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: Wind speed forecasting is gaining importance as the share of wind energy in electricity systems in-
Received 18 August 2020 creases. Numerous forecasting approaches have been used to predict wind speeds. However, considering
Received in revised form the differences in wind speed time-series, there is no universal approach that has proven to be accurate
16 October 2020
under all circumstances. In our study, a combined prediction system is proposed, which consists of four
Accepted 13 November 2020
Available online 19 November 2020
parts: optimal sub-model selection, point prediction based on a modified multi-objective optimization
algorithm, interval forecasting based on distribution fitting, and forecasting system evaluation. The
developed combined system integrates the merits of the sub-models and provides accurate point and
Keywords:
Artificial intelligence
interval forecasting performance. The experimental results reveal that the proposed combined fore-
Combined forecasting system casting system can provide effective wind speed point and interval forecasts. The absolute percentage
Data preprocessing error values of the proposed system for point forecasting are 2.9220%, 3.1696%, and 4.8358% at Site 1 and
Sub-model selection strategy 2.2719%, 2.5882%, and 3.4799% at Site 2 for one-, two-, and three-step forecasts, respectively. Therefore,
Wind speed forecasting the proposed system is deemed more useful for the scheduling and management of electric power
systems than other benchmark models.
© 2020 Elsevier Ltd. All rights reserved.
https://doi.org/10.1016/j.energy.2020.119361
0360-5442/© 2020 Elsevier Ltd. All rights reserved.
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Table 1
Literature review of existing wind speed prediction approaches.
WRF [9] Wind data at turbine heights of 70 or 80 m from the The hourly wind speed forecasting accuracy was improved by 8.7% in the
ground surface in Henan, Hubei, Hunan, Anhui, and calibration stage and 7.58% in the validation stage.
Jiangxi provinces
MM5 [37] Wind speed data in a wind park located in southeastern The hybrid MM5eneural network approach could obtain good short-term
Spain predictions of wind speed at specific points.
Kalman filtering [38] Two-year-long wind speed data sets provided by an Provided significant forecast improvement with respect to the wind
NWP model and two anemometric stations located in speed model direct output, especially in very short-term forecasts
eastern Liguria (Italy)
AR [39] Hourly mean wind speed data collected from three The hybrid model based on AR model and Gaussian process regression
wind farms in China (GPR) could improve the point forecasting and interval forecasting
performance.
ARMA [40] Seven years of hourly wind speed data collected at four No single model structure outperforms the others at all heights.
different heights from an observation site in Colorado,
USA
ARIMA [41] Wind speed series from wind farms in China The accuracy of the adaptive model is better than that of some well
recognized single models and some of the latest published hybrid models.
Spatial correlation model [42] Wing power dataset in June 2009 The spatial correlation model could accurately characterize the
correlations among outputs of wind turbines and reduce the error in
short-term wind power predictions.
Predictive deep convolutional neural 10 10 wind turbine array in Indiana (85.1855W, PDCNN could capture spatioetemporal correlations effectively, and it
network [43] 40.4092 N) outperformed conventional machine learning models.
Correlation-Constrained Sparsity- Wind power time series in 2006 from 25 randomly The proposed CCSC-VAR had better overall performance than both the
Controlled Vector Autoregressive selected wind farms across Denmark original SC-VAR and other benchmark methods.
(CCSC-VAR) model [44]
Extreme learning machine (ELM) [45] Wind energy datasets collected from a certain wind The developed system could not only be used as an effective tool for wind
farm situated in Chengde, Hebei Province, China energy deterministic forecasting and uncertainty analysis, but also for
other engineering application areas in the future.
Least square support vector machine Wind speed and wind power data obtained from four The proposed model provided more accurate and stable forecasts
(LSSVM) [46] databases compared to the other methods.
Long short-term memory (LSTM) [47] Four wind speed datasets collected from Liaotung The proposed model was effective for two case study datasets (wind
Peninsula, China, and a Queensland electrical power speed data and electrical load data).
load dataset
Note: With advances in forecasting technology, few researchers have adopted a single model to forecast wind speeds. In this table, different wind speed forecasting methods
used in previous research are introduced in detail, including physical forecasting methods, statistical forecasting methods, spatial correlation models, and artificial intelligence
forecasting technologies.
disadvantages of all of the abovementioned approaches, artificial initial parameters. In addition, these approaches are costly in
intelligence approaches, such as artificial neural networks (ANNs) terms of resources and running time [48].
[16e19], adaptive network-based fuzzy inference systems [20], (2) Statistical approaches are used to forecast wind speeds based
support vector machines (SVMs) [21,22], and deep learning on the assumption of linearity, which is inconsistent with the
methods [23], have been developed and applied in many fields, irregular and nonlinear characteristics of wind speed time-
including wind power forecasting [24], pollutant forecasting [25], series data [49]. Thus, the forecasting results are not always
future price forecasting [26,27], thermal conductivity forecasting satisfactory.
[28e31], and physical modeling [32e35]. However, single artificial (3) When spatial correlation approaches are employed in wind
intelligence approaches may have difficulty escaping from local speed forecasting, a significant amount of information is
optima and have low convergence rates; thus, they are not suitable required, and good forecasting performance for wind speeds
for wind speed prediction. A detailed literature review including cannot be achieved in multi-spatially correlated sites [50].
each of these model types is presented in Table 1. Moreover, the (4) Artificial intelligence approaches are effective for depicting
limitations of the abovementioned wind speed forecasting ap- the nonlinear characteristics of wind speed time-series;
proaches are summarized as follows.1 however, they are not always successful owing to inherent
properties such as easily falling into partial optima, over-
(1) Numerical simulation approaches are appropriate for con- fitting, and low convergence speeds [51].
ducting long-term rather than short-term wind speed pre-
dictions.1 The forecasting ability depends strongly on the Considering the limitations of the above approaches, individual
forecasting approaches are insufficient for accurately forecasting
wind speeds. Therefore, many studies have proposed the applica-
tion of hybrid models to predict wind speeds [52,53].
1
According to the prediction time horizon and the purpose of wind forecasting, However, we should note that wind speed data characteristics
wind speed forecasting can be roughly classified into three categories [36]: short- are different at each site, which cannot be captured by certain
term (over minutes, hours or days), medium-term (over weeks to months) and
long-term (over years) forecasting.
forecasting approaches. Thus, hybrid forecasting models may
2
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
provide excellent prediction performance for some wind speed estimation approaches are employed to enhance the IP
data, but may not be effective for the wind speed datasets in other precision.
areas. To improve the applicability, combined models (CMs) [54] (5) The PCFS focuses on wind speed forecasting with a lead time
have been developed and shown to be effective methods for wind of 10 min, 20 min, and 30 min based on real data with a
speed prediction [55]. The principle of CMs is to minimize the sum length of 21 d. Referring to real wind speed data and
of the squared error of the training set to obtain optimal weights comparative forecasting results, the PCFS can provide a
[56]. Referring to previous research, Niu et al. [57] employed a reference for the scheduling and management of smart grids
combined wind speed prediction model that integrated a data based on the wind speed forecasting results.
preprocessing method, a multi-objective optimization algorithm,
and several widely used benchmark models to enhance the wind The remainder of this paper is organized as follows: Sections 2
speed forecasting capability. Li et al. [21] proposed a variable and 3 describe the methodology and process of the PCFS. The
weighting combination strategy that combined three different experimental preparations and three experiments are presented in
hybrid approaches to forecast wind speed. The simulation results Section 4. A deeper discussion of the PCFS is introduced in Section
indicated that the proposed model was superior to the benchmark 5. Section 6 presents the conclusions.
model for wind speed prediction. However, there are typically
several challenges when a CM is used to predict wind speed: (i) 2. Relevant methods in the proposed combined forecasting
multi-objective intelligent optimization algorithms are always system
employed to determine the weight coefficients for the prediction
results of each sub-model. However, multi-objective optimization The methods included in the combined forecasting system are
algorithms are prone to local optimization, and the late conver- detailed in this subsection, including singular spectrum analysis
gence rate is slow. (ii) Once the sub-models are determined, they (SSA), the modified dragonfly algorithm (MDA), the modified multi-
are applied to all datasets. This will result in the most suitable sub- objective dragonfly algorithm (MMODA), several sub-models, and
model not being selected according to the actual wind speed the OSS strategy.
characteristics at different observation sites when determining the
sub-models of the CM. 2.1. Singular spectrum analysis
Therefore, a novel combined system incorporating four
partsdoptimal sub-model selection (OSS), point forecasting (PF), SSA, which is adopted to disintegrate a primordial sequence into
interval prediction (IP), and system evaluationdis proposed for explicable sub-sequences, has extensive applications in many fields
short-term wind speed forecasting. In the OSS, a comprehensive such as biology [58], engineering [59], climatology [60], and eco-
evaluation indicator (CEI) is adopted to determine five optimal sub- nomics [61]. The flow of SSA can be summarized as follows:
models. Then, the selected sub-models are combined using a Step 1 (Embedding): The original time series, c ¼ ðc1 ;c2 ;/;cN Þ,
modified multi-objective optimization algorithm for combined PF. is converted into a sequence, z ¼ ðz1 ; z2 ; /; zK Þ, via Eq. (1).
