Professional Documents
Culture Documents
Energy
journal homepage: www.elsevier.com/locate/energy
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.
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
thermal collectors to utilize the energy. Furthermore, with rapid forecasting for solar radiation. Second, the optimization algorithm
economic development, many industrial sectors need more and called CS (Cuckoo Search) optimization algorithm can be used to
more electricity power supply. Solar energy can be converted into confirm the weight parameters. With proper selection of the pa-
electricity, either using PV (photovoltaic) directly, or using CSP rameters, redundant information can be eliminated or overlooked,
(concentrated solar power) indirectly. PV converts light into a more effective model can be built. Third, the model CS-OP-ELM
electric current using the photoelectric effect and produces plenty creates admirable improvements which are satisfactory for the
of waste heat, which can be recovered for thermal use by current research. In particular, OP-ELM model can build a neural
attaching PV board with recuperating tubes filled with carrier network with large number of neurons, rank neurons and deter-
fluids. CSP systems use lenses or mirrors and tracking systems to mine the number of neurons. The optimized hybrid strategy by CS
heat a fluid such as water in a boiler to produce steam used to algorithm, based on Direct strategy and Direct-Recursive strategy,
produce power in a steam turbine coupled to an electrical can lessen system loss. So there is no surprise that the hybrid
generator [4]. model can take full advantages of these models. Fourth, the pro-
Many researchers have developed several methods to forecast posed CS-OP-ELM model is automatic in essence, and it does not
the solar radiation. These approaches can be classified into three require to make complex decision about each case of the modeling
different types [5]: (1) The first one allows forecasting the future processes.
solar radiation at the time (t þ p) based on the past observed data at The structure of this work is presented. Section 2 introduces
the time (t þ p1, t þ p2, …, t). The aim is to find a relationship the hybrid forecasting theory, CS (Cuckoo Search) optimization
between the inputs and the outputs data. So models like AR algorithm and two kinds of forecasting strategies: Direct strategy
(Autoregressive) [6], ARMA (Autoregressive moving average) [7], and Direct-Recursive strategy. Section 3 presents the methodol-
ARIMA (autoregressive integrated moving average) [8] and Markov ogy including OP-ELM (Optimally Pruned Extreme Learning
chain [9], have been found to be very suitable; (2) The second one is Machine), ELM (Extreme Learning Machine), MRSR (Multi-
based on meteorological parameters at the time (t þ p1, t þ p2, response Sparse Regression), and LOO-CV (leave-one-out cross-
…, t) such as air temperature, relative humidity, wind speed, wind validation) algorithm in detail. In Section 4, in order to illustrate
direction, cloud, sunshine duration, clearness index, pressure, etc. the effective of the optimized hybrid model, a case study about
and geographical coordinates as latitude and longitude. The aim is solar radiation data forecast in four sites of the United States can
to find a relationship between these factors and solar radiation. be simulated, BP (Back Propagation) and ARMA model can also be
MLP (multilayer perceptron) network [10e13], RBF (Radial Basis applied to compare. Finally, Section 5 concludes the results of
Function) network [14], and fuzzy logic [15] are suitable; (3) The this work.
last one combines the two previous approaches. In this case, the
input data are the past observed solar radiation data and other 2. The optimized hybrid forecasting method by CS (Cuckoo
meteorological factors at the time (t þ p1, t þ p2, …, t). To sum Search) algorithm based on two kinds of strategies
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 This work proposes a new optimized hybrid method called CS-
reasonable to employ the third one. OP-ELM. At first, two kinds of strategies based on OP-ELM (Opti-
In this paper, we use the third type of method-an optimized mally Pruned Extreme Learning Machine) can be applied to forecast
hybrid model called CS-OP-ELM to forecast solar radiation. There solar radiation. Then a hybrid model is built, and CS (Cuckoo
are four advantages of this optimized approach. To begin with, Search) algorithm is employed to determine the weight coefficients
high-dimension data can be taken into account as input datasets. of the hybrid model. Finally, the optimized hybrid model can be
Specifically, different from traditional method like ARMA model, presented as the developed new method to forecast solar radiation.
the proposed model integrates some meteorological factors to do The detailed process is presented in Fig. 1.
J. Wang et al. / Energy 81 (2015) 627e644 629
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). Levy 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
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
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)
! !
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.
