You are on page 1of 7

Journal of International Council on Electrical Engineering Vol. 2, No. 4, pp.351~357, 2012 351 http://dx.doi.org/10.5370/JICEE.2012.2.4.

351

The Analysis of Output from PV Power Station to Estimate Generation Reserve for Frequency Regulation
Motoki Akatsuka, Ryoichi Hara*, Hiroyuki Kita* Katsuyuki Takitani** and Koji Yamaguchi**
Abstract As a solution for the global environmental issues and energy resource depletion issue, photovoltaic generation (PV) has been installed rapidly. However, PV generation output heavily depends on climatic conditions and it would give negative impacts on balancing between supply and demand in power systems. In order to secure the power quality, the unstable output from the natural energy resource driven generators should be considered directly in usual power system operations such as supply and demand balancing. The solar radiation forecast would take a key role in the future operation of power system; therefore, investigation of performance and its characteristics of the forecast accuracy is important. In this paper, standard deviation and time series characteristics of solar radiation forecast error are statistically estimated. This paper also develops a First-order Markov process model for solar radiation forecast error. Keywords: PV power station, Solar radiation forecast, Statistical estimation, Time series analysis

1. Introduction
Recently, renewable energy driven generations have received great attentions from the viewpoint of CO2 emission mitigation and energy security enhancement. Photovoltaic generation (PV) is one of major renewable energy resources and its installed capacity has been rapidly growing due to financial supports by governments [1]. In Japan, the installation target of PV in 2030 is set to 40 times of which in 2005. This challenging target accelerates further installations in Japan. The penetration of PV brings some advantages; however, unstable and intermittent generation output may give some negative impacts on stable power system operation such as voltage variation, frequency variation and energy surplus. Therefore, assessment of those impacts is essential to maintain stable system operation in the near future. For the undesired assessment results, precautions must be considered. The authors have focused on the impact of PV installations on frequency regulation. As well known, frequency variation is caused by imbalance of total supply and total demand across the power system and is kept
Corresponding Author: Graduate School of Information Science and Technology, Hokkaido University, Japan (akatsuka@ee4-si.eng.hok udai.ac.jp) * Graduate School of Information Science and Technology, Hokkaido University, Japan ** Japan Weather Association, Japan (taki@sapporo.jwa.or.jp) Received: July 17, 2012; Accepted: September 15, 2012

within the admissible level by controlling generations output to follow total demand power momentarily. In detail, the balancing control is achieved in two stages; demand forecast based generation scheduling such as unit commitment (UC) and economic load dispatch (ELD) and real-time regulation so-called load frequency control (LFC) and governor free control (GF) for forecast error and shortterm fluctuation. When large amount of PVs are installed to the power system, solar radiation forecast should be also considered in UC and/or ELD process. Consequently, LFC and GF should compensate the solar radiation forecast error, too. That is, importance of forecast accuracy improvement is needless to say, forecast error estimation is also important to allocate adequate regulation reserve. This paper analyzes the actual solar radiation forecast error observed in the demonstration project named Verification of Grid Stabilization with Large-scale PV Power Generation promoted by NEDO in Wakkanai, Japan [2]. More specifically, static characteristic of the forecast error is analyzed and discussed in terms of standard deviation. Dynamic characteristic of forecast error is also modeled by means of the first-order Markov process in this paper.

2. Preprocess of Solar Radiation


Trend of solar radiation forecast error depends on season, time of day and weather condition. Therefore, the forecast error should be investigated and modeled with being

352

The Analysis of Output from PV Power Station to Estimate Generation Reserve for Frequency Regulation

1.4 Solar radiation [kW/m ] 1.2 1.0 0.8 0.6 If IET


2

forecast is almost flat from 8 to 16 oclock. This observation implies that the normalization process can eliminate both of seasonal and temporal trends in forecast error.
1.2 Solar radiation [kW/m2] 1.0 0.8 0.6 0.4 0.2 0 -0.2 4 8 12 Time [h] 16 20 If I eIf IET