Finally, the characteristics of the wind speed forecasting time-
series are measured using seven distribution functions (DFs): the c ¼ ðc1 ; c2 ; /; cN Þ/z ¼ ðz1 ; z2 ; /; zK Þ (1)
extreme value (EV), gamma, logistic, log-logistic, log-normal,
Rician, and Weibull distributions. Based on the combined PF result Here, the dimension of embedding in the SSA is denoted by L.
of the proposed combined forecasting system (PCFS) and the For K ¼ N L þ 1, the L-lagged vectors can be defined as zi ¼
optimal distribution of the wind speed forecasting time-series, the ðci ; ciþ1 ; /; ciþL1 ÞT 2RL ; K ¼ N L þ 1, L2½2; N.
IP results of the PCFS are obtained. By combining these four mod- Further, z ¼ ðz1 ; z2 ; /; zK Þ will be embodied as the trajectory
ules, the PCFS can correct the shortcomings of traditional combined matrix in Eq. (2).
models and obtain more accurate forecasting results. 0 1
The contributions of this study are as follows: c1 c2 ,,, cK
B c2 c3 ,,, cKþ1 C
z ¼ ðz1 ; z2 ; /; zK Þ ¼ B
@ ,,,
C (2)
(1) A PCFS combining PF and IP is developed. Most previous ,,, ,,, ,,, A
studies only focused on wind speed PF or IP, which cannot cL cLþ1 ,,, cN
forecast wind speeds effectively and reliably. PF plays a sig- Step 2 (Singular value disintegration): For the covariance
nificant role in the scheduling and management of power matrix, s ¼ z,zT , singular value disintegration is adopted to create L
systems, and IP is important in arranging the rotary reserve
characteristic values (l1 ; l2 ;,,,; lL ) and characteristic vectors (u1 ;u2 ;
capacity, which can significantly enhance the stability of the pffiffiffiffi
,,,; uL ). If t ¼ maxði; suchthat li > 0Þ and vi ¼ zT ui li ði ¼ 1; 2; …;
electric power system and the safety of the power supply.
tÞ, then the singular value disintegration (SVD) of the trajectory
(2) In the OSS module, the CEI is proposed to determine the
matrix can be expressed as follows:
optimal sub-models. Referring to the forecasting indexes of
the sub-models, the CEI can be used to adaptively determine
z ¼ z1 þ z2 þ ,,, þ zt (3)
the optimal sub-models at multiple angles, which will ach-
ieve better forecasting results. pffiffiffiffi
where zi ¼ li ui vi , and the rank of zi is 1. Therefore, v1 ; v2 ; ,,,; vt
(3) In the PF module, an improved weight determination pffiffiffiffi
method for the CM is developed. To overcome the disad- are principal components, and ð li ; ui ; vi Þ is the characteristic
vantages of traditional multi-objective optimization algo- solution of the SVD of z.
rithms, a modified multi-objective optimization algorithm is Step 3 (Grouping): The extent will be disintegrated into subsets
adopted to determine the weights of selected sub-models to a1 ; a2 ; ,,,; am without connection to each other. For a ¼ ðn1 ; n2 ;
enhance the global search capacity and convergence speed. , , ,; np Þ, the consequence matrix, ya , is expressed as ya ¼ ya1 þ
(4) In the IP module, different DFs are used to measure the ya2 þ ,,, þ yap , and the trajectory matrix will be disintegrated as
features of the PF values. A modified intelligent optimization y ¼ ya1 þ ya2 þ ,,, þ yam .
algorithm is used to analyze the characteristics of the wind In this study, we partition the set of indices, a ¼ f1; 2;…;tg, into
speed forecasting values. Seven distributions and two two discrete subsets: a ¼ f1; 2; …; bg and a ¼ fb þ 1; b þ 2; …; tg.
3
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Here, ya is defined as follows: flies to the population inside, which is expressed as follows:
Table 3
Assessment criteria employed in this study.
Metric Equation
R2 P PM
R2 ¼ 1 M i¼1 ðPi Ai Þ= i¼1 ðPi PÞ
MAE PM
MAE ¼ jP
i¼1 i A i j =M
P
MAPE MAPE ¼ f½ M i¼1 jðPi Ai Þ =Ai j =Mg 100%
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
RMSE P
RMSE ¼ ½ M i¼1 ðPi Ai Þ =M
2
SSE PM 2
SSE ¼ i¼1 ðPi Ai Þ
pffiffiffiffiffi
Ui Ui ¼ Pi þ ðK10:5*a sÞ= M
pffiffiffiffiffi
Li Li ¼ Pi ðK0:5*a sÞ= M
P
FICP FICP ¼ ð M i¼1 ci =MÞ 100%
P
FINAW FINAW ¼ M ðU Li Þ=NR
8 i¼1 i
AWDi < ðLi Ai Þ=ðUi Li Þ; A i < Li
AMDi ¼ 0; Ai 2½Li ; Ui
:
ðAi Ui Þ=ðUi Li Þ; Ai > Ui
AWD 1 XM
AMD ¼ AMDi
NR i¼1
Note: In this table, Pi and Ai represent the i-th PF value and i-th actual value of the wind speed data, respectively; K and s
denote the quantile and scale parameters of the DFs, respectively; M denotes the testing sample number; and NR represents
the range of forecasting values. If the i-th actual value Ai 2½Li ; Ui , then ci ¼ 1; otherwise, ci ¼ 0.
6
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
on MMODA is employed to obtain the optimal weights of the five Part III: Interval forecasting. In Part II, we obtain 720 com-
selected sub-models, which is important for improving the pre- bined predicted values from the weighted calculation of the pre-
diction accuracy and stability simultaneously. Furthermore, the diction results of each optimal sub-model. Because the distribution
first 576 forecasting values of each optimal sub-model are of 144 of the predicted values is unknown, seven basic DFs, i.e., the
employed to determine the weight coefficients for the integration EV, gamma, logistic, log-logistic, log-normal, Rician, and Weibull
of the five sub-models. In the MMODA, the first 576 of the 720 DFs, are employed to measure the real-time data characteristics of
predicted values from each SSA-based sub-model (obtained in Part the first 576 values. In this study, the optimal distribution of the
I) are adopted as the training set to determine the weights, and the first 576 values is applied as the optimal distribution of the
remaining 144 predicted values are adopted as the testing set. Then, remaining 144 values. The optimal distribution of the wind speed
the eventual prediction consequences are obtained by combining varies with time, and thus the distributions should be fitted in real-
the abovementioned hybrid models via the optimal weights. time for each forecast. Moreover, the effectiveness of the fitting
Therefore, we can obtain the combined forecasting time-series, depends on the estimated accuracy of the distribution parameter.
which contains 720 values. The first 576 values are the training Hence, the maximum likelihood estimation (MLE) and MDA are
fitting values of the PCFS, and the remaining 144 values are the final used to determine the parameters of the DFs. Considering that the
forecast values of the PCFS. The detailed process of Part II is shown goodness-of-fit index (R2) is sufficiently effective to represent the
in Fig. 3. The optimization objectives of the MMODA are as follows: distribution fitting performance, a single-objective optimization
algorithm, MDA rather than MMODA, is adopted to obtain the
8 N
> 1 X
bi
xi x
100%
optimal parameters of the DFs. In this study, the lower and upper
> f
> 1 ðxÞ ¼ MAPE ¼
< N i¼1
xi
bounds of the wind speed forecasting sequence are established by
min (21) matching the PF time series and optimal fitting DF. The PCFS pro-
>
>
>
: f ðxÞ ¼ std x vides informative and reliable IP performance. The detailed process
2 b i x i ; i ¼ 1; 2; …; N of Part III is shown in Fig. 4.
Part IV: System evaluation. The PF effectiveness of each model
b i denotes the i-th pre-
where xi denotes the i-th real value, and x is assessed using four indicators, while three indicators are
dicted value. employed to evaluate the IP effectiveness (see Table 3).
7
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
4. Experimental setup and analysis benchmark models is compared to determine five optimal sub-
models for the subsequent combined prediction. For the seven
In this section, the experimental setup and analysis are sub-models considered, the numbers of input and output nodes are
described in detail, including the experimental dataset, employed 5 and 1, respectively. In addition, for ARIMA and LSTM, to train the
forecasting criteria, and empirical analyses. neural networks fully, the number of iterations is set to 10000. The
key parameters, p and q, of ARIMA are determined via the Akaike
4.1. Experimental dataset information criterion (AIC) and Bayesian information criterion
(BIC). For LSTM, to reduce the running time, the number of itera-
Experimental datasets were collected from two wind speed tions is set to 2500. The sub-model comparison and selection re-
observation points in the Shandong Peninsula, China. The sampling sults are presented in Table 4.
lag is 10 min for each dataset, signifying that there are 144 records For Site 1, SSA-ELM has the best one-step prediction. When the
per day. Table 2 lists the dominant statistical characteristics of the number of forecasting steps is increased, the forecasting ability of
two datasets. SSA-DBN is superior to that of the other hybrid sub-models. This
indicates that different prediction step sizes will have different
optimal sub-models. The BP, DBN, ELM, and LSTM based on SSA can
4.2. Performance verification
always achieve high forecasting accuracy for one-step to three-step
predictions. Thus, they provide good prediction ability as sub-
Multiple assessment indexes are employed to compare the
models in the PCFS. SSA-ENN is added to the PCFS for the one-
prediction ability of the PCFS and other models [74]. In our study,
step and three-step predictions, while SSA-ARIMA is added to the
the PF accuracy and stability are assessed via four indexes, whereas
PCFS for two-step forecasting. This demonstrates that the OSS
the IP ability is assessed via three indexes. The model evaluation
strategy operates effectively under different forecasting
indexes adopted in this study are listed in Table 3.
mechanisms.
