present
different forecasting performances. But it is found that the fore-
casting precision of the proposed hybrid CS-OP-ELM model is
almost the highest in all models, following by Model 3, while BP
shows the worst forecasting results (see Table 8). In other words,
CS-OP-ELM is the most suitable model for forecasting clear sky or
real sky global horizontal radiation.
643
5. Conclusion
This paper proposes an intelligent optimized hybrid model
called CS-OP-ELM to forecast clear sky global horizontal radia-
tion (CSKY Glo) and real sky global horizontal radiation (RSKY
Glo). In order to overcome the disadvantage and limitation of
traditional models, CS-OP-ELM makes the hybrid weights change
over time using Cuckoo Search (CS) algorithm. OP-ELM model
can be first built applying the ELM method with a lot of neurons.
Then MRSR can get a ranking of the neurons and the actual
pruning can be performed by evaluating the LOO error versus
the number of neurons used. Based on two kinds of strategies,
Direct strategy and Direct-Recursive strategy, Cuckoo Search al-
gorithm (CS) is employed to determine the weight parameters.
Furthermore, for the sake of testing the availability of CS-OP-
ELM method, the forecasting accuracy has been compared with
other four models, including two common approaches (ARMA
and BP), two different strategies (OP-ELM by Direct strategy and
OP-ELM by Direct-Recursive strategy). Based on two evaluation
criteria (RMSE and MRE), CS-OP-ELM outperforms the other four
models and shows the highest forecasting precision for fore-
casting hourly and daily global horizontal radiation in clear sky
and real sky condition.
Acknowledgment
This research was supported by the National Natural Science
Foundation of China (Grant No. 71171102/G0107).
References
[1] Guo ZH, Dong Y, Wang JZ, Lu HY. The forecasting procedure for long-term
wind speed in the Zhangye area. Math Probl Eng 2010;2010:17. Article ID
684742.
[2] International Energy Agency (IEA). CO2 emission from fuel combustion
Highlights 2013. IEA Statistics; 2013. p. 158.
[3] U.S.-China joint announcement on climate change, Beijing, China. 12
November 2014.
[4] Tian Y, Zhao CY. A review of solar collectors and thermal energy storage in
solar thermal applications. Appl Energy 2013;104:538e53.
[5] Mellit A, Pavan AM. A 24-h forecast of solar irradiance using artificial neural
network: application for performance prediction of a grid-connected PV plant
at Trieste, Italy. Sol Energy 2010;84(5):807e21.
[6] Aguiar R, Collares-Pereira M. Tag: a time-dependent, autoregressive, Gaussian
model for generating synthetic hourly radiation. Sol Energy 1992;49(3):
167e74.
[7] Mora-Lopez L, de Cardona MS. Multiplicative ARMA models to generate
hourly series of global irradiation. Sol Energy 1998;63(5):283e91.
[8] Santos J, Pinazo J, Canada J. Methodology for generating daily clearness index
index values kt starting from the monthly average daily value K t determining
the daily sequence using stochastic models. Renew Energy 2003;28(10):
1523e44.
[9] Maafi A, Adane A. A two-state markovian model of global irradiation suitable
for photovoltaic conversion. Sol Wind Technol 1989;6(3):247e52.
[10] Dorvio ASS, Jervase JA, Al-Lawati A. Solar radiation estimation using artificial
neural networks. Appl Energy 2002;74:307e19.
[11] Sozen A, Arcaklýog¢lu E, Ozalp M. Estimation of solar potential in Turkey by
artificial neural networks using meteorological and geographical data. Energy
Convers Manage 2004;45:3033e52.
[12] Lopez G, Batlles FJ, Tovar-Pescador J. Selection of input parameters to model
direct solar irradiance by using artificial neural networks. Energy 2005;30:
1675e84.
[13] Elminir HK, Azzam YA, Younes FI. Prediction of hourly and daily diffuse
fraction using neural network, as compared to linear regression models. En-
ergy 2005;32:1513e23.
[14] Mohandes M, Balghonaim A, Kassas M, Rehman S, Halawani TO. Use of radial
basis functions for estimating monthly mean daily solar radiation. Sol Energy
2000;68:161e8.