0.4 0.2 0 Jan. 1st Date Dec. 31

Fig. 1. 30-minutes solar radiation and extraterrestrial solar radiation at around noon in FY2009. classified by these factors. Fig. 1 shows the solar radiation forecast (If) predicted at around noon in FY2009 and the corresponding extraterrestrial solar radiation (IET) [3]. Here, the forecast was made for 30 minutes average of solar radiation at the Wakkanai PV power station. IET can be calculated by the following equation.
I ET = I o cos i

Fig. 2. Actual and forecasted solar radiation, forecast error and extraterrestrial solar radiation.
0.8 KTf 0.6 Clearness index KT eKTf

(1)

0.4 0.2 0 -0.2 4 8 12 Time [h] 16

where, Io is the solar constant [kW/m2], i is the incidence angle of direct solar radiation which can be calculated theoretically from the date, time, latitude and longitude of measuring point. Fig. 1 implies the similarity of seasonal trends of If and IET; therefore, it is expected that normalization of solar radiation data based on IET can effectively eliminate the seasonal trend of If. For this reason, the observed solar radiation data were analyzed after normalization in this paper. The normalization was performed by the following equations [3];
KTf = I f I ET
eKT = eI I ET

20

Fig. 3. Actual and forecasted clearness index and normalized forecast error.

3. Static Characteristic
The mean value and standard deviation of clearness index forecast error (eKT), represented as KT and KT in this paper, were statistically estimated in the following procedure. First, past annual eKT data were classified by time of day (30 minutes interval was considered) and KTf since statistical characteristic of eKT depends on these factors. Then, KT and KT were evaluated for each classified group containing more than 20 samples (groups with less than 20 samples were ignored from the credibility perspective). Fig. 4 shows the estimated KT and KT. As shown in Fig. 4(a), absolute values of KT are smaller than 0.05 and negligible in almost all groups; therefore, KT is approximated by 0 in this paper. As shown in Fig. 4(b), KT from 8 to 16 oclock tends to small when KTf is small (cloudy sky) or large (clear sky). On the other hand, for the medium KTf (fine), KT becomes larger. For example, KT is 0.11 (KTf: 0.2), 0.19 (KTf: 0.4) and 0.07 (KTf: 0.8) at from 11:00 to 11:30. Since the largest KT is 0.20 in Fig. 4(b), the

(2) (3) (4) (5)

eI = I f I
KT = I / I ET

where, KTf and eKT are the normalized solar radiation forecast and forecast error, eI is the solar radiation forecast error [kW/m2] defined as equation (4), I is 30 minutes average of the actual solar radiation observed [kW/m2]. KT is the normalized solar radiation observed, and is so-called clearness index. As an example, If, I, eIf and IET in a certain day are shown in Fig. 2. Corresponding normalized values, KTf , KT and eKT are also shown in Fig. 3. As shown in Fig. 2, solar radiation forecast in the example day smoothly varies according to the diurnal motion. In Fig. 3, on the other hand, normalized

Motoki Akatsuka, Ryoichi Hara, Hiroyuki Kita, Katsuyuki Takitani and Koji Yamaguchi

353

possible largest I is estimated as 0.24[kW/m2] (multiplying 0.20 and maximum annual IET at Wakkanai (1.21[kW/m2])). This value provides an important criterion for discussion on regulation margin allocation, etc.
1.0 Clearness index forecast 0.8 0.6 0 0.4 0.2 0 4 8 12 Time [h] 16 20 -0.05 -0.10 0.10

0.05

(=1, 2, ... , N) represents the forecast target period (9:00 9:30, 9:30 - 10:00, , 14:30 - 15:00), N is the total number of forecast intervals (=12), d represents the date having same average KTf, D is the total number of days which have same average KTf. Estimated C() and R() are illustrated in Fig. 5. As shown in Fig. 5, C() and R() exponentially decrease with the increase in . Fig. 5(b) also reveals that the autocorrelation coefficients for different average KTf become similar and can be approximated by exp(-0.46). Since R(0.5) is close to 0.8, correlation coefficient between successive two eKT can be estimated at about 0.8 (value of R(0.5) shown in Fig. 5(b)), that is, we can say that forecast errors appeared in two successive time intervals are strongly associated.
0.04 0.03 0.02 0.01 0 -0.01 0 2 Lag [h] 4 6