Furthermore, the forecasting abilities of the sub-models at Site 2
4.3. Experiment I: sub-model selection: comparison of the sub- are similar to those at Site 1; SSA-GRNN, which has CEI values of
model prediction performance 1.000 for one-step and multi-step predictions, is not appropriate for
predicting wind speed.
In this experiment, the forecasting performance of seven
8
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Table 4
Forecasting performance results of each sub-model in Experiment I.
MAE MAPE RMSE SSE CEI MAE MAPE RMSE SSE CEI MAE MAPE RMSE SSE CEI
Site 1 SSA-ARIMA 0.2711 4.3519 0.3342 96.6228 0.3714 0.3459 6.4148 0.4492 145.2782 0.3993 0.5589 10.0163 0.7219 375.2347 0.8632
SSA-BP 0.2148 3.4119 0.2813 56.9677 0.0729 0.2811 4.6217 0.3659 96.3908 0.1619 0.4401 7.3336 0.5785 240.9404 0.5060
SSA-DBN 0.2029 3.2232 0.2661 50.9850 0.0111 0.2597 4.1384 0.1228 88.4384 0.0000 0.3546 6.5633 0.4823 160.4746 0.0000
SSA-ELM 0.2005 3.1512 0.2651 50.5928 0.0000 0.2760 4.4877 0.3620 94.3670 0.1467 0.4380 7.2626 0.5779 240.4700 0.5009
SSA-ENN 0.2359 3.8529 0.3056 67.2627 0.1873 0.3590 5.8847 0.4780 164.4766 0.4184 0.5160 8.7737 0.6801 333.0130 0.7313
SSA-GRNN 0.3747 6.2910 0.4861 170.1230 1.0000 0.5277 9.2113 0.6757 328.7276 1.0000 0.6101 10.5917 0.7840 442.5257 1.0000
SSA-LSTM 0.2293 3.6794 0.3030 66.1217 0.1588 0.3128 4.9766 0.4252 130.1950 0.2711 0.4470 7.1657 0.6227 279.1916 0.5497
Site 2 SSA-ARIMA 0.3590 4.5007 0.4735 161.4236 0.7445 0.3782 4.6981 0.5475 215.8044 0.5007 0.5423 7.8186 0.7431 397.5700 0.4528
SSA-BP 0.2157 2.8414 0.2935 62.0087 0.0843 0.2727 3.5703 0.3696 98.3625 0.1292 0.5474 7.5106 0.7496 404.5852 0.4515
SSA-DBN 0.1992 2.5567 0.2781 55.6924 0.0138 0.3007 4.1259 0.4028 116.7934 0.2259 0.4441 5.8962 0.6031 261.8901 0.2799
SSA-ELM 0.1955 2.5049 0.2753 54.5492 0.0000 0.2750 3.6085 0.3723 99.7835 0.1366 0.4441 5.8969 0.6048 263.3342 0.2809
SSA-ENN 0.2206 2.9229 0.3017 65.5372 0.1112 0.3867 5.1937 0.5300 202.2717 0.5206 0.5387 7.3329 0.7353 389.3289 0.4343
SSA-GRNN 0.3943 5.3847 0.5360 206.8518 1.0000 0.5201 7.1187 0.6995 352.3336 1.0000 0.8960 13.8774 1.0887 853.4446 1.0000
SSA-LSTM 0.2791 3.5566 0.1627 117.1362 0.1913 0.2605 3.3452 0.1401 100.8901 0.0025 0.3951 5.3393 0.5967 219.9902 0.0000
Remark. In this study, the OSS strategy can select optimal sub- three hyper-parameters of CEEMD are set as 0.05, 50, and 500, and
models adaptively to ensure the forecasting precision of the PCFS. the two hyper-parameters (window length and primary ingredient
disintegration number) of SSA are set as 50 and 20, respectively.
The PF results are presented in Table 5 and Fig. 5.
4.4. Experiment II: point forecasting: comparison of the PCFS and Specifically, the performances of the benchmark forecasting
reference models models, including the persistence model, are similar and inferior to
that of the PCFS. The MAPE values in the one-step forecasts range
To verify the PCFS, several benchmark models and CMs based on from 5% to 8%, and the MAPE values increase to some extent in the
different data processing techniques and optimization algorithms two-step and three-step forecasts. Moreover, the forecasting ability
are used for comparison with the proposed PCFS, including of the CMs based on EMD, CEEMD, MODA, and MOGOA are better
empirical mode decomposition (EMD), complete ensemble empir- than those of the benchmark forecasting models, including the
ical mode decomposition (CEEMD), MODA, and the multi-objective persistence model, but worse than that of the PCFS. It is worth
grasshopper optimization algorithm (MOGOA). For the three multi- noting that the mean MAPE values of the CEEMD-MMODA-CM and
objective optimization algorithms considered, the number of iter- SSA-MOGOA-CM for the one-step, two-step, and three-step pre-
ations is set as 500, the archive size is set as 200, and the population dictions are 3.0952%, 3.9509%, and 4.7803% and 2.6889%, 3.2796%,
number is set as 50. EMD has no parameters that need to be set. The
Table 5
Comparison of the PF performance of the PCFS and benchmark models.
Models 1-Step Forecasting based on Site 1 2-Step Forecasting based on Site 1 3-Step Forecasting based on Site 1
MAE MAPE RMSE SSE MAE MAPE RMSE SSE MAE MAPE RMSE SSE
ARIMA 0.4426 7.3228 0.5786 241.0027 0.5306 9.2221 0.6914 344.2251 0.6050 10.5668 0.7797 437.6642
BP 0.4238 6.9192 0.5482 216.4159 0.5727 9.9995 0.7300 383.6550 0.6391 11.2849 0.8128 475.6569
DBN 0.4158 6.6755 0.5475 215.7874 0.4248 6.7568 0.3098 223.0368 0.5546 8.8483 0.5260 378.6903
ELM 0.4133 6.7153 0.5403 210.1576 0.5057 8.4661 0.6692 322.4004 0.5829 9.8805 0.7578 413.4213
ENN 0.4112 6.6819 0.5390 209.1715 0.5125 8.5817 0.6749 327.9648 0.5812 9.8964 0.7556 411.0331
GRNN 0.4318 7.0539 0.5706 234.4383 0.5344 9.1468 0.6983 351.0695 0.6397 11.5324 0.8125 475.3453
LSTM 0.4214 6.7090 0.3041 218.9556 0.5051 8.4240 0.6721 325.1950 0.5826 9.9136 0.7627 418.8074
Persistence Model 0.3056 6.6930 0.3900 21.9000 0.4285 9.4054 0.5550 44.3500 0.4854 10.6989 0.6314 57.4100
EMD-MMODA-CM 0.2221 4.7358 0.2692 10.4344 0.2305 4.9427 0.2807 11.3482 0.2553 5.5099 0.3217 14.8984
CEEMD-MMODA-CM 0.1562 3.4178 0.1932 5.3723 0.2110 4.5937 0.2687 10.3954 0.2485 5.3598 0.3255 15.2569
SSA-MODA-CM 0.1395 3.0629 0.1720 4.2583 0.1692 3.6349 0.2218 7.0837 0.2429 5.0853 0.3061 13.4892
SSA-MOGOA-CM 0.1415 3.0412 0.1839 4.8702 0.1650 3.5716 0.2116 6.4494 0.2383 5.0639 0.2865 11.8233
PCFS 0.1347 2.9220 0.1722 4.2724 0.1472 3.1696 0.1923 5.3244 0.2279 4.8358 0.2914 12.2280
Models 1-Step Forecasting based on Site 2 2-Step Forecasting based on Site 2 3-Step Forecasting based on Site 2
MAE MAPE RMSE SSE MAE MAPE RMSE SSE MAE MAPE RMSE SSE
ARIMA 0.4519 5.9896 0.6097 267.6592 0.5927 7.5694 0.8050 466.6241 0.6339 8.9150 0.8303 496.3977
BP 0.4119 5.4399 0.5669 231.3914 0.5283 7.1029 0.7196 372.8134 0.6078 8.4614 0.8004 461.3135
DBN 0.4111 5.3327 0.5777 240.2924 0.5234 7.0459 0.7149 367.9652 0.5634 7.3053 0.6227 448.3214
ELM 0.4086 5.3540 0.5692 233.2812 0.5261 7.0755 0.7182 371.4142 0.6061 8.3474 0.7998 460.6030
ENN 0.4121 5.4864 0.5684 232.6435 0.5331 7.2220 0.7214 374.6787 0.6108 8.4434 0.8050 466.5272
GRNN 0.4902 6.3091 0.6617 315.2218 0.8247 12.4417 0.9951 712.9374 0.6458 8.9792 0.8361 503.3797
LSTM 0.4130 5.3348 0.3385 243.7030 0.4147 5.3479 0.3443 247.9294 0.6039 8.1944 0.8059 467.6700
Persistence Model 0.2958 5.3166 0.3744 20.1800 0.4069 7.2774 0.5332 40.9400 0.4674 8.4284 0.6289 56.9500
EMD-MMODA-CM 0.1843 3.2596 0.2292 7.5660 0.1974 3.5029 0.2457 8.6923 0.2422 4.2577 0.3142 14.2150
CEEMD-MMODA-CM 0.1567 2.7727 0.1923 5.3247 0.1851 3.3081 0.2349 7.9455 0.2357 4.2009 0.3052 13.4168
SSA-MODA-CM 0.1338 2.4158 0.1662 3.9795 0.1755 3.2115 0.2166 6.7542 0.2166 3.8866 0.2659 10.1795
SSA-MOGOA-CM 0.1297 2.3367 0.1610 3.7328 0.1664 2.9876 0.2091 6.2938 0.2051 3.8020 0.2581 9.5946
PCFS 0.1260 2.2719 0.1572 3.5587 0.1435 2.5882 0.1747 4.3956 0.1946 3.4799 0.2412 8.3759
9
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
and 4.4330%, respectively, while the corresponding values for the for policy makers in conducting risk management and evaluations
PCFS are 2.5969%, 2.8789%, and 4.1579%, respectively. Thus, the of electrical systems to ensure their dependability and operability.
