Home Search Collections Journals About Contact us My IOPscience

Ultra-Short-Term Wind Power Prediction Using a Hybrid Model

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2017 IOP Conf. Ser.: Earth Environ. Sci. 63 012005

([Link]

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: [Link]
This content was downloaded on 12/05/2017 at 03:30

Please note that terms and conditions apply.

You may also be interested in:

Autoregressive modelling for rolling element bearing fault diagnosis

H Al-Bugharbee and I Trendafilova

Time-Varying Autoregressive Modeling for Analyzing Transient Frequencies

Satoshi Kawata, Toshitaka Shigeoka and Osamu Nakamura

Using word-pictorial presentation model to simplify understanding concept test of Newton’s law
Jeffry Handhika, C Cari, A Suparmi et al.

Wind power makes its mark in Europe

Philip Surman

Wind power forecast error smoothing within a wind farm

Nadja Saleck and Lueder von Bremen

Causal relations in molecular dynamics from the multi-variate autoregressive model

A. Gorecki, J. Trylska and B. Lesyng

Enhanced regional forecasting considering single wind farm distribution for upscaling
Lueder von Bremen, Nadja Saleck and Detlev Heinemann

Hybrid modelling framework by using mathematics-based and information-based methods

J Ghaboussi, J Kim and A Elnashai

Effect of silanized-chitosan on flammability, mechanical, water absorption and biodegradability

properties of pseudo-stem banana fiber and montmorillonite filled waste polypropylene biocomposite
W E Prasetyo, A Prihandoko, S Pujiasih et al.
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

Ultra-Short-Term Wind Power Prediction Using a Hybrid

Model

E. Mohammed, [Link] and J. Yu

School of Electrical Engineering and Automation, Harbin Institute of Technology,
Harbin 150001, China

moheissa183@[Link]

Abstract. This paper aims to develop and apply a hybrid model of two data analytical methods,
multiple linear regressions and least square (MLR&LS), for ultra-short-term wind power
prediction (WPP), for example taking, Northeast China electricity demand. The data was
obtained from the historical records of wind power from an offshore region, and from a wind
farm of the wind power plant in the areas. The WPP achieved in two stages: first, the ratios of
wind power were forecasted using the proposed hybrid method, and then the transformation of
these ratios of wind power to obtain forecasted values. The hybrid model combines the
persistence methods, MLR and LS. The proposed method included two prediction types, multi-
point prediction and single-point prediction. WPP is tested by applying different models such
as autoregressive moving average (ARMA), autoregressive integrated moving average
(ARIMA) and artificial neural network (ANN). By comparing results of the above models, the
validity of the proposed hybrid model is confirmed in terms of error and correlation coefficient.
Comparison of results confirmed that the proposed method works effectively. Additional,
forecasting errors were also computed and compared, to improve understanding of how to
depict highly variable WPP and the correlations between actual and predicted wind power.

1. Introduction
Wind power prediction (WPP) is one of the most important issues of wind-power grid integration. An
accurate WPP is suggested to be practical to operators in power systems. From the applied time-series
data analysis, ultra-short-term WPP (USTWPP) is very important and contributes to the stability to the
power system. The accurate prediction is not only useful for ensuring operation security for the power
system, but also for enhancing the economic feasibility and performance of the power system [1].
In this paper, we developed and applied a modern hybrid method to predict high resolution by
applied time-series analysis in data for wind power system, which focus on understanding the current
methods to forecast on different time series (different time window and different length of prediction
sample). The developed method is multiple linear regressions, and least square (MLR&LS) based a
hybrid model which utilizes a new principle. As a proposal is a concern, there are two main ways to
predict wind power. The first way is data transformation: convert raw data of wind power into some
ratio and use this ratio as the value to be predicted. The second way is data inverse-transformation:
convert the predicted ratio value into actual value and then predict total wind power. In many studied
cases, there are sets of alternative forecast strategies. Forecast with a single model or method (SF) may
not always lead to fitting values, while the forecast with hybrid model or method (HF) is more

