Content uploaded by Anna Bruce
Author content
All content in this area was uploaded by Anna Bruce on Feb 01, 2018
Content may be subject to copyright.
Kanyawee Keeratimahat
Short term variability of utility-scale PV in the Australian National
Electricity Market
K. Keeratimahat1, A. G. Bruce1,2, and I. MacGill2,3
1School of Photovoltaic and Renewable Energy Engineering, UNSW, Sydney, Australia
2Centre for Energy Environmental Markets, UNSW, Sydney, Australia
3School of Electrical Engineering and Telecommunications, UNSW, Sydney, Australia
E-mail: k.keeratimahat@unsw.edu.au
Abstract
Increasing penetrations of utility-scale PV introduce more operational uncertainty to power
systems, as it firstly creates more short-term imbalance between demand and supply due to
the variable and only partially predictable behaviour of PV over short time frames, and
secondly displaces conventional generation which traditionally assists in managing system
imbalance and provides inertia to resist short-term frequency deviations. In the Australian
electricity market, frequency control ancillary services (FCAS) are used to purchase
regulation services to correct short-term uncertainty within the 5-minute dispatch period, and
contingency services to correct manage sudden unexpected changes in demand or supply.
However, it is unclear if and how these arrangements might need to change with more PV.
One risk is that penetrations of PV might be unnecessarily unlimited given the present
uncertainties of its impacts.
This paper seeks to address some of these questions by analysing the short-term operational
characteristics of utility-scale PV in the Australian National Electricity Market, at the
timescale of seconds. Several studies have been carried out to understand the short-term
characteristics of PV output, including the magnitude and frequency of output fluctuations
and ramp rates over different timescales in different regions. However, the limited availability
of high resolution data for utility-scale PV has hampered such work.
Our study analyses four second-resolution output from four utility-scale PV plants (20 MW –
100 MW) that are registered generators in the Australian National Electricity Market (NEM).
The statistical analysis examines the significance of PV output fluctuations at this 4-second
timescale. The results show that the variability over a four second time interval is
insignificant, but that extreme events, with ramp rates of the size of the entire utility PV plant,
potentially sufficient to trigger frequency regulation and contingency services, do occur albeit
infrequently. These events are much more likely to be caused by operation issues rather than
cloud transients, a challenge shared by conventional generation which is also subject to
sudden unexpected forced outages. Although the current level of PV penetration in the NEM
is low such that the impacts are barely visible, the variability of combined PV plants suggests
that there is more frequency occurrence of large magnitude change in aggregated PV power
output. This study represents the first stage of work assessing the scale of the variability
challenge imposed by utility PV, and hence appropriate management options.
1. Introduction
The increasing penetration of utility-scale photovoltaic (PV) generation in electricity
industries around the world is assisting in reducing the sector’s environmental harms but does
introduce a number of challenges for power system operation. Available PV generation from
such plant is inherently highly variable and only partially predictable over all relevant time
periods of operational decision making. This adds to the challenge of maintaining supply-
demand balance at all times and locations across the network. Furthermore, these plants
displace conventional generation which typically has greater dispatchability to assist in
managing variability in both demand and other generation. And of particular relevance to
short-term frequency management, the power electronics interface of PV does not offer the
inertia of conventional plant with synchronous generators, that inherently and immediately
acts to resist frequency deviations (AEMO, 2015).
Electricity market arrangements already require a range of ancillary services to assist in short-
term frequency management given the variability and only limited predictability of energy
demand (regulation), and major disturbances arising from the unexpected failure of large
conventional generation or network elements (contingencies). However, utility PV has novel
characteristics that are not yet fully understood and, hence, there are growing efforts to better
understand the variability and uncertainty of such plants, and explore revised market
arrangements to assist in managing this. One approach already in use for wind generation in a
number of electricity industries is to set a minimum synchronous generation requirement for
market dispatch. This inherently limits the instantaneous proportion of generation arising
from variable renewables. Such arrangements can, of course, have adverse impacts on the
dispatch and revenue of renewables generators and raises questions of how these limits should
be set, and whether other management approaches might be more appropriate.