PCFS can improve the wind speed forecasting performance to Thus, IP is carried out based on the PF results to provide more in-
varying degrees with high forecasting stability and accuracy. formation for wind farm operation and management. To verify the
superiority of the PCFS for IP, four CMs are selected as benchmark
Remark. Compared with the benchmark models and the CMs
models because the PF results of these four CMs were superior to
based on various data processing techniques and optimization al-
those of the other reference models.
gorithms, PCFS always provides the best prediction ability, which
To improve the IP accuracy and evaluate an appropriate fore-
verifies that the combined strategy of the PCFS is credible and valid
casting interval, seven DFs were used to fit the wind speed pre-
and can be adopted for wind speed prediction.
diction series. MDA and MLE were used to determine the
parameters of the seven DFs. Considering that R2 is sufficiently
4.5. Experiment III: Interval forecasting: comparison of the PCFS effective to represent the distribution fitting performance, MDA
and reference models rather than MMODA is used to obtain the optimal parameters of the
DFs. Based on R2, the optimal distribution of the forecast values of
Compared with PF, IP results contain more data and are useful the wind speed series can be obtained to determine the forecasting
10
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
interval. The optimal distributions of the four CMs and the PCFS at three-step predictions.
each site and for each forecasting step can be determined based on
Remark. The comparative results show that the PCFS can provide
the R2 values. The R2 values of the distributions for the PCFS are
the best IP results relative to the reference models.
presented in Appendix A. From Table 1 in Appendix A, we can
observe that at Site 1, the Weibull distribution is the optimal dis-
tribution for the PCFS, whereas at Site 2, the log-logistic distribu- 5. Discussion
tion is the most appropriate distribution, and these can be used to
determine the forecasting interval. In this section, the improvements of the PCFS, sensitivity anal-
FINAW, FICP, and AWD were employed to evaluate the IP ability ysis, and applications in power systems are discussed.
of the PCFS and four CMs. The abovementioned indicators and
lower and upper limits of the IP are listed in Table 3; the IP results 5.1. Improvements of the proposed system
are presented in Fig. 6 and Tables 6 and 7. Moreover, the expecta-
tion probability is set at 90%, 80%, 70%, and 60%, expressed as ð1 To investigate the improvement provided by the PCFS, four
aÞ 100%, to evaluate the IP ability of each model. In Tables 6 and 7, criteria are used to analyze the precision of the PCFS [75]. Detailed
the PCFS achieves the best IP results in all datasets for one-step to definitions of these four criteria are listed in Table 8. Moreover, by
11
P. Jiang, Z. Liu, X. Niu et al.
Table 6
Comparison of the IP performance between the PCFS and reference models at Site 1.
One-Step 90% 40.9722 0.2676 0.0503 62.5000 0.1723 0.0297 65.2778 0.1517 0.0199 63.8889 0.1495 0.0070 78.0556 0.1388 0.0018
80% 44.4444 0.2762 0.0355 67.3611 0.1785 0.0213 72.9167 0.1506 0.0127 68.0556 0.1561 0.0048 84.3056 0.1411 0.0008
70% 45.8333 0.2864 0.0288 68.7500 0.1867 0.0176 79.8611 0.1461 0.0094 71.5278 0.1582 0.0038 87.0833 0.1446 0.0002
60% 47.2222 0.2917 0.0243 71.5278 0.1881 0.0148 82.6389 0.1475 0.0079 72.9167 0.1624 0.0032 89.1667 0.1472 0.0002
Two-Step 90% 39.5833 0.2800 0.0869 53.4722 0.1996 0.1036 56.9444 0.1818 0.0287 59.0278 0.1630 0.0079 71.1111 0.1526 0.0070
80% 45.1389 0.2732 0.0598 56.2500 0.2119 0.0798 61.8056 0.1854 0.0203 65.9722 0.1625 0.0050 76.6667 0.1560 0.0046
70% 46.5278 0.2823 0.0496 60.4167 0.2106 0.0652 65.9722 0.1844 0.0161 70.8333 0.1613 0.0037 80.1389 0.1582 0.0036
60% 50.0000 0.2749 0.0410 61.8056 0.2158 0.0579 66.6667 0.1906 0.0140 70.8333 0.1689 0.0032 82.2222 0.1610 0.0030
Three-Step 90% 38.8889 0.2961 0.1656 48.6111 0.2149 0.0879 38.1944 0.2463 0.0282 33.3333 0.2706 0.0138 65.1389 0.2106 0.0104
80% 42.3611 0.3009 0.1253 51.3889 0.2265 0.0683 43.7500 0.2373 0.0192 39.5833 0.2521 0.0084 71.3889 0.2055 0.0030
70% 43.7500 0.3095 0.1077 52.7778 0.2352 0.0589 44.4444 0.2477 0.0162 40.9722 0.2586 0.0068 73.4722 0.2103 0.0019
60% 45.8333 0.3086 0.0940 54.1667 0.2400 0.0524 47.2222 0.2431 0.0136 43.0556 0.2569 0.0056 75.5556 0.2117 0.0016
12
Table 7
Comparison of the IP performance between the PCFS and reference models at Site 2.
One-Step 90% 84.7222 0.1743 0.0004 91.6667 0.1597 0.0008 96.5278 0.1525 0.0004 96.5278 0.1506 0.0005 97.2222 0.1475 0.0004
80% 86.1111 0.1862 0.0002 94.4444 0.1684 0.0004 97.2222 0.1603 0.0003 97.9167 0.1611 0.0003 99.3056 0.1598 0.0002
70% 89.5833 0.1881 0.0001 95.1389 0.1757 0.0002 98.6111 0.1680 0.0002 98.6111 0.1681 0.0002 99.3056 0.1653 0.0001
60% 90.9722 0.1920 0.0001 96.5278 0.1796 0.0002 98.6111 0.1739 0.0001 99.3056 0.1730 0.0002 99.3056 0.1712 0.0001
Two-Step 90% 83.3333 0.1792 0.0023 83.3333 0.1846 0.0055 87.5000 0.1576 0.0006 87.5000 0.1761 0.0029 94.4444 0.1559 0.0005
80% 87.5000 0.1854 0.0016 86.8056 0.1919 0.0037 90.9722 0.1662 0.0003 88.8889 0.1863 0.0021 95.8333 0.1656 0.0003
70% 88.8889 0.1918 0.0012 90.2778 0.1936 0.0029 91.6667 0.1733 0.0002 90.2778 0.1916 0.0017 97.9167 0.1705 0.0001
60% 90.9722 0.1943 0.0010 90.2778 0.2005 0.0024 93.7500 0.1765 0.0001 90.2778 0.1978 0.0014 97.9167 0.1761 0.0001
Three-Step 90% 70.8333 0.2228 0.0121 72.2222 0.2196 0.0120 76.3889 0.1913 0.0040 84.0278 0.1834 0.0087 84.7222 0.1769 0.0037
80% 77.0833 0.2200 0.0095 75.0000 0.2276 0.0095 79.8611 0.1970 0.0031 84.7222 0.1902 0.0067 86.8056 0.1899 0.0029
Table 8
Four improvement percentage indexes.