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

appropriate because the hybrid strategy can take the advantages of each SF and is more likely to get
better results.
In recent years, wind power as one of the most important renewable-energy sources is growing fast.
Utilization of wind energy showed rapid growth in the latest years which span many regions around
the world [2, 3]. In China, the growth of wind power is revealed to be faster than accounting for a
growth rate of 21.2% in 2012 [4]. Most of the wind-power projects are achieved based on national
goals of reducing an emission, and producing high-output electricity.
It has been shown that, wind power output has been strong fluctuating by characteristics and only is
partly controllable in nature [5, 6]. Short-term or ultra-short-term WPP is required for controlling and
scheduling of power generation, for self-dispatching of power transmission market [7] and for rating
improvement of the overhead line [8].
High penetrated wind power to the power grid brings many challenges to power system operators
[9] if not managed properly. Accurate load forecasting is more important for electric utilities planning
for the future [10]. More research on wind power technology is then necessary to supplement the
conventional energy sources. However, wind power is a highly fluctuating resource. The fluctuations
at large wind farms have the impact to the control and management strategies of their power output in
the power grid [5].
Several studies have investigated the accuracy of short-term WPP through three approaches:
physical numerical weather prediction models, statistical models based historical data and statistical
models with numerical weather prediction data as additional exogenous inputs [4]. Estimating wind
power output through accurate forecasting could significantly reduce the uncertainty of system
operators by using incorporated meteorological data. Short-term WPP and USTWPP are required for
monitoring and scheduling of power grid [7].
Power forecasting has become one of the major areas of research in engineering, particularly for
short-term and ultra-short-term forecasting that increasingly become important since the rise of the
competitive power system in markets. Although power forecasting is the important, great challenge is
expected due to influence of several exogenous variables of the power [11]. The main purposes of
ultra-short-term and short-term forecasting that ranges from minutes to hours are for real time control,
operation and security evaluation of the power system [12].

2. Related work and model

2.1. Time series models in forecasting

Time series is the sequence of data points, typically measured at sequential times, predominantly with
uniform time intervals [13]. There are many different modern and traditional approaches to time series
modeling. In this paper, the time-series analysis is performed using the autoregressive moving
average(ARMA) model, autoregressive integrated moving average(ARIMA) model and MLR&LS
model.
There are two types of goals for time-series analysis. The first goal is to identify the nature of the
phenomenon represented by the sequence of observations, and the second one is to forecast by
predicting future values of the time-series variable. Both goals require identifying and more or less
formally describing the pattern of observed time-series data. Time-series data of wind power is
developed on understanding the behavior of wind power signals at different time scales. As it is well
known, output power of wind farm is high variable and only partial controllable [14]. In the time-
series analysis, it is assumed that, the data consist of a systematic pattern (usually a set of identifiable
components) and noise (error) which usually makes the pattern difficult to be identified. Most time
series analytical techniques involve some form of filtering out of noise to make the pattern more
salient [11, 14-16].

2.2. ARMA model

2
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

ARMA model is considered as a basic sequential method and a practical model for comprehensive
time series [17]. ARMA model is a kind of time-series analysis statistical model, historically
suggested by the American G. Box and British G. Jenkins of statisticians [18]. For the random and
dynamic wind power data, ARMA model with time series reflects a larger advantage. The problem for
this model is to determine the order [16].
ARMA model uses autocorrelation or persistence in a time-series. The general form of ARMA
model is as following:
p q
wt  c   i wt i   j  t  j (1)
i 1 j 1
Where, t is the time to be forecasted; wt represents the forecasting value at time t; wt i represents the
historical sample data at time t  i ; i is the autoregressive coefficient; j is the moving average
coefficient; c is the constant value;  t j represents error at time t  j , is a random variable; p is the order
of the autoregressive terms; q is the order of the moving-average process.
Forecast the future values; they predict future values can be using the realized ARMA model, in
equation (2) is applied to predict the minutes ahead forecasting values (m = 1; 2; 3 minutes).
p q
w(t  m )   i wt i   j  t  j (2)
i 1 j 1

Where, w(t  m) is the predict wind power generation at time (t  m) .