These questions provide the motivation for the study presented in this paper. In particular, the
variability of utility PV is still only partially understood. This variability can also be expected
to vary according to the location (e.g. micro-climate) and chosen technologies (e.g. fixed plate
versus tracking) of the plants. Several studies (Curtright and Apt, 2008; Mills et al., 2011;
Sayeef et al., 2012; Shedd et al., 2012; Heslop and MacGill, 2014; Bucciarelli et al., 2015)
have been carried out to understand the short-term characteristics of PV output, including the
magnitude and frequency of output fluctuations and ramp rates over different timescales in
different regions.
There is considerable historical data available for large utility scale PV generation over
typical time frames of wholesale market operation (5-30 minutes). There are also some data
sets of high temporal resolution (seconds) PV generation for particular, generally small to
medium size, PV plants. However, there is currently only very limited data from large utility
size plants over the very short time frames relevant to frequency control ancillary services
(FCAS). For example, the Australian National Electricity Market (NEM) operates regulation
and fast contingency services at available SCADA (four seconds) time intervals, up to the five
minute period of spot market dispatch. The limited availability of high resolution data for
utility-scale PV has limited the usefulness of variability assessments.
The FCAS markets in the Australian NEM actually provide the basis for this work, as they
provide public four second generation output data for all scheduled generating units as part of
their calculation of how much each participant should pay for these services.
In order to better understand the variability of utility-scale PV, this paper analyses four
second-resolution output from the first four utility-scale PV plants (20MW – 100MW) in the
NEM. Variability here is specifically defined as the change in power output over time (Ela et
al., 2013; Riesz and Milligan, 2015).
As far as we are aware, this study presents the first analysis of this four second data for four
utility PV plants currently operating in the NEM. Statistical analysis is undertaken to examine
the significance of PV output fluctuations at the 4-second timescale.
A description of the datasets and methodology used for the analysis is presented in Section 2.
Section 3 and Section 4 describes the outcomes of the variability analysis and the
interpretation of the results including the investigation of extreme ramp events. Section 5
presents some preliminary conclusions of the study and directions for future work.
2. Methodology
The instantaneous power output at 4-second intervals of four utility-scale (20-102MW) PV
plants in Australia was extracted from the AEMO Ancillary Services Market Causer Pays
Data which is publicly available from the AEMO website (AEMO, 2016). A summary of the
utility PV plants, size, capacity factor, technology and time period of available data analysed
in this paper is presented in Table 1.
Table 1. Existing PV plants with AEMO 4-second ouput (causer pays) data available
Plant (latitude, longitude)
AC Capacity
(MW)
Capacity Factor
(as of 2016)
Data available
from
Mounting
Technology
Nyngan, NSW (-31.57, 147.08)
102
25%
May, 2015
Fixed-tilt
Moree, NSW (-29.57, 149.87)
56
26%
February, 2016
Single-axis tracking
Broken Hill, NSW (-31.99, 141.39)
53
22%
August, 2015
Fixed-tilt
Royalla, NSW* (-35.49, 149.14)
20
22%
April, 2016
Fixed-tilt
* Non-scheduled generator
2.1. Data cleaning
Statistical analysis was limited to daylight hours between sunrise and sunset. Therefore, the
number of data points on each day for each individual PV plant is different (different starting
timestamp and ending timestamp). In practice, there can be some complexities in analysis at
the start and end of days depending on the plant configuration, weather conditions and
inverter characteristics. For a comparative analysis, all the datasets must be time
synchronised. To achieve this, any missing data for a particular timestamp was filled by
assuming the value of the previous timestamp. Long periods of missing data were not
observed in these datasets. However, the dataset did include days where communication
errors were observed, such as a day that showed power output at one plant continue until
9PM, and these days were discarded from the analysis.