Bias of forecast error

(a) Bias
Autocorrelation Function C()
1.0 Clearness index forecast 0.8 0.6 0.10 0.4 0.2 0 4 8 12 Time [h] 16 20 0.05 0 0.20 0.15

K =0.0-0.3
Tf Tf Tf Tf Tf Tf

K =0.3-0.4 K =0.4-0.5 K =0.5-0.6 K =0.6-0.7 K =0.7-1.0

(a) Autocorrelation function


Autocorrelation coeffieicnt R() 1.0 0.8 0.6 0.4 0.2 -0 -0.2 -0.4 0 2 Lag [h] 4 6
K =0.0-0.3
Tf Tf Tf Tf Tf Tf

(b) Standard deviation Fig. 4. Statistics of clearness index forecast error.

K =0.3-0.4 K =0.4-0.5 K =0.5-0.6 K =0.6-0.7 K =0.7-1.0

4. Autocorrelation Function
In this chapter, time sequential characteristic of solar radiation forecast error is discussed in terms of autocorrelation function. Fig. 4(b) indicates that the forecast error from 9 to 15 oclock show similar static characteristic, therefore, the authors have estimate the autocorrelation functions and autocorrelation coefficients of eKT from 9 to 15 oclock classified by the daily average of KTf. Here, the autocorrelation function C() and the autocorrelation coefficient R() are defined as follows.
C ( ) = C ( mt ) = E eKT ( k ) eKT ( k + m ) 1 D 1 N m d d = ( k + m) eKT ( k ) eKT D d =1 N m k =1

(b) Autocorrelation coefficient Fig. 5. Autocorrelation function and autocorrelation coefficient of clearness index forecast error.

5. Markov Process Model


5.1 Modeling As well known, an autocorrelation coefficient of the firstorder Markov process can be represented as an exponential function. Therefore, we can assume from Fig. 5 that the forecast error might be expressed as the first-order Markov process. Markov process representation of forecast error

(6)

R ( ) = C ( ) C ( 0 )

(7)

where, t is the length of forecast interval (=0.5[h]), k

354

The Analysis of Output from PV Power Station to Estimate Generation Reserve for Frequency Regulation

brings some useful applications such as a time-domain frequency excursion analysis. The first-order Markov process is expressed by equation (8) [4] whose parameters are related to the autocorrelation function and coefficient as shown in equations (9) and (10).
eKT ( k + 1) = eKT ( k ) + n ( k )

from the normal distribution. Autocorrelation coefficients of r(k) are also shown in Fig. 7. As shown in Fig. 7, autocorrelation coefficient for lags longer than zero is almost zero; this means that r(k) is independent from the past ones. Here, the irregularity observed at 5[h] lag and KTf >0.7 is a result of few samples.
0.6 0.5 relative frequency 0.4 0.3 0.2 0.1 0 -3 -2 -1 0 1 normalized residual 2 3 KTf=0.0-0.3 KTf=0.4-0.5 KTf=0.7-1.0 normal distribution

(8) (9) (10)

= 1 t
2 = (1 2 ) C ( 0 )

Here, n(k) is the white noise which follows the normal distribution N(0,). Designed is 0.77 (coefficient 0.46 in the approximation function for R() is substituted to ) and designed 2 are shown in Table 1. The above Markov process model of forecast error is expressed in terms of clearness index, however, we can easily convert to the solar radiation forecast error by equation (11).
eI (k ) = eKT ( k ) I ET (k )

Fig. 6. Relative frequency of residual from equation (9).