2.3. ARIMA model

ARIMA model is one of the most popular and frequently used stochastic time series models in
forecasting. ARMA model is used only for stationary time series, but ARIMA model could be used in
cases of non-stationary time series. Applying generalized mathematical formulation with parameters
such as , ARIMA model can make a non-stationary time series become stationary. ARIMA
model is based on the assumption of linearity of the predicted values [19].
The general form of ARIMA model is as following:
p q
(1    i Li )(1  L) d w(t  m)  (1    j L j ) t (3)
i 1 j 1

Where, L represents the lag operator, and d is the order of the difference.

2.4. ANN Model

Artificial neural network (ANN) model is considered an advanced technology of short-term
forecasting, so it becomes an alternative technique to time series forecasting, and gained immense
popularity in last few periods. ANN has a flexible structure to model a wide variety of nonlinear
problems. The advantage of ANN model is that it has the potential to approximate a large class of
functions with a high degree of accuracy [19]. The ANN with multi-layer structure is very adaptable
because it can incorporate several levels in a single model [20]. The goal of use ANN is to find the
neural network weights that reduce the model errors.
In this paper, ANN structure used is a three-layer feed-forward back propagation neural network
which includes input layer, hidden layer and output layer. Actual input data in the following
simulation used two variables, recorded wind power and time for the forecasted horizon. Recorded
data obtains from the historical records of offshore region wind farms. The composition of the sample
set is as follows: training, validation and testing samples account for 70%, 15% and 15% respectively.
In general, ANN forecasting model can be written as follows:
W  F[ H1 (w), H 2 (w),..., H q (w)]   (4)
Where, W represents the dependent output variables (forecasted wind power values); w represents a
set of input variables (historical wind power data); F represents output layer activation functions, H i

3
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

represents hidden layer activation functions;  is the error term; forecasted power will be calculated by
equation (5) with time series values:
q p
W (t  m)   j .g (C0 j   ij .wt 1 )   t (5)
j 1 i 1
Where,  ij (i  0,1,2,..., p, j  1,2,..., q)  j ( j  1,2,..., q) are connected weights, p is the number of input
neurons and q is the number of hidden neurons.

2.5. Multiple Linear Regressions

MLR analysis in WPP technique used the weighted least square estimation method, based on
statistical analysis, can calculate the relationship between other factors [21]. The least square method
is a wide-range used parameter estimating method in the MLR. The basis of this method is that it can
lead to the least mean square error between the actual value and the estimated value [17].

3. Methods for USTWPP

In this section, we will describe the detail procedure of USTWPP with maximum width of the
prediction time window no more than [Link] output power of a wind farm or multiple wind farms
in a power grid will be determined by historical wind power data through hybrid time series model.

3.1. Procedure of USTWPP

Wind power has a random and intermittent nature characteristic, and the output power of wind
represents this characteristic. Hence it is very vital to reduce these influences for improving the
accuracy of WPP. This issue will be discussed from the following aspects:
- For different width of the prediction time window, what is the most appropriate length of
prediction?
In this paper, the width of the prediction time window will be assigned to 20, 30 and 60min; the
length or step of prediction will be assigned to 5 and 10min.
- What's the performance difference between SF and HF?
In this paper, SF is implemented by ARMA, ARIMA and ANN respectively; HF is implemented
by the MLR&LS model.
- What's the performance difference between single-point prediction and multi-point prediction?
In this paper, single-point prediction is to forecast the power or power ration of the next one length,
and multi-point prediction is to forecast the power or power ration of the next few lengths within the
width of the prediction time window.
- For single wind farm, what's the performance difference between direct prediction for single wind
power and indirect prediction for power ration to total wind power?
Power ration to total wind power is defined as equation (6):
Pi
wi  (6)
Ptotal
Where, Pi is the wind power of the i th wind farm, and Ptotal is the sum of wind power of multiple
wind farms including the i th wind farm.
After predicting power ratios wi and total power Ptotal , the WPP of the i th wind farm is as following:

Pi pred  wi  Ptotal (7)

- For multiple wind farms, what's the performance difference between direct prediction for total
wind power and indirect prediction for single wind power?
For indirect prediction, first predict the wind power of each wind farm, and then sum them to
obtain the total wind power of multiple wind farms. The wind power of each wind farm can be from
the direct WPP.