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%
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%
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%
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%
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%
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%
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, of the cuckoo's eggs (the number of components about objective
unit: Wh/m2), real sky diffuse horizontal radiation (RSKY Dif, unit: function corresponding variables) is 2, the egg laid by a cuckoo is
Wh/m2), clear sky direct normal radiation (CSKY Dir, unit: Wh/m2) discovered by the host bird with a probability pa(pa ¼ 0.2), the
and clear sky diffuse horizontal radiation (CSKY Dif, unit: Wh/m2) maximum iteration is 500, and the expected error is 0.001.
can be used as input variables. For instance, for five models in site 1, Step 2: Calculate the objective function f(x) (x ¼ (x1,x2,/,x30)T)
we use eight groups of data at 10:00e19:00 on Feb 9 in 2008 and (or fitness function) of each nest and record the current optimal
2009 as input variables, CSKY Glo or RSKY Glo at 10:00e19:00 on solution. The objective function is mean absolute error between
Feb 9 in 2008 and 2009 as output variables to build models, which actual value and forecasting value.
is training set. After building models, we apply these eight groups Step3: Retain the previous generation of the optimal nest's
of data at 10:00e19:00 on Feb 9 in 2010 as input variables, the location, update other nest's location according to the location
output variable would be the forecasting values of CSKY Glo or update formula Eq. (4) and Eq. (5) and get the new generation.
RSKY Glo at 10:00e19:00 on Feb 9 in 2010, which is test set. Step 4: Compare the previous generation of the optimal nest's
Table 1 shows the location of six sites. The latitude and the location with the new generation of the optimal nest's location,
longitude are in decimal degrees and negative indicates west of the and keep the better one of the location (the value of the
prime meridian, the time zone is the hourly offset from GMT objective function is the least).
(Greenwich Mean Time) (Eastern Standard Time is designated as Step 5: Random number R is the probability that the host bird
5). Because the locations of each site are different, the length of discovers the egg laid by a cuckoo. If R > pa, then randomly
day in the same month is not the same. Meanwhile, only can solar change the location of the nest and get a new set of the nest.
radiation in the daytime be predicted since solar radiation is pretty Step 6: if the value of objective function is less than the expected
weak at night. value 0.001, the iteration will stop, otherwise go to Step 2.
Step 7: Output the global optimal location.
4.2. Evaluation criterion
Fig. 1 also indicates the whole process of the optimized hybrid
Two loss functions serve as the criteria to evaluate the predic- model.
tion performance relative to solar radiation value, namely, RMSE
(root mean square error) and MRE (mean relative error) [34,35]. 4.3.2. Parameters determination of BP neural network and ARMA
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi model
u BP neural network [36] and ARMA model [37] are competitive
u1 X N
RMSE ¼ t
2
ðy b yiÞ (29) techniques that have been widely applied for nonlinear modeling.
N i¼1 i
Specifically, BP has strong nonlinearity mapping ability that can
summarize and get the rule from acquired data automatically.
N ARMA is one of the most significant time series methods, and it is
1 X y i
yi b
MRE ¼ (30) applied to analyze information for a long-term research. Conse-
N i¼1 yi
quently, it is interesting to compare them with our proposed hybrid
method. However, it is crucial to determine the parameters for good
where yi is the actual value, b
y i is the forecasting value and N is the models. In BP neural network, the number of hidden layers and the
number of the forecasting samples. The forecasting performance is number of neuron nodes need to be adjusted. In order to get the
better when the loss function value is smaller [34,35]. best results, PSO (Particle Swam Optimization) algorithm can be
used. Through experimental simulation, the hidden layer is 2, the
4.3. Experimental analysis and discussion 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
4.3.1. Experiments parameters p and q. In this study, the dynamic method is employed.
This work proposes an optimized hybrid model called CS-OP- Different forecasting points have different ARMA models. First we
ELM model and adopts multiple variables dataset as input vari- set the maximum values of p and q (pmax ¼ 4, qmax ¼ 4), then let p
ables to forecast clear sky global horizontal radiation (CSKY Glo) and q do cycle test from 1 to the maximum value. Finally, p and q
and real sky global horizontal radiation (RSKY Glo). The proposed corresponding with the minimum AIC (Akaike Information Crite-
CS-OP-ELM algorithm can be built on the basis of the OP-ELM. The rion) will be considered as the optimal parameter values.
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 4.3.3. Results comparisons of five kinds of models
to prune the possibly meaningless neurons of the FFNN: MRSR can The Forecasting performances of hourly and daily MRE and daily
get a ranking of the neurons and the actual pruning can be per- RMSE of five kinds of models in six sites are present in Tables 2e7.