1.0 Autocorrelation coefficient KTf=0.0-0.3 KTf=0.4-0.5 0.5 KTf=0.7-1.0

(11)

Table 1. Variance of random number for each average of clearness index forecast
average 0.0 0.3 0.4 0.5 0.6 0.7 of KTf 0.3 0.4 0.5 0.6 0.7 1.0 2 0.008 0.013 0.014 0.012 0.007 0.003

-0.5 0

2 Lag [h]

5.2 Residual Analysis n(k) in equation (8) is assumed to be a normal random number. For the validity of this assumption, the residual of observed eKT has to distribute normally. Here, the residual r(k) is defined as equation (12).
r ( k ) = eKT ( k + 1) eKT ( k )

Fig. 7. Autocorrelation coefficient of residual from equation (9). 5.3 Evaluation of Generated Forecast Error Validity of the developed Markov process model is ascertained through computational comparisons in this section. In our investigation, time sequential solar radiation forecast errors were generated by the developed Markov process model with initial value set by the normal random number following N(0,2). The actual daily average of KTf was applied in the error generation process for convenience of comparison. Fig. 8 shows generated eI from 9 to 15 oclock with the actual observed eI. As shown in Fig. 8, instantaneous values of observed and generated error are different due to random factor in the Markov model. For stochastic discussions, we have evaluated and compared the relative frequencies of generated and observed eI along a year. Fig. 9 shows the obtained relative frequencies and cumulative relative frequencies. From the static viewpoint, relative frequencies at around maximum

(12)

In order to validate the assumption on n(k), distribution of the observed r(k) was investigated. In our investigation, r(k) observed from 9 to 15 oclock are classified by the daily average KTf, as we did for eKT in the previous chapter. After the classification, we count the relative frequencies of r(k). The obtained relative frequencies are normalized by the standard deviation of classified r(k) and plotted in Fig. 6. Relative frequency of the normal distribution N(0,1) is also shown in Fig. 6 for comparison. As shown in Fig. 6, relative frequency of r(k) is close to the normal distribution when the daily average KTf is from 0.4 to 0.5. However, for the daily average larger then 0.7, the relative frequency differs

Motoki Akatsuka, Ryoichi Hara, Hiroyuki Kita, Katsuyuki Takitani and Koji Yamaguchi

355

and minimum eI become one of main interests. From Fig. 9(b), we can find that 95% of forecast errors become larger than -0.21[kW/m2] and smaller than 0.21[kW/m2] in case of the generated errors. Likewise, 95% of observed forecast errors are larger than -0.23[kW/m2] and smaller than 0.23[kW/m2]. This result indicates that the developed Markov model can provide good estimations of possible forecast error magnitude.
Solar radiation forecast error [kW/m2] 0.3 0.2 0.1 0 -0.1 -0.2 -0.3 9 10 11 12 Time [h] 13 14 15 observed simulated

observed eI classified by the daily average KTf. Comparison of subfigures in Fig. 10 indicates that the developed Markov model results in relatively worse accuracy against large daily average KTf. The developed Markov process underestimates the possible magnitude estimation by about 0.09[kW/m2] against the 95% threshold.

6. Conclusion
The static and dynamic characteristic of solar radiation were analyzed in this paper. In the analysis, the solar radiation forecast error was normalized based on the extraterrestrial solar radiation to eliminate both seasonal and temporal trends. Our static analysis revealed that possible largest standard deviation of solar radiation forecast error is 0.24[kW/m2] which is equivalent to about 24% of rate capacity of PV. In the dynamic characteristics analysis, autocorrelation functions and autocorrelation coefficients of the solar radiation forecast error were calculated. Furthermore, this paper also developed the Markov process model for solar radiation forecast error. The future work is more detailed verification of the developed Markov process model, especially from the viewpoint of long-term accumulation of forecast error which provides an important insight for the energy required to compensate the forecast error.

Fig. 8. Example of observed and simulated solar radiation forecast error.