4
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

Figure 1 shows the detail procedure of USTWPP designed in this paper.

In figure 1, the procedure of USTWPP collects data from SCADA systems of wind farms, and
power grid dispatches center, and predicts WPP of single wind farm, total WPP of multiple wind
farms, and power rations of different wind farms at different prediction conditions and with different
prediction ways. The predicted value of the box②can be obtained from the box③and④according to
the formula(14), and the predicted value of the box⑤can be obtained from the sum of the box①.
This paper will observe and compare the error of USTWPP respectively from the above aspects.

Wind Power Data

Length of Prediction Length of Prediction

5 min 10 min
Prediction
Conditions
Time Time Time Time Time Time
Window Window Window Window Window Window
60min 30min 20min 60min 30min 20min

Prediction
Single-point Prediction Multi-point Prediction Ways

WPP of Single Wind Power Ration Total WPP of Multiple Predicted

Farm Prediction Wind Farms Variables
③
Direct Indirect Direct Indirect Prediction
Prediction Prediction Prediction Prediction Types

① ② ④ ⑤

Figure 1. Detail procedure of USTWPP.

3.2. Forecasting frame-work by MLR & LS hybrid model

MLR is a new analysis tool, and has the statistical learning mechanism to explain the behavior of one
of the dependent variables as a function of many other predictor variables. MLR allows finding the
best linear prediction equation between dependent and other predictor variables. The basic idea of the
hybrid model in this paper is to combine MLR and LR, which retains advantages to each approach and
access to the largest combination advantages. The main goal of doing so is to improve the forecast
accuracy.
Next, we will review the hybrid model forecasting method. The general form of MLR model is as
following:
p
 0    j w j   i  Wi (8)
j 1

Where, Wi is the variable to be predicted; w j is the predictor variable; p is the number of predictor
variables;  is the regression error;  is the regression coefficient that reflects the weight of each
predictor variable denotes the regression error.
The regression coefficients are calculated by using the least square function according to the
following formulas:
n p

 (Wi   0    j wij ) 2  min (9)

i 1 j 1

The optimal solution of  is:

̂  (w w) 1w W (10)

5
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

The fitted values are Wˆ  w̂ , the residuals are ˆ  W  Wˆ .

Equation (8) can be represented by the following matrix to estimate the regression coefficients.
 0 
1 w11 w12 ... w1 p      1  W1  (11)
1 
 w21 w22 ... w2 p   1   2  W2 
 2       
... ... ... ... ...     ...   ... 
  ...    
 1 wn1 wn 2 ... wnp     n  Wn 
 p 
 
When MLR model is used to WPP, the first part to the left of equation (11) showed the historical
wind power or power ration data matrix with n(p+1)dimensions, the right of the equation (11) show
the predicted wind power or power ration (response vector) with n dimension. Equation (11) above
can be simplified to: w    W
From equation (6) and (7), one can find that following equation holds in the matrix of equation (11)
The summation of power ration values wi ,t , mast be equal one, as equation (12)
n

w i ,t 1 (12)
i 1

Since [W1, W2….Wn] is a predicted vector, the summation of Wimight not be equal to 1. Thus, a
correction is needed, as in equation (13)
Wi
Pi p red  n (13)
W
i 1
i

After predicting the total power Ptotal and power ration by correction equation Pi pred , the indirect
WPP of the i th wind farm is as following:
Pˆi  Pi pred  Ptotal (14)
Where, P̂i is the indirect WPP of the i th
wind farm which satisfies n
Pˆtotal   Pˆi
i 1

Figure 2 shows the process of combine method for direct and indirect of USTWPP.
Wind Power Data [w] Sum of
Prediction
Power ratio wi Power ratio wi
WPP of Single Wind Indirect
Direct WPP
Total WPP Ptotal Direct prediction Wi Farm WPP

Calculate correction ratio W

i
Total WPP of
Pˆ i
Power Ratio
Indirect prediction value Multiple Wind Farm

Figure 2. Process of combine method for direct and indirect of USTWPP.

4. Simulation and results

In this section, simulations are carried out for USTWPP using the time-series analysis. The test data is
divided into two sets: one is the comparing results, and the other is the confirmed results.