2.2. Variability
Variability here is defined as the change in PV plant power output over time. Variability of
PV generators was therefore calculated from the difference between the instantaneous power
at the next time frame and the current time frame as shown in Equation 1.
(1)
Where P(t) is the power output recorded at timestamp (t) and (t + 1) is the next timestamp.
The sampling frequency of power output, and hence shortest timeframe for variability
analysis, was 4 seconds, which is the SCADA time period for NEM operation.
2.3. Categorisation of sunny and cloudy days
In order to explore the variability observed under different general weather conditions, the
analysis established sunny and cloudy day categories. It has also sought to identify periods
where the plant output was likely being impacted by operational issues such as partial or
complete forced outages, or curtailment requirements. In previous studies, categorising sunny
and cloudy days has done by using solar resource data to create clear sky profiles and
comparison of the output from PV plants with the clear sky profile (Ibanez et al., 2012;
Gibescu, Nijhuis and Rawn, 2014). In the absence of solar resource data, categorisation for
this study is done by analysing the output fluctuation of the PV plants. Haghdadi et al. (2017)
proposed a method for classifying clear sky periods by discarding all periods which exceed
the aggregated variability threshold over a period (e.g. 1 minute) to achieve such
categorisation. The clear sky periods were retained and fitted with 3rd degree polynomial
equations to create a clear sky profile for each day. The study used several years of PV output
data in order to establish the profile, which could then be compared to the output data as a
basis for accurately determining cloudy periods in the data. However, while 3rd degree
polynomial equations can be fitted to the daily output of fixed PV arrays such as those studied
by Haghdadi, Figures 1-2, they cannot be applied to the generation profile from a single-axis
tracking system, Figure 3. Another approach is to categorise the data by aggregated ramp
rates over the whole day, as proposed by van Haaren et al. (2014). Although solar irradiance
is used by van Haaren et al. to calculate variability, this method does not rely on the sky
clearness index and therefore the plant power output can also be used. One drawback of this
method is that it cannot distinguish between clear sky and completely overcast days as the
aggregated variability of both days are similar yet the magnitude of power output would be
different. Therefore, sub criteria must be applied.
In this paper, we adopted the van Haaren method to classify the operational characteristics of
the PV plants, with the addition of sub criteria. Two factors were considered in the
classification, (1) the aggregated variability over the whole day, normalised to the plant
capacity and (2) the average power output between 10AM to 2PM, when the output of the
plants will be typically relatively stable with regard to changes in the sun’s position,
normalised to the plant capacity. The analysis also established criteria for detecting days with
operational issues such as plant failure or perhaps market dispatch instructions. The criteria
are presented in Table 2. The results are shown in Figure 1 to Figure 3 for the three larger PV
plants. It is interesting to note that the curtailed output at Nyngan is often around 50% of the
plant capacity (Figure 1). The curtailed output from Moree also shows the same pattern
(Figure 3) whereas the curtailment at Broken Hill appears to have three steps of 25% of the
plant capacity (Figure 2). These are likely to outcome of inverter, wiring or network
connection configuration of the plants. Significant clipping during the middle of sunny days is
apparent at the three larger plants, suggestive of a significant DC/AC rating ratio.
Table 2. Criteria to categorise types of day
Category -
Label
Description
Normalised daily
aggregated variability
Normalised average power
output between 10AM - 2PM
Category 1 -
ClearSky
Clear sky
<= 3
>= 0.7
Category 2** -
Curtailed
Output is clipped or
curtailed
<= 6
0.4 <= x <0.7
Category 3 -
OpIssue
Operation issue (operating
at very low output)
<= 3
<= 0.1
Category 4 –
PartlyCloudy
Partly cloudy
3 < x <= 6
>= 0.7
Category 5 –
Overcast
Cloudy/overcast
3 < x <= 6
< 0.4
Category 6 -
ScatteredCloud
Scattered moving clouds
resulting in high fluctuation
of solar irradiation.