0.25 Relative frequency 0.20 0.15 0.10 0.05 0 -0.8 observed simulated

Acknowledgements
-0.6 -0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2] 0.8

(a) Relative frequency


Cumurative relative frequency 1.0 0.8 0.6 0.4 0.2 0 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2] 0.8 observed simulated

This work employed the data acquired in demonstration project Verification of Grid Stabilization with Largescale PV Power Generation by NEDO, Japan. This work was supported in part by Global COE Pr ogram Center for Next-Generation Information Technol ogy based on Knowledge Discovery and Knowledge Fe deration, MEXT, Japan.

References
[1] IEA-PVPS, Trends in Photovoltaic Applications Survey report of selected IEA countries between 1992 and 2008, Report IEA-PVPS T1-18, 2009 [2] R. Hara, H. Kita, T. Tanabe, H. Sugihara, A. Kuwayama, S. Miwa, Testing the technologies Demonstration Grid-Connected Photovoltaic Projects in Japan, IEEE Power and Energy Magazine, Vol.7,

(b) Cumulative relative frequency Fig. 9. Relative frequency distribution and cumulative relative frequency of observed and simulated solar radiation forecast error. Fig. 10 shows the relative frequencies of generated and

356

The Analysis of Output from PV Power Station to Estimate Generation Reserve for Frequency Regulation

0.4 observed simulated Relative frequency

0.3 observed simulated Relative frequency


-0.6 -0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2] 0.8

0.3 0.2 0.1 0 -0.8

0.2

0.1

0 -0.8

-0.6

-0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2]

0.8

(a) Average of clearness index forecast: 0.0 - 0.3


0.2 observed simulated

(b) Average of clearness index forecast: 0.3 - 0.4


0.2 observed simulated Relative frequency

Relative frequency

0.1

0.1

0 -0.8

-0.6

-0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2]

0.8

0 -0.8

-0.6

-0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2]

0.8

(c) Average of clearness index forecast: 0.4 - 0.5


0.4 observed simulated Relative frequency

(d) Average of clearness index forecast: 0.5 - 0.6


0.5 Relative frequency 0.4 0.3 0.2 0.1 0 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 2 Solar radiation forecast error [kW/m ] 0.8 observed simulated

0.3 0.2 0.1 0 -0.8

-0.6

-0.4 -0.2 0 0.2 0.4 0.6 Solar radiation forecast error [kW/m2]

0.8

(e) Average of clearness index forecast: 0.6 - 0.7

(f) Average of clearness index forecast: 0.7 - 1.0.

Fig. 10. Relative Frequency of Observed and Generated Solar Radiation Forecast Error for Each Daily Average Clearness Index Forecast. No.3, pp.77-85, 2009 [3] T. Muneer, Solar Radiation and Daylight Models, ELSEVIER, 2004 [4] G. E. P. Box, G. M. Jenkins, G. C. Reinsel, Time Series Analysis Forecasting and Control, WILEY, 2008 Ryoichi Hara received the Ph.D degree from Hokkaido University, Sapporo, Japan, in 2003. He has been an associate professor at Hokkaido University. His research interests are analysis, operation and control of electric power system. He is particularly interested in technological and economical harmonization of the bulk power system and distributed energy resources.

Motoki Akatsuka received the B.E. degree and M.E. degree from Hokkaido University, Hokkaido, Japan in 2007 and 2009, respectively. He is currently a Ph.D. candidate at Hokkaido University. He is interested in analysis and operation of PV system.

Hiroyuki Kita received the Ph.D degree from Hokkaido University, Sapporo, Japan, in 1994. He has been a professor at Hokkaido University. His research interests include the planning, analysis and control of electric power system.

Motoki Akatsuka, Ryoichi Hara, Hiroyuki Kita, Katsuyuki Takitani and Koji Yamaguchi

357

Katsuyuki Takitani received the B.E. degree from Hirosaki University, Aomori, Japan in 1981. He joined Japan Weather Association in April 1981. His research interest includes the meteorological information and meteorological forecast.

Koji Yamaguchi received the M.E. degree from Osaka Prefecture University, Osaka, Japan, in 1999. He joined Japan Weather Association in April 1999 and has engaged in development & research of the solar radiation forecast method.

You might also like