4.1. Comparisons between single-point prediction and multi-point prediction

The prediction accuracy is a function for the difference between the predicted value and actual values.
In this section, the comparison of a numerical result of the single-point and multi-point prediction
ways will be displayed on the table 1. In this table, data is taken from the actual wind power data, and
the width of prediction time window T is assigned to 60, 30 and 20 min; the length of prediction t is
assigned to 5 and 10 min. MAPE is the mean absolute percentage error, and MSE is the
mean square error.

6
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

Table 1. MAPE and MSE comparisons of two ways.

Error Single-Point Multi-Point Single-Point Multi-Point
T=60min, t=5min T=60min, t=10min
MAPE 2.01 2.32 2.18 2.45
MSE 2.09 2.78 2.61 3.30
T=30min, t=5min T=30min, t=10min
MAPE 1.41 1.50 1.52 2.27
MSE 1.09 1.27 1.45 2.32
T=20min, t=5min T=20min, t=10min
MAPE 0.97 1.07 1.13 1.40
MSE 0.51 0.65 0.69 1.02

Table 1 show that the single-point prediction way will give the better results, and is slightly more
prone to MSE. This way could outperform others if fewer period data sets were used and fewer time
intervals ahead were predicted.
Figure 3 gives the total WPP comparison of the two prediction ways, at T = 60 min, t = 5min.

Figure 3. Total WPP with single-point & multi- point predictions.

From figure 3, the WPP curve with the single-point prediction way is closer to the actual wind
power curve, and correlation coefficient between the two curves is about 0.874; while the WPP curve
in the multi-point prediction way is less accurate with correlation coefficient being about 0.844.

4.2. Comparisons of multi-point USTWPP and its errors with different time
The graphical view about the multi-point USTWPP for three cases, anyone at different t with the
same T is shown in figure4 - 6. From these figures, it is seen that the total WPP curve accurately
follows the actual curve at the smaller time horizon, and at the end of T, only minor deviations are
seen fort= 10 min.

Figure 4. Multi-point USTWPP curve at Figure 5. Multi-point USTWPP curve at t = 5

t = 5 and 10min with T = 60min. and 10min with T = 30min.

Figure [Link]-point USTWPP curve at t= 5 and Figure 7. Multi-point USTWPP error curve at

7
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

10min with T = 20min. t= 5 and 10 min with T = 60 min.

To illustrate the performance of the multi-point USTWPP in a short run, the predicted result up to
60 min are shown in table 1, and the corresponding error curves are shown in figure7 - 9. From these
cases, the better accuracy for the USTWPP corresponds to the short width of prediction time window
T and sampling time period.

Figure 8. Multi-point USTWPP error curve at t Figure 9. Multi-point USTWPP error curve at t =
= 5 and 10 min with T = 30 min. 5 and 10 min with T = 20 min.
As shown in figure9, the multi-point USTWPP error is quite low and is the smallest one, which
concentrates within the prediction range.

4.3. Comparisons of two wind farms at WPP by direct and indirect prediction types
This part will show the comparisons between single WPP at two wind farm by direct and indirect
prediction types. According to the values of WPP in figure 10&11, direct and indirect WPP of two
wind farms respectively, WPP of two wind farms by direct type of prediction at t= 5 min and T = 60
min. The actual and prediction curves at two wind farm powers are shown in figure 10. It was
observed that the predicted values by the direct prediction are not close to the actual values.

Figure 10. Actual & predicted single WPP by Figure 11. Actual & predicted single WPP by
direct prediction type. indirect prediction type.
The WPP of two wind farms by indirect type of prediction at t= 5 min and T = 60 min, as it can
be seen from figure 11, the predictive power result at two wind farm is more satisfactory than the
direct predict type.
As part of the comparison between direct and indirect WPP, to find out, why the prediction by the
indirect type it is the high-accuracy value, in the case of indirect WPP, using the correction values to
prediction wind power, causing him to better than direct prediction.