[15] Sen Z. Fuzzy algorithm for estimation of solar irradiation from sunshine
duration. Sol Energy 1998;63:39e49.
[16] Yang Y, Dong Y, Chen YH, Li CH. Intelligent optimized hybrid model based on
GARCH and SVM for forecasting electricity price of New South Wales,
Australia. Abstr Appl Anal 2014;2014:9. Article ID 504064.
[17] Payne RB, Sorenson MD, Klitz K. The cuckoos. Oxford University Press; 2005.
644 J. Wang et al. / Energy 81 (2015) 627e644
[18] Chakraverty S, Kumar A. Design optimization for reliable embedded system
using cuckoo search. In: International Conference on Electronics Computer
Technology, vol. 1; 2011. p. 264e8.
[19] Yang XS, Deb S. Cuckoo search via Le
vy Flights. In: Proc. For World Congress
on Nature & Biologically Inspired Computing; 2009. p. 210e4.
[20] Sorjamaa A, Hao J, Reyhani N, Ji Y, Lendasse A. Methodology for long-term
prediction of time series. Neurocomputing 2007;70:2861e9.
[21] Sorjamaa A, Lendasse A. Time series prediction using DirRec strategy. In:
Verleysen M, editor. European symposium on artificial neural networks; 2006.
p. 143e8.
€ AM, Se
[22] Grigorievskiy A, Miche Y, Ventela
verin E, Lendasse A. Long-term time
series prediction using OP-ELM. Neural Netw 2014;51:50e6.
[23] Xue XW, Yao M, Wu ZH, Yang JH. Genetic ensemble and extreme learning
machine. Neurocomputing 2014;129(10):175e84.
[24] Huang GB, Zhou H, Ding X, Zhang R. Extreme learning machine for regression
and multiclass classification. IEEE Trans Syst Man Cybern Part B: Cybern
2012;42(2):513e29.
[25] Huang GB, Zhu QY, Siew CK. Extreme learning machine: theory and applica-
tions. Neurocomputing 2006;70(1e3):489e501.
[26] Huang GB, Wang DH, Lan Y. Extreme learning machine: a survey. Int J Mach
Learn Cybern 2011;2(2):107e22.
[27] Efron P, Hastie, Johnstone I, Tibshirani R. Least angle regression. Ann Stat
2004;32(2):407e99.
[28] Blatman G, Sudret B. Adaptive sparse polynomial chaos expansion based on
least angle regression. J Comput Phys 2011;230(6):2345e67.
€ T, Tikka J. Multiresponse sparse regression with application to multi-
[29] Simila
dimensional scaling. In: Artificial neural networks: formal models and their
applicationsdICANN, vol. 3697; 2005. p. 97e102.
[30] Chen S, Hong X, Harris CJ, Wang XX. Identification of nonlinear systems using
generalized kernel models. IEEE Trans Control Syst Technol 2005;13(3):
401e11.
[31] Dong MG, Wang N. Adaptive network-based fuzzy inference system with
leave-one-out cross-validation approach for prediction of surface roughness.
Appl Math Model 2011;35(3):1024e35.
[32] Liu XY, Li P, Gao CH. Fast leave-one-out cross-validation algorithm for
extreme learning machine. J Shanghai Jiaot Univ 2011;45(8):1140e5 [in
Chinese].
[33] National Solar Radiation Database (NSRDB). 1991e2010. Update, available
from: http://rredc.nrel.gov/solar/old_data/nsrdb/1991-2010/.
[34] Dong Y, Wang JZ, Jiang H, Wu J. Short-term electricity price forecast based on
the improved hybrid model. Energy Convers Manage 2011;52(8e9):2987e95.
[35] Dong Y, Wang JZ, Jiang H, Shi XM. Intelligent optimized wind resource
assessment and wind turbines selection in Huitengxile of Inner Mongolia,
China. Appl Energy 2013;109:239e53.
[36] Zhang Z, Wang F, Su S. Solar irradiance short-term prediction model based on
BP neural network. Energy Proc 2011;12:488e94.
[37] Voyant C, Muselli M, Paoli C, Nivet ML. Numerical weather prediction (NWP)
and hybrid ARMA/ANN model to predict global radiation. Energy 2012;39(1):
341e55.