Note: MAE1 , MAPE1 , RMSE1 , and SSE1 are the values of the PCFS prediction accuracy evaluation indexes, and MAE2 , MAPE2 , RMSE2 , and SSE2 are the values
for the reference models.
calculating the mean error index values for each site, the in the included algorithms is discussed. Table 10 lists the assess-
improvement percentages are calculated, as summarized in Table 9. ment metrics used to measure sensitivity, which are defined by
From the results, we can observe the following: calculating the standard deviation of each error criterion [5]. It is
clear that the robustness will decrease with an increase in the
(a) The PCFS can significantly enhance the forecasting precision assessment metrics. Table 11 lists the assessment metric values
relative to the individual models and CMs, as the PCFS has calculated by changing one parameter of the SSA or MMODA, while
the best evaluation criteria values for the one-step to three- keeping the remaining parameters constant.
step predictions. Furthermore, the improvement percent- A detailed analysis is conducted in two ways: one transforms the
age of the PCFS over the benchmark models is larger than parameters in the SSA and the other transforms the parameters in
that over the CMs using different data preprocessing tech- the MMODA. Specifically, the value of the window length in the SSA
niques and optimization algorithms. is varied as 40, 45, 50, 55, and 60, while the value of the primary
(b) Specifically, compared with the individual forecasting stra- ingredient disintegration number is varied as 10, 15, 20, 25, and 30.
tegies, the IMAE, IMAPE, IRMSE, and ISSE values are very large. For Similarly, in the MMODA, the dragonfly number is varied as 30, 50,
the one-step prediction, the IMAPE values are 60.9845%, 70, and 90; the number of iterations is defined as 100, 150, 200, and
57.9750%, 56.7470%, 56.9659%, 57.3160%, 61.1325%, 56.8749%, 250; and the archive size is 300, 400, 500, 600, and 700. The specific
and 56.7523%, indicating that the forecasting ability of the analysis of Table 11 is as follows.
PCFS is better than that of the benchmark models.
(c) The improvement percentages of the PCFS over the EMD- (1) The robustness of the developed system is suitable for vari-
MMODA-CM and CEEMD-MMODA-CM are similar. More- ations in the window length and primary ingredient disin-
over, the improvement in the prediction performance of the tegration number in the SSA. For instance, in the one-step
PCFS over the SSA-MODA-CM and SSA-MOGOA-CM is prediction at Site 1, the evaluation indexes of the window
smaller than that over the benchmark models and CMs with length are 0.0015, 0.0273, 0.0017, and 0.0861, which are
various data pretreatment technologies. small and vary over a small range. The same applies to Site 2.
(2) When the parameters in the MMODA are varied, the SMAE,
SMAPE, SRMSE, and SSSE values still vary within a small range,
which confirms that the sensitivity of the developed system
5.2. Sensitivity analysis to optimization parameters is low, and the steadiness of the
PCFS is satisfactory.
In this section, the sensitivity of the PCFS to parameter changes
Table 9
Improvement percentages of the PCFS over reference models.
IMAE IMAPE IRMSE ISSE IMAE IMAPE IRMSE ISSE IMAE IMAPE IRMSE ISSE
ARIMA 70.8578 60.9845 72.2746 98.4605 74.1192 65.7095 75.4755 98.8013 65.9009 57.3153 66.9201 97.7942
BP 68.8077 57.9750 70.4568 98.2512 73.5982 66.3331 74.6815 98.7151 66.1181 57.8872 66.9869 97.8010
DBN 68.4745 56.7470 70.7194 98.2830 69.3422 58.2847 64.1829 98.3553 62.2112 48.5210 53.6333 97.5086
ELM 68.2842 56.9659 70.3056 98.2340 71.8266 62.9519 73.5472 98.5990 64.4684 54.3791 65.8073 97.6426
ENN 68.3370 57.3160 70.2507 98.2275 72.1991 63.5664 73.7158 98.6166 64.5584 54.6575 65.8716 97.6521
GRNN 71.7259 61.1325 73.2651 98.5753 78.6117 73.3291 78.3269 99.0865 67.1357 59.4585 67.6964 97.8948
LSTM 68.7550 56.8749 48.7300 98.3074 68.3940 58.1913 63.8919 98.3040 64.3926 54.0771 66.0478 97.6758
Persistence Model 56.6528 56.7523 56.8967 81.3901 65.2031 65.4862 66.2732 88.6035 55.6588 56.5243 57.7414 81.9833
EMD-MMODA-CM 35.8571 35.0392 33.8989 56.4951 32.0754 31.8239 30.2824 51.4980 15.0741 14.8641 16.2404 29.2290
CEEMD-MMODA-CM 16.6763 16.0982 14.5274 26.7923 26.5955 27.1325 27.1211 47.0034 12.7431 13.0216 15.5629 28.1437
SSA-MODA-CM 4.5940 5.1982 2.5878 4.9379 15.6826 15.8999 16.2795 29.7580 8.0538 7.3140 6.8820 12.9488
SSA-MOGOA-CM 3.8497 3.4209 4.4811 8.9724 12.2856 12.2169 12.7619 23.7234 4.7065 6.2062 2.2194 3.8008
Table 10
Evaluation criteria for the sensitivity analysis.
SMAE Std deviation of the MAE for n predictions SMAE ¼ StdðMAE1 ; MAE2 ; …; MAEn Þ
SMAPE Std deviation of the MAPE for n predictions SMAPE ¼ StdðMAPE1 ; MAPE2 ; …; MAPEn Þ
SRMSE Std deviation of the RMSE for n prediction SRMSE ¼ StdðRMSE1 ; RMSE2 ; …; RMSEn Þ
SSSE Std deviation of the SSE for n predictions SSSE ¼ StdðSSE1 ; SSE2 ; …; SSEn Þ
13
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Table 11
Sensitivity analysis results for the PCFS.
1-Step SSA Window Length 0.0015 0.0273 0.0017 0.0861 0.0008 0.0139 0.0007 0.0324
Primary Ingredient Disintegration Number 0.0017 0.0295 0.0016 0.0781 0.0007 0.0127 0.0009 0.0410
MMODA Dragonfly Number 0.0016 0.0285 0.0017 0.0854 0.0027 0.0439 0.0027 0.1274
Iteration Number 0.0018 0.0319 0.0017 0.0865 0.0025 0.0454 0.0026 0.1202
Archive Size 0.0018 0.0315 0.0023 0.1166 0.0015 0.0278 0.0025 0.1158
2-Step SSA Window Length 0.0035 0.0825 0.0046 0.2987 0.0025 0.0516 0.0036 0.2249
Primary Ingredient Disintegration Number 0.0065 0.1305 0.0066 0.4389 0.0047 0.1060 0.0082 0.5068
MMODA Dragonfly Number 0.0024 0.0453 0.0017 0.1109 0.0080 0.1384 0.0118 0.7204
Iteration Number 0.0050 0.0980 0.0057 0.3682 0.0018 0.0313 0.0025 0.1466
Archive Size 0.0022 0.0515 0.0031 0.1967 0.0027 0.0435 0.0038 0.2300
3-Step SSA Window Length 0.0074 0.1502 0.0069 0.6400 0.0048 0.0701 0.0034 0.2756
Primary Ingredient Disintegration Number 0.0077 0.2001 0.0052 0.4810 0.0033 0.0377 0.0034 0.3257
MMODA Dragonfly Number 0.0052 0.1160 0.0060 0.4958 0.0084 0.0727 0.0056 0.5777
Iteration Number 0.0031 0.0611 0.0046 0.4245 0.0055 0.1020 0.0071 0.6153
Archive Size 0.0018 0.0678 0.0041 0.3361 0.0040 0.0747 0.0053 0.4892
5.3. Applications in power systems model, CEEMD-MMODA-CM, and SSA-MOGOA-CM are 98.3520%,
98.2558%, 98.0490%, 98.1585%, 98.1654%, 98.0957%, 83.9923%,
A stable wind power system depends, to a large extent, on wind 33.9798%, and 12.1655%, respectively. For IP, when the expectation
speed predictions [76]. To enhance existing wind power systems, probability is 90%, the mean FICP values are 71.4352 and 92.1296 for
the PCFS is proposed by integrating PF and IP to realize forecasting Sites 1 and 2, respectively, while the mean FINAW values are 0.1673
precision and stability for safe wind power scheduling and opera- and 0.1601 and the mean AWD values are 0.0064 and 0.0015,
tion. The roles of PF and IP in the power system are as follows: respectively. Based on simulation analyses, we found that the PCFS
is superior to each of the reference methods in terms of PF and IP. To
(1) Precise prediction of wind speeds is fundamental for accu- verify the impact of parameter changes on the prediction perfor-
rately determining the upper bound of the wind energy mance of the PCFS, a sensitivity analysis is conducted, as presented
output. Considering that wind energy is proportional to the in the discussion section, which further verifies the stability of our
wind speed, the PF precision for wind speed must be proposed combined system. Moreover, the practical applications of
improved, which would allow wind power productivity to be PF and IP in power systems are introduced to help decision makers
estimated and an intelligent grid to be effectively pro- launch an appropriate number of wind turbines and arrange the
grammed by policy makers [77]. rotary reserve capacity. In summary, PCFS successfully improves
(2) Further, to effectively schedule and manage power systems, short-term wind speed prediction precision and stability and can
the wind speed PF precision should be improved. Superflu- provide assistance in power grid programming.
ous electrical production is detrimental in actual application There are some shortcomings of our proposed combined sys-
scenarios and may lead to a decrease in supply quality, un- tem. For instance, other influencing factor that might affect wind
safe power systems, and increased operational costs [78]. speed forecasting are neglected, and only the application of the
Therefore, it is essential to develop a more precise and PCFS to power systems is considered. Thus, considering additional
steadier wind speed prediction system to aid policy makers factors and more application fields are the two main directions for
in formulating decisions in season; consequently, the future research. Financial time series modeling [79], stock index
abovementioned shortcomings could be addressed. forecasting [80], and policy effectiveness assessment [81] may be
(3) Wind speed PF errors may cause over- or under-estimation of key areas in the field of economic management.
the number of wind turbines launched. To overcome this
shortcoming, IP is developed to facilitate the arrangement of
the rotary reserve capacity, which can significantly enhance Credit author statement
the stability of the electrical system and the security of the
electrical supply. Ping Jiang: Conceptualization, Supervision, Writing-Reviewing
and Editing. Zhenkun Liu: Software, Writing e original draft. Xin-
song Niu: Software. Lifang Zhang: Methodology.