4.4. Comparisons between proposed method & ARMA, ARIMA and ANN methods
This paper studies USTWPP using four methods: ARMA, ARIMA, ANN, and MLR&LS. In this part,
the simulation results for USTWPP using the statistical methods are presented.
Figure 12 gives the comparison procedures of the four methods.

8
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

Wind Power Data(Input)

MLR & LS Model ARMA ARIMA ANN

(Proposal Method) Model Model Mothod

Total WPP with Order (p,q) Order (p,d,q)

Training
Power Ration Selection Selection

Transform Ratio Total WPP Total WPP Validating

into Total WPP by ARMA by ARIMA and Testing

Comparisons of 4 Methods

Figure [Link] procedures among four methods.

The ARMA model is known to have good results, so it is used as the benchmark in this paper.
Figure 13 with using the ARMA model shows that the predicted values are almost close with the
actual values in the previous stage of the width of the prediction time window, and the error is very
small.

Figure 13. Actual & predicted total WP by Figure 14. Actual & predicted total WP by
ARMA. ARIMA.
Figure 14 with using the ARIMA model shows that the predicted curve is different from the actual
one. The predicted values by ARIMA model depend upon the parameter.
Simulation results of ANN model are demonstrated after many experiments with various network
architectures based on BP algorithm as shown in figure 15. The accurate predicted values and the
correlation between the actual values and the predicted values indicate high prediction accuracy.
Devaluation expected error also indicates high prediction accuracy. ANN has been successfully used
to forecast integrated wind power.

Figure 15. Actual & predicted total WP by ANN. Figure 16. Actual & predicted total WP by
MLR&LS.
The proposed MLR&LS hybrid model for USTWPP has been successfully implemented in this
paper. Figure 16 shows a time series of WPP that is closest to the time series of actual wind power. In
this study, the final predicted value by MLR&LS is more accurate than that by ARMA or other models.

4.5. Comparison of the predicted results

9
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

Comparison of ARMA, ARIMA, ANN and MLR&LS model is depicted in table 2. From table 2
and figure 13 - 16, it can be clearly observed the WPP accuracy level with ARIMA compared with
ARMA is not quite significant. It can be said both methods achieved a good forecast response as errors
of them are low. This result is consistent with most of previous studies. Moreover, we observed the
WPP accuracy level with ANN compared with ARIMA is not quite significant. However, it is
observed, the MLR & LS has a more high-accuracy level for total WPP, compared with other methods
in this paper. Numerical results in table 2 and figure 17 showed the MLR&LS is best than other
methods. We also noted the pattern of ARMA, and ARIMA shows the linear pattern. This finding is
consistent with a majority of the studies. Finally, all analysis and simulation results prove the proposed
approach gives better performance.

4.6. Forecasting accuracy comparisons of the four approaches

The forecasting accuracy is a function as the difference between the predicted values and the actual
values. Table 2 gives the error of predicted values compared to actual values for the four methods. It
can be inferred that the predicted values are close to the actual values when used MLR&LS. In these
calculations, it is also shown that there is high correlation between the actual curve and predicted
curve in all models, but the proposed hybrid model gives the better result.
Table 2. Error and correlation coefficient of the 4 methods for USTWPP at t= 5 min and T = 60 min.
Model MAPE MSE 
ARMA 2.01 2.09 0.86
ARIMA 3.09 5.63 0.65
ANN 1.89 1.67 0.90
MLR &LR 1.36 0.88 0.95
It is observed from table 2 above. That WPP by the proposed hybrid model outperformed all other
models. The hybrid model has both the lowest error and the maximum correlation coefficient.
Compared with ARMA, ARIMA and ANN, the hybrid model has the superiority of high USTWPP
accuracy. Through table 2 and figure 13-16, we evaluated the predictive performances of the four
models in details. In the total WPP of multiple wind farms, it is shown that the hybrid model is better
than ANN, ARMA and ARIMA.
Northeast
China

ing
del
Mo
eries
e S
Tim

AN
NE
rro
r
r
ro
Er
A

AR
M

IM
AR

AE
rro
r
MLR & LS based hybrid model

MLR & LS Error

Figure 17. Simulation chart of comparison between four methods to total WPP of multiple wind farms.
Based on the simulation training about the historical data of eight wind farms, the USTWPP is
reasonable. To avoid volatility due to the use of the MLR & LS hybrid model, all the forecasting were