> 6
No limit
** An additional criteria was added to Category 2 to distinguish between cloudy days when the output fluctuated
but still resulted in the same average power output as the day with curtailment. This is to consider the 90th
percentile of the power output throughout the day. If the 90th percentile of the day falls below the threshold, the
day is discard and classified as Category 5. The threshold was determined from 90th percentile value of day that
the output of the plant was curtailed. The day with curtailment was observed by visual inspection. Note that the
level of plant curtailment is different from plant to plant.
Figure 1. Nyngan power output profiles in 2016
Figure 2. Broken Hill power output profiles in 2016
Figure 3. Moree power output profiles in 2016
3. Variability analysis
3.1. Variability of individual PV plants
The results from the analysis of the variability over 4 seconds of the output at PV farms in
Australia in 2016 are shown in Table 3. Normalised change in power output per unit is a
common way to present output variability and allow comparison to be made across different
plant sizes. However, normalised ramp decreases as the plant capacity increases. As this study
focus on the magnitude of change in power, presenting the normalised ramp could be
misleading. Changes in power output over 4 seconds in this study were therefore presented as
absolute magnitudes (MW) to demonstrate the potential impacts of PV variability on the
system. Table 3 shows that the output variability from all individual PV plants remains within
1MW over 95% of the time. Intuitively, larger PV plants exhibit a higher frequency of large
fluctuations. All PV plants exhibited distributions with long tails – i.e. extreme events where
the changes in power output are close to the size of the plant capacity do occur. For this study,
this will be referred to as extreme ramp events. Events in the order of 50-100MW are of
sufficient scale to potentially cause challenges for system operation in the NEM, although it
should be noted that contingency reserves are generally set by the size of the largest
dispatched generator or transmission network element, which are typically in the hundreds of
MWs. The analysis also shows that the larger downward ramps occur more frequently that
upward ramps, suggestive of the role of plant failures in these events. This analysis suggests
that more regulation reserves may be required as utility-scale PV penetrations increase, in line
with other literature exploring this question (Riesz et al., 2011).
Table 3. Normalised frequency of occurrences of changes in PV power output at different magnitudes
Figure 4 shows the cumulative frequency of the absolute changes in output at each PV plant,
confirming that more than 85% of the time, the 4 seconds variability remains at zero. The
impact of spatial smoothing across a large plant due to the fact that cloud boundaries do not
affect the whole plant simultaneously can be observed from the graph. While 99.99% of the
variability at Royalla (20MW) is less than 75% of the plant capacity, 99.99% of the variability
at Nyngan (102MW) is within 30% of its plant capacity. This agrees with the results from
other studies (Murata et al., 2009; Marcos et al., 2012; ARENA, 2015) on geographical
smoothing of large-scale PV plants, which conclude that larger PV plants will not experience
a sudden change in cloud cover simultaneously. The trend towards oversizing the DC side of
PV plants will also act to reduce the effect of cloud transients, since the output will be capped
at inverter capacity (e.g. peak output clipping in Figure 1 to Figure 3).
Figure 4. Cumulative probability of the change in power output at each PV plant
Change in output (MW) (number of events in 2016)
-105 to -45
-45 to -10
-10 to -1
-1 to 1
1 to 10
10 to 45
45 to 105
Royalla 20MW
N/A
0.0036% (91)
0.51%
99.02%
0.47%
0.0034% (87)
N/A
Broken Hill 53MW
0.000080% (3)
0.00064% (24)
0.76%
98.59%
0.64%
0.00061% (23)
0.000080% (3)
Moree 56MW
0.00021% (7)
0.0059%(198)
1.17%
97.78%
1.05%
0.0049% (165)
0.00% (0)
Nyngan 102MW
0.00023% (9)
0.035% (1,339)
2.53%
96.34%
1.07%
0.031% (1,209)
0.00010% (4)
All plants
0.00045% (18)
0.042% (1,702)
3.76%
93.13%
3.03%
0.040% (1,623)
0.00012% (5)
3.2. Variability of combined PV plants
Figure 5 shows the tails of the frequency distributions of variability for all four PV plants.