6. Conclusion
In this study, the PCFS was proposed, which integrates optimal Declaration of competing interest
sub-model selection, point forecasting, and interval forecasting.
Moreover, three experiments and three analyses were conducted The authors declare that they have no known competing
based on two datasets from the Shandong Peninsula in China. For financial interests or personal relationships that could have
PF, the MAPE values for the one-step, two-step, and three-step appeared to influence the work reported in this paper.
predictions of the PCFS are 61.1325%, 73.3291%, and 59.4585%
better than those of the GRNN; 35.0392%, 31.8239%, and 14.8641%
better than those of the EMD-MMODA-CM; and 5.1982%, 15.8999%, Acknowledgement
and 7.3140% better than those of the SSA-MODA-CM, respectively.
Moreover, the mean improvement percentages in the SSE of the This research was supported by the National Natural Science
PCFS relative to the ARIMA, BP, DBN, ELM, ENN, LSTM, persistence Foundation of China (Grant No. 71573034).
14
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Appendix A
Table 1
Predicted values distribution fitting results of the PCFS
Dataset Forecasting Step Method EV Gamma Logistic Loglogistic Lognormal Rician Weibull Optimal Distribution
Site 1 1-Step MDA 0.9700 0.9414 0.9511 0.9306 0.9325 0.9628 0.9767 Weibull
MLE 0.9438 0.9075 0.8967 0.8661 0.8865 0.9366 0.9710 Weibull
2-Step MDA 0.9629 0.9409 0.9443 0.9224 0.9261 0.9587 0.9709 Weibull
MLE 0.9345 0.8999 0.8855 0.8541 0.8782 0.9291 0.9638 Weibull
3-Step MDA 0.9649 0.6097 0.9441 0.9210 0.9251 0.9569 0.9712 Weibull
MLE 0.9371 0.8984 0.8879 0.8550 0.8763 0.9290 0.9647 Weibull
Site 2 1-Step MDA 0.7943 0.8959 0.9102 0.9338 0.9162 0.8809 0.8273 Loglogistic
MLE 0.5304 0.8864 0.8680 0.9189 0.9076 0.8211 0.7088 Loglogistic
2-Step MDA 0.7917 0.8726 0.9031 0.9276 0.9078 0.8710 0.8200 Loglogistic
MLE 0.5007 0.8717 0.8555 0.9119 0.8961 0.7998 0.6861 Loglogistic
3-Step MDA 0.8083 0.7271 0.8965 0.9164 0.8907 0.8624 0.8237 Loglogistic
MLE 0.4469 0.8414 0.8406 0.9002 0.8697 0.7647 0.6307 Loglogistic
Table 2
List of terminologies
15
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
Table 2 (continued )
AWDi Cumulative breadth error of i-th testing sample Pi i-th PF value of wind speed data
Ai i-th actual value of wind speed data NR range of forecasting values
Ui interval upper limit of i-th testing sample Li interval lower limit of i-th testing sample
IMAE improvement percentage of MAE. IMAPE improvement percentage of MAPE.
IRMSE improvement percentage of RMSE. ISSE improvement percentage of SSE.
SMAE Std value of MAE of n times prediction. SMAPE Std value of MAPE of n time prediction.
SRMSE Std value of RMSE of n time prediction. SSSE Std value of SSE of n time prediction.
Table 3
Parameters Setting of Each Model in Experiment I.
References
Model Parameters Value [1] Global Wind Energy Council. Global wind statistics. 2019. p. 2019. www.gwec.
net/wpcontent/uploads/vip/GWEC_PRstats2018_EN_WEB.pdf.
SSA Window Length 50 [2] Jiang P, Liu Z. Variable weights combined model based on multi-objective
Primary Ingredient Disintegration Number 20 optimization for short-term wind speed forecasting. Appl Soft Comput J
BPNN Input layer nodes number 5 2019. https://doi.org/10.1016/j.asoc.2019.105587.
Output layer nodes number 1 [3] Sun C, Bie Z, Xie M, Jiang J. Assessing wind curtailment under different wind
Hidden layer nodes number 5 capacity considering the possibilistic uncertainty of wind resources. Elec Po-
Learning rate 0.1 wer Syst Res 2016. https://doi.org/10.1016/j.epsr.2015.10.028.
Iterations number 10000 [4] Luo GL, Li YL, Tang WJ, Wei X. Wind curtailment of China’s wind power
Training requirements precision 0.0001 operation: evolution, causes and solutions. Renew Sustain Energy Rev 2016.
DBN Iterations Number 10000 https://doi.org/10.1016/j.rser.2015.09.075.
Number of input nodes 5 [5] Liu Z, Jiang P, Zhang L, Niu X. A combined forecasting model for time series:
Number of hidden layers 2 application to short-term wind speed forecasting. Appl Energy 2020. https://
doi.org/10.1016/j.apenergy.2019.114137.
Nodes number of the first hidden layer 30
[6] Zhang W, Zhang L, Wang J, Niu X. Hybrid system based on a multi-objective
Nodes number of the second hidden layer 30
optimization and kernel approximation for multi-scale wind speed fore-
ELM Iterations Number 10000
casting. Appl Energy 2020. https://doi.org/10.1016/j.apenergy.2020.115561.
Input layer nodes number 5 [7] Niu X, Wang J. A combined model based on data preprocessing strategy and
Output layer nodes number 1 multi-objective optimization algorithm for short-term wind speed fore-
Hidden layer nodes number 5 casting. Appl Energy 2019. https://doi.org/10.1016/j.apenergy.2019.03.097.
ENN Iterations Number 10000 [8] Gsella A, De Meij A, Kerschbaumer A, Reimer E, Thunis P, Cuvelier C. Evalu-
Input layer nodes number 5 ation of MM5, WRF and TRAMPER meteorology over the complex terrain of
Output layer nodes number 1 the Po Valley, Italy. Atmos Environ 2014. https://doi.org/10.1016/
Hidden layer nodes number 5 j.atmosenv.2014.03.019.
Context layer nodes number 5 [9] Di Z, Ao J, Duan Q, Wang J, Gong W, Shen C, et al. Improving WRF model
The minimum of the input element 0 turbine-height wind-speed forecasting using a surrogate- based automatic
The maximum of the input element 1 optimization method. Atmos Res 2019. https://doi.org/10.1016/
j.atmosres.2019.04.011.
GRNN Iterations Number 10000
[10] Wang J, Niu T, Lu H, Yang W, Du P. A novel framework of reservoir computing
Input layer nodes number 5
for deterministic and probabilistic wind power forecasting. IEEE Trans Sustain
Pattern layer nodes number 5
Energy 2020. https://doi.org/10.1109/TSTE.2019.2890875.
Summation layer nodes number 1 [11] Ebrahim Banihabib M, Mousavi-Mirkalaei P. Extended linear and non-linear
Output layer nodes number 1 auto-regressive models for forecasting the urban water consumption of a
LSTM Iterations Number 2500 fast-growing city in an arid region. Sustain Cities Soc 2019. https://doi.org/
Number of input nodes 5 10.1016/j.scs.2019.101585.
Number of hidden layers 2 [12] Wei LY, Cheng CH, Wu HH. A hybrid ANFIS based on n-period moving average
Nodes number of the first hidden layer 30 model to forecast TAIEX stock. Appl Soft Comput J 2014. https://doi.org/
Nodes number of the first hidden layer 30 10.1016/j.asoc.2014.01.022.
[13] Aasim, Singh SN, Mohapatra A. Repeated wavelet transform based ARIMA
Note: The primary parameter values in the related algorithms, including SSA, BPNN, model for very short-term wind speed forecasting. Renew Energy 2019.
DBN, ELM, ENN, GRNN, and LSTM, are presented in this table. Moreover, the pa- https://doi.org/10.1016/j.renene.2019.01.031.
rameters p and q of ARIMA are determined via AIC and BIC roles. [14] Wang J, Du P, Hao Y, Ma X, Niu T, Yang W. An innovative hybrid model based
on outlier detection and correction algorithm and heuristic intelligent opti-
mization algorithm for daily air quality index forecasting. J Environ Manag
2020. https://doi.org/10.1016/j.jenvman.2019.109855.