10
2017 International Conference on Environmental and Energy Engineering (IC3E 2017) IOP Publishing
IOP Conf. Series: Earth and Environmental Science1234567890
63 (2017) 012005 doi:10.1088/1755-1315/63/1/012005

repeated with T = 60 min. From table 2 and figure 17, the hybrid model clearly performs much better
than ARMA, ARIMA and ANN model for single-point prediction and multi-point prediction. More
accurately, in calculation of total WPP with ARMA, ARIMA, ANN and hybrid model, MAPE and
MSE of the hybrid model are reduced, and  of the hybrid; models are increased.

5. Conclusion
For USTWPP, this paper compared the performances of four methods, including the MLR & LS
hybrid model. The WPP obtained from ANN, and ARMA are accurate compared against the ARIMA
model. Broadly, the hybrid model is the best one than ANN and ARMA. In the case of WPP, wind
power pursues the peak and valley of actual power components of the patterns. The shape of WPP
obtained from the hybrid model is more accurate than that obtained from the other methods for this
study. And the USTWPP by the proposed method is close to the actual power. So, we can easily
conclude that the forecasting accuracy with the combination method performs better than the
individual ones in general. Our simulation results show that the proposed method has best
performance than the same method for individual forecasting process and contributes for a reasonable
improvement in the final WPP. After several experiments with different width as the prediction time
window on the MLR & LS hybrid model, the smallest MSE was noted to give the best prediction
accuracy for the test and validity data. The MSE recorded during the experiment is presented in table 2
and figures17 to show that.

6. Acknowledgment
This work is supported by the National Natural Science Foundation of China under Grant 51377035.

7. References
[1] Gargeya M A and Pabba S P 2013 J. IJAREEIE. 2 1203-1212
[2] Wu Y K, LeeT C, Hsieh T Y and Lin W M 2014 J. JPEE. 2 11-17
[3] Soder L, Hofmann L, Orths A, Holttinen H, Wan Y H and Tuohy A [Link] Trans. Energy
conversion.22 4-12
[4] Chen N, Qian Z, Nabney I T and Meng X 2014. IEEE Trans. Power Syst. 29656-665
[5] Sorensen P, Cutululis N A, Vigueras-Rodríguez A, Jensen L E, Hjerrild J, Donovan M H
andMadsen H 2007. IEEE Trans. Power Syst. 22 958-965
[6] de Haan J E, Frunt J and Kling W L [Link] [Link] Europe. 1-8
[7] Taylor J W and McSharry P E 2007. IEEE Trans. Power Syst. 222213-2219
[8] Albizu I, Fernandez E, Mazon A J, Sagastabeitia K J, Bedialauneta M T and Olazarri J G
[Link] 2015 ICCEP. 382-388
[9] Chang W Y 2013 J. Energies. 64879-4896
[10] Ismail M M and Hassan M M2013. IEEE ICMIC. 112-118
[11] Nataraja C, Gorawar M B, Shilpa G N and Harsha J S 2012 J. IJESIT. 1 45-53
[12] Jain A and Satish B 2008. IEEERegion 10 Conference. 1-6
[13] Botterud A, Wang J, Miranda V and Bessa R J 2010 J. Electricity Journal. 23 71-82
[14] Coughlin K 2011 J. Lawrence Berkeley National Laboratory
[15] Chang WY 2014 J. JPEE. 2 161-168
[16] Li L, Wang M, Zhu F and Wang C 2009 J. ISCSCT. 293-297
[17] Hongfang C and Rui S 2010. IEEE First [Link] Symposium on CDEE. 22-26
[18] Zhang J and Wang C 2013. IEEE International Conference [Link].361-364
[19] AreekulPh, Senjyu T, Toyama Hand Yona A [Link] Trans. Power Syst. 25 524-530
[20] Defilippo S B, Neto G G and Hippert H S 2011 J. 44 JAIIO - SIO 2015 - ISSN: 2451-7569
[21] Alfares H K and Nazeeruddin M 2002 J. International Journal of Systems Science. 3323-34

11