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 Table 8
The percentage of the lowest MRE and RMSE in six sites.
kernels for robustness and more generality, while the original ELM
only uses sigmoid kernels. Based on the theory of the OP-ELM al- Clear sky Real sky
gorithm, two kinds of strategies, Direct strategy and Direct- Models The lowest The lowest The lowest The lowest
Recursive strategy, can be used to establish a hybrid model. Then MRE RMSE MRE RMSE
CS (Cuckoo Search) algorithm is employed to determine the weight ARMA 4.17% 8.33% 4.17% 4.17%
BP 0 8.33% 0 0
parameters. The detailed optimized process of the hybrid model is
OP-ELM by 20.83% 12.5% 16.67% 16.67%
given as follows: Direct strategy
OP-ELM by 0 8.3% 4.17% 8.33%
Step 1: Initialize the cuckoo flock and randomly generate the Direct-Recursive
initial locations xi (i ¼ 1,2,…,30) of the cuckoo's nests. In this strategy
CS-OP-ELM 75% 62.5% 75% 70.83%
experiment, the number of the cuckoo's nests is 30, the number
J. Wang et al. / Energy 81 (2015) 627e644 641
[18] Chakraverty S, Kumar A. Design optimization for reliable embedded system € T, Tikka J. Multiresponse sparse regression with application to multi-
[29] Simila
using cuckoo search. In: International Conference on Electronics Computer dimensional scaling. In: Artificial neural networks: formal models and their
Technology, vol. 1; 2011. p. 264e8. applicationsdICANN, vol. 3697; 2005. p. 97e102.
[19] Yang XS, Deb S. Cuckoo search via Levy Flights. In: Proc. For World Congress [30] Chen S, Hong X, Harris CJ, Wang XX. Identification of nonlinear systems using
on Nature & Biologically Inspired Computing; 2009. p. 210e4. generalized kernel models. IEEE Trans Control Syst Technol 2005;13(3):
[20] Sorjamaa A, Hao J, Reyhani N, Ji Y, Lendasse A. Methodology for long-term 401e11.
prediction of time series. Neurocomputing 2007;70:2861e9. [31] Dong MG, Wang N. Adaptive network-based fuzzy inference system with
[21] Sorjamaa A, Lendasse A. Time series prediction using DirRec strategy. In: leave-one-out cross-validation approach for prediction of surface roughness.
Verleysen M, editor. European symposium on artificial neural networks; 2006. Appl Math Model 2011;35(3):1024e35.
p. 143e8. [32] Liu XY, Li P, Gao CH. Fast leave-one-out cross-validation algorithm for
€ AM, Se
[22] Grigorievskiy A, Miche Y, Ventela verin E, Lendasse A. Long-term time extreme learning machine. J Shanghai Jiaot Univ 2011;45(8):1140e5 [in
series prediction using OP-ELM. Neural Netw 2014;51:50e6. Chinese].
[23] Xue XW, Yao M, Wu ZH, Yang JH. Genetic ensemble and extreme learning [33] National Solar Radiation Database (NSRDB). 1991e2010. Update, available
machine. Neurocomputing 2014;129(10):175e84. from: http://rredc.nrel.gov/solar/old_data/nsrdb/1991-2010/.
[24] Huang GB, Zhou H, Ding X, Zhang R. Extreme learning machine for regression [34] Dong Y, Wang JZ, Jiang H, Wu J. Short-term electricity price forecast based on
and multiclass classification. IEEE Trans Syst Man Cybern Part B: Cybern the improved hybrid model. Energy Convers Manage 2011;52(8e9):2987e95.
2012;42(2):513e29. [35] Dong Y, Wang JZ, Jiang H, Shi XM. Intelligent optimized wind resource
[25] Huang GB, Zhu QY, Siew CK. Extreme learning machine: theory and applica- assessment and wind turbines selection in Huitengxile of Inner Mongolia,
tions. Neurocomputing 2006;70(1e3):489e501. China. Appl Energy 2013;109:239e53.
[26] Huang GB, Wang DH, Lan Y. Extreme learning machine: a survey. Int J Mach [36] Zhang Z, Wang F, Su S. Solar irradiance short-term prediction model based on
Learn Cybern 2011;2(2):107e22. BP neural network. Energy Proc 2011;12:488e94.
[27] Efron P, Hastie, Johnstone I, Tibshirani R. Least angle regression. Ann Stat [37] Voyant C, Muselli M, Paoli C, Nivet ML. Numerical weather prediction (NWP)
2004;32(2):407e99. and hybrid ARMA/ANN model to predict global radiation. Energy 2012;39(1):
[28] Blatman G, Sudret B. Adaptive sparse polynomial chaos expansion based on 341e55.
least angle regression. J Comput Phys 2011;230(6):2345e67.