Table 4 [15] Chen Y, Zhang S, Zhang W, Peng J, Cai Y. Multifactor spatio-temporal corre-
Parameters Setting of Each Algorithm in Experiment II. lation model based on a combination of convolutional neural network and
long short-term memory neural network for wind speed forecasting. Energy
Mothed Parameter Value Convers Manag 2019. https://doi.org/10.1016/j.enconman.2019.02.018.
MODA Iteration Number 500 [16] Eksioglu M, Fernandez JE, Twomey JM. Predicting peak pinch strength: arti-
Archive Size 200 ficial neural networks vs. regression. Int J Ind Ergon 1996. https://doi.org/
10.1016/0169-8141(95)00106-9.
Population Size 50
[17] Bagherzadeh SA, D’Orazio A, Karimipour A, Goodarzi M, Bach QV. A novel
MOGOA Iteration Number 500
sensitivity analysis model of EANN for F-MWCNTseFe 3 O 4/EG nanofluid
Archive Size 200
thermal conductivity: outputs predicted analytically instead of numerically to
Population Size 50 more accuracy and less costs. Phys A Stat Mech Its Appl 2019. https://doi.org/
MMODA Iteration Number 500 10.1016/j.physa.2019.01.048.
Archive Size 200 [18] Ghasemi A, Hassani M, Goodarzi M, Afrand M, Manafi S. Appraising influence
Population Size 50 of COOH-MWCNTs on thermal conductivity of antifreeze using curve fitting
EMD No parameters and neural network. Phys A Stat Mech Its Appl 2019. https://doi.org/10.1016/
CEEMD STD of added noise 0.05 j.physa.2018.09.004.
Realization Number 50 [19] Moradikazerouni A, Hajizadeh A, Safaei MR, Afrand M, Yarmand H,
Maximum Sifting Iteration 500 Zulkifli NWBM. Assessment of thermal conductivity enhancement of nano-
SSA Window Length 50 antifreeze containing single-walled carbon nanotubes: optimal artificial
Primary Ingredient Disintegration Number 20 neural network and curve-fitting. Phys A Stat Mech Its Appl 2019. https://
doi.org/10.1016/j.physa.2019.01.051.
Note: Parameter settings of the relevant benchmark models in Experiment II are [20] Alrashed AAAA, Gharibdousti MS, Goodarzi M, de Oliveira LR, Safaei MR,
the same as in Table 3. Bandarra Filho EP. Effects on thermophysical properties of carbon based
16
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
nanofluids: experimental data, modelling using regression, ANFIS and ANN. [42] Ye L, Zhao Y, Zeng C, Zhang C. Short-term wind power prediction based on
Int J Heat Mass Tran 2018. https://doi.org/10.1016/ spatial model. Renew Energy 2017. https://doi.org/10.1016/
j.ijheatmasstransfer.2018.04.142. j.renene.2016.09.069.
[21] Li H, Wang J, Lu H, Guo Z. Research and application of a combined model [43] Zhu Q, Chen J, Zhu L, Duan X, Liu Y. Wind speed prediction with spatio-
based on variable weight for short term wind speed forecasting. Renew En- temporal correlation: a deep learning approach. Energies 2018. https://
ergy 2018. https://doi.org/10.1016/j.renene.2017.09.089. doi.org/10.3390/en11040705.
[22] Karimipour A, Bagherzadeh SA, Taghipour A, Abdollahi A, Safaei MR. A novel [44] Zhao Y, Ye L, Pinson P, Tang Y, Lu P. Correlation-constrained and sparsity-
nonlinear regression model of SVR as a substitute for ANN to predict con- controlled vector autoregressive model for spatio-temporal wind power
ductivity of MWCNT-CuO/water hybrid nanofluid based on empirical data. forecasting. IEEE Trans Power Syst 2018. https://doi.org/10.1109/
Phys A Stat Mech Its Appl 2019. https://doi.org/10.1016/j.physa.2019.01.055. TPWRS.2018.2794450.
[23] Liu H, Mi X, Li Y, Duan Z, Xu Y. Smart wind speed deep learning based multi- [45] Yang W, Wang J, Lu H, Niu T, Du P. Hybrid wind energy forecasting and
step forecasting model using singular spectrum analysis, convolutional Gated analysis system based on divide and conquer scheme: a case study in China.
Recurrent Unit network and Support Vector Regression. Renew Energy 2019. J Clean Prod 2019. https://doi.org/10.1016/j.jclepro.2019.03.036.
https://doi.org/10.1016/j.renene.2019.05.039. [46] Hong DY, Ji TY, Li MS, Wu QH. Ultra-short-term forecast of wind speed and
[24] Jiang P, Li R, Li H. Multi-objective algorithm for the design of prediction in- wind power based on morphological high frequency filter and double simi-
tervals for wind power forecasting model. Appl Math Model 2019. https:// larity search algorithm. Int J Electr Power Energy Syst 2019. https://doi.org/
doi.org/10.1016/j.apm.2018.10.019. 10.1016/j.ijepes.2018.07.061.
[25] Li H, Wang J, Li R, Lu H. Novel analysiseforecast system based on multi- [47] Wei D, Wang J, Ni K, Tang G. Research and application of a novel hybrid model
objective optimization for air quality index. J Clean Prod 2019. https:// based on a deep neural network combined with fuzzy time series for energy
doi.org/10.1016/j.jclepro.2018.10.129. forecasting. Energies 2019. https://doi.org/10.3390/en12183588.
[26] Tian C, Hao Y. Point and interval forecasting for carbon price based on an [48] Yang H, Zhu Z, Li C, Li R. A novel combined forecasting system for air pol-
improved analysis-forecast system. Appl Math Model 2020. https://doi.org/ lutants concentration based on fuzzy theory and optimization of aggregation
10.1016/j.apm.2019.10.022. weight. Appl Soft Comput J 2020. https://doi.org/10.1016/j.asoc.2019.105972.
[27] Wang J, Niu X, Liu Z, Zhang L. Analysis of the influence of international [49] Yang W, Wang J, Niu T, Du P. A novel system for multi-step electricity price
benchmark oil price on China’s real exchange rate forecasting. Eng Appl Artif forecasting for electricity market management. Appl Soft Comput J 2020.
Intell 2020. https://doi.org/10.1016/j.engappai.2020.103783. https://doi.org/10.1016/j.asoc.2019.106029.
[28] Bahrami M, Akbari M, Bagherzadeh SA, Karimipour A, Afrand M, Goodarzi M. [50] Li C, Zhu Z, Yang H, Li R. An innovative hybrid system for wind speed fore-
Develop 24 dissimilar ANNs by suitable architectures & training algorithms casting based on fuzzy preprocessing scheme and multi-objective optimiza-
via sensitivity analysis to better statistical presentation: measure MSEs be- tion. Energy 2019. https://doi.org/10.1016/j.energy.2019.02.194.
tween targets & ANN for FeeCuO/EgeWater nanofluid. Phys A Stat Mech Its [51] Wang R, Wang J, Xu Y. A novel combined model based on hybrid optimization
Appl 2019. https://doi.org/10.1016/j.physa.2018.12.031. algorithm for electrical load forecasting. Appl Soft Comput J 2019. https://
[29] Peng Y, Parsian A, Khodadadi H, Akbari M, Ghani K, Goodarzi M, et al. Develop doi.org/10.1016/j.asoc.2019.105548.
optimal network topology of artificial neural network (AONN) to predict the [52] Wu C, Wang J, Chen X, Du P, Yang W. A novel hybrid system based on multi-
hybrid nanofluids thermal conductivity according to the empirical data of objective optimization for wind speed forecasting. Renew Energy 2020.
Al2O3 e Cu nanoparticles dispersed in ethylene glycol. Phys A Stat Mech Its https://doi.org/10.1016/j.renene.2019.04.157.
Appl 2020. https://doi.org/10.1016/j.physa.2019.124015. [53] Zhou Q, Wang C, Zhang G. Hybrid forecasting system based on an optimal
[30] Hemmat Esfe M, Saedodin S, Naderi A, Alirezaie A, Karimipour A, model selection strategy for different wind speed forecasting problems. Appl
Wongwises S, et al. Modeling of thermal conductivity of ZnO-EG using Energy 2019. https://doi.org/10.1016/j.apenergy.2019.05.016.
experimental data and ANN methods. Int Commun Heat Mass Tran 2015. [54] Bates JM, Granger CEJ. Combination of forecasts. Oper Res Q 1969. https://
https://doi.org/10.1016/j.icheatmasstransfer.2015.01.001. doi.org/10.2307/3008764.
[31] Karimipour A, Bagherzadeh SA, Goodarzi M, Alnaqi AA, Bahiraei M, Safaei MR, [55] He Z, Chen Y, Shang Z, Li C, Li L, Xu M. A novel wind speed forecasting model
et al. Synthesized CuFe2O4/SiO2 nanocomposites added to water/EG: evalu- based on moving window and multi-objective particle swarm optimization
ation of the thermophysical properties beside sensitivity analysis & EANN. Int algorithm. Appl Math Model 2019. https://doi.org/10.1016/j.apm.2019.07.001.
J Heat Mass Tran 2018. https://doi.org/10.1016/ [56] Zhang Y, Wang J, Lu H. Research and application of a novel combined model
j.ijheatmasstransfer.2018.08.112. based on multiobjective optimization for multistep-ahead electric load fore-
[32] Alrashed AAAA, Karimipour A, Bagherzadeh SA, Safaei MR, Afrand M. Electro- casting. Energies 2019. https://doi.org/10.3390/en12101931.
and thermophysical properties of water-based nanofluids containing copper [57] Bo H, Niu X, Wang J. Wind speed forecasting system based on the variational
ferrite nanoparticles coated with silica: experimental data, modeling through mode decomposition strategy and immune selection multi-objective drag-
enhanced ANN and curve fitting. Int J Heat Mass Tran 2018. https://doi.org/ onfly optimization algorithm. IEEE Access 2019. https://doi.org/10.1109/
10.1016/j.ijheatmasstransfer.2018.07.123. ACCESS.2019.2957062.
[33] Safaei MR, Hajizadeh A, Afrand M, Qi C, Yarmand H, Zulkifli NWBM. Evaluating [58] Hassani H, Ghodsi Z. A glance at the applications of Singular Spectrum
the effect of temperature and concentration on the thermal conductivity of Analysis in gene expression data. Biomol Detect Quantif 2015. https://doi.org/
ZnO-TiO2/EG hybrid nanofluid using artificial neural network and curve 10.1016/j.bdq.2015.04.001.
fitting on experimental data. Phys A Stat Mech Its Appl 2019. https://doi.org/ [59] Bo H, Nie Y, Wang J. Electric load forecasting use a novelty hybrid model on
10.1016/j.physa.2018.12.010. the basic of data preprocessing technique and multi-objective optimization
[34] Wu H, Bagherzadeh SA, D’Orazio A, Habibollahi N, Karimipour A, Goodarzi M, algorithm. IEEE Access 2020. https://doi.org/10.1109/access.2020.2966641.
et al. Present a new multi objective optimization statistical Pareto frontier [60] Unnikrishnan P, Jothiprakash V. Daily rainfall forecasting for one year in a
method composed of artificial neural network and multi objective genetic single run using Singular Spectrum Analysis. J Hydrol 2018. https://doi.org/
algorithm to improve the pipe flow hydrodynamic and thermal properties 10.1016/j.jhydrol.2018.04.032.
such as pressure drop and heat transfer coefficient for non-Newtonian binary [61] de Carvalho M, Rua A. Real-time nowcasting the US output gap: singular
fluids. Phys A Stat Mech Its Appl 2019. https://doi.org/10.1016/ spectrum analysis at work. Int J Forecast 2017. https://doi.org/10.1016/
j.physa.2019.122409. j.ijforecast.2015.09.004.
[35] Hemmat Esfe M, Wongwises S, Naderi A, Asadi A, Safaei MR, Rostamian H, [62] Mirjalili S. Dragonfly algorithm: a new meta-heuristic optimization technique
et al. Thermal conductivity of Cu/TiO2-water/EG hybrid nanofluid: experi- for solving single-objective, discrete, and multi-objective problems. Neural
mental data and modeling using artificial neural network and correlation. Int Comput Appl 2016. https://doi.org/10.1007/s00521-015-1920-1.
Commun Heat Mass Tran 2015. https://doi.org/10.1016/ [63] Song J, Li S. Elite opposition learning and exponential function steps-based
j.icheatmasstransfer.2015.05.014. dragonfly algorithm for global optimization. IEEE Int. Conf. Inf. Autom. ICIA
[36] Wang J, Qin S, Zhou Q, Jiang H. Medium-term wind speeds forecasting uti- 2017 2017;2017. https://doi.org/10.1109/ICInfA.2017.8079080.
lizing hybrid models for three different sites in Xinjiang, China. Renew Energy [64] Zhou Y, Wang R, Luo Q. Elite opposition-based flower pollination algorithm.
2015. https://doi.org/10.1016/j.renene.2014.11.011. Neurocomputing 2016. https://doi.org/10.1016/j.neucom.2015.01.110.
[37]
Salcedo-Sanz S, Angel MPB, Ortiz-García EG, Portilla-Figueras A, Prieto L, [65] Sun Y, Wang X, Chen Y, Liu Z. A modified whale optimization algorithm for
Paredes D. Hybridizing the fifth generation mesoscale model with artificial large-scale global optimization problems. Expert Syst Appl 2018. https://
neural networks for short-term wind speed prediction. Renew Energy 2009. doi.org/10.1016/j.eswa.2018.08.027.
https://doi.org/10.1016/j.renene.2008.10.017. [66] Guo ZH, Wu J, Lu HY, Wang JZ. A case study on a hybrid wind speed fore-
[38] Cassola F, Burlando M. Wind speed and wind energy forecast through Kalman casting method using BP neural network. Knowl Base Syst 2011. https://
filtering of Numerical Weather Prediction model output. Appl Energy 2012. doi.org/10.1016/j.knosys.2011.04.019.
https://doi.org/10.1016/j.apenergy.2012.03.054. [67] Box GEP, Jenkins GM, Reinsel GC. Time series analysis: forecasting and control.
[39] Zhang C, Wei H, Zhao X, Liu T, Zhang K. A Gaussian process regression based fourth ed. 2013. https://doi.org/10.1002/9781118619193.
hybrid approach for short-term wind speed prediction. Energy Convers [68] Mudelsee M. Climate time series analysis: classical statistical and bootstrap
Manag 2016. https://doi.org/10.1016/j.enconman.2016.08.086. methods. https://doi.org/10.1007/978-90-481-9482-7; 2010.
[40] Liu H, Erdem E, Shi J. Comprehensive evaluation of ARMA-GARCH(-M) ap- [69] Hinton GE. Learning multiple layers of representation. Trends Cognit Sci 2007.
proaches for modeling the mean and volatility of wind speed. Appl Energy https://doi.org/10.1016/j.tics.2007.09.004.
2011. https://doi.org/10.1016/j.apenergy.2010.09.028. [70] Huang G Bin, Zhou H, Ding X, Zhang R. Extreme learning machine for
[41] Zhang J, Wei Y, Tan Z. An adaptive hybrid model for short term wind speed regression and multiclass classification. IEEE Trans Syst Man Cybern B Cybern
forecasting. Energy 2020. https://doi.org/10.1016/j.energy.2019.06.132. 2012. https://doi.org/10.1109/TSMCB.2011.2168604.
17
P. Jiang, Z. Liu, X. Niu et al. Energy 217 (2021) 119361
[71] Shen W, Fu X, Wang R, Yin Y, Zhang Y, Singh U, et al. A prediction model of improved regularized extreme learning machine. Energy 2018. https://
NH3 concentration for swine house in cold region based on Empirical Mode doi.org/10.1016/j.energy.2018.09.180.
Decomposition and Elman neural network. Inf Process Agric 2019. https:// [77]
Zuluaga CD, Alvarez MA, Giraldo E. Short-term wind speed prediction based
doi.org/10.1016/j.inpa.2018.12.001. on robust Kalman filtering: an experimental comparison. Appl Energy 2015.
[72] Ni YQ, Li M. Wind pressure data reconstruction using neural network tech- https://doi.org/10.1016/j.apenergy.2015.07.043.
niques: a comparison between BPNN and GRNN. Meas J Int Meas Confed 2016. [78] Zhang X, Wang J, Gao Y. A hybrid short-term electricity price forecasting
https://doi.org/10.1016/j.measurement.2016.04.049. framework: cuckoo search-based feature selection with singular spectrum
[73] Hochreiter S, Schmidhuber J. Long short-term memory. Neural Comput 1997. analysis and SVM. Energy Econ 2019. https://doi.org/10.1016/
https://doi.org/10.1162/neco.1997.9.8.1735. j.eneco.2019.05.026.
[74] Hao Y, Tian C, Wu C. Modelling of carbon price in two real carbon trading [79] Alhnaity B, Abbod M. A new hybrid financial time series prediction model. Eng
markets. J Clean Prod 2020. https://doi.org/10.1016/j.jclepro.2019.118556. Appl Artif Intell 2020. https://doi.org/10.1016/j.engappai.2020.103873.
[75] Wang Y, Wang J, Li Z. A novel hybrid air quality early-warning system based [80] Wang Y, Wang L, Yang F, Di W, Chang Q. Advantages of direct input-to-output
on phase-space reconstruction and multi-objective optimization: a case study connections in neural networks: the Elman network for stock index fore-
in China. J Clean Prod 2020. https://doi.org/10.1016/j.jclepro.2020.121027. casting. Inf Sci 2020. https://doi.org/10.1016/j.ins.2020.09.031.
[76] Sun N, Zhou J, Chen L, Jia B, Tayyab M, Peng T. An adaptive dynamic short- [81] Guan H, Zhao A, Shi G. Research on E-commerce precision poverty alleviation.
term wind speed forecasting model using secondary decomposition and an Economic Science Press; 2019 (In Chinese).
18