When the outputs from all the PV plants are combined, there is a slightly higher frequency of
large magnitude changes. This can be seen by the frequency distribution in Table 3, which
shows a higher frequency of 4 second ramp events greater than 1MW than for individual PV
plants. The combined output (“All plants”) also results in one ramp event that is larger than
any of the ramps of the individual plants. The combined changes in output also show that a
few additional extreme ramp events have been created. Due to the limited current installed
capacity of utility scale PV, this impact is as yet quite minor, however the results show more
frequent occurrence of large magnitude aggregate changes in PV output with higher
penetration utility scale PV. This result agrees with the trend summarised in Fattori and
Anglani (2015). These results indicate that very short-term variability is likely to become
more of a concern when the level of PV penetration is higher, regardless of the impact of
spatial dispersion and uncorrelation of PV output over a long distance for short time periods.
Figure 5. Tails of frequency distribution of variability in all four PV plants
4. Investigation of the extreme ramp events
In this section, extreme ramp events of the size greater than 100 MW that were observed in
the variability analysis and which are particularly likely to be of interest to the system
operator (Summers, 2017) are examined in more detail. The number of extreme ramp events
are summarised in Table 4. The high number of events observed at Nyngan in 2015 is likely a
result of the commissioning phase and a similar effect can be observed with the PV plant at
Broken Hill. Observing the output of the plants on days with extreme ramp events show that it
is unlikely for cloud transients to be the cause of these events, and that they are more likely
the consequence of operational issues. Examples can be seen in Figure 6(a) and Figure 6(b),
which show that Nyngan plant output tripping over 4 seconds on a day that appeared to be
less cloudy. The investigation of extreme aggregate ramps created by combining all time
synchronised PV plant outputs shows one trip with a magnitude of 107 MW (Figure 6(c)).
This is equivalent to a small contingency event. If the difference between PV output and
dispatch target was greater than 100 MW, it would constitute a small contingency event
(Summers, 2017). Note that semi-schedule generators are not usually required to meet their
dispatch target which are set by Australian Solar Energy Forecasting System (ASEFS) but
may be constrained below the target. The dispatch target or forecasted value is calculated
based on the actual MW output of the previous dispatch interval, in normal operation (i.e.no
down-regulation order). If the MW output of the previous dispatch interval is down-regulated,
the dispatch target is calculated from the available solar resource. The impact of another
extreme ramp event with a size of 100 MW was investigated using frequency deviation (The
nominal frequency is 50Hz) and no distinct correlation was found, as shown in Figure 6(d).
This is likely to be because the dispatch target was already set at a low output, but may also
be the results of automatic governor control of load and other generators.
Table 4. Summary of 4-second change in power output with the size larger than 50 MW
PV plant
Year
Number of events
Nyngan
2015 (Available from May)
28
2016
6
Broken Hill
2015 (Available from August)
5
2016
2
Moree
2016 (Available from February)
6
Combined output
2016
14
(a) (b)
(c) (d)
Figure 6. (a) The drop of 100% of power output at Nyngan; (b) The drop of 100% of power output at
Moree; (c) the drop in combined power output from all PV plants of 107 MW; (d) the drop of 100%
power output at Nyngan with time synchronised system frequency.
5. Conclusion
This study has analysed the short-term variability of the existing PV plants registered in the
NEM. The analysis used 4-second data recorded by AEMO for FCAS causer pays
calculations. The results of the variability study show that over 95% of the time, these very
short-term output changes are within a 1 MW fluctuation and that the frequency of down-
ramp events is slightly higher than up-ramp events. Although the effect of PV variability is
limited by low PV penetration, the variability of the combined PV output shows an increasing
frequency of higher magnitudes of variability. This suggests that at higher penetration levels,
despite diversity smoothing effects, larger ramping events can be expected from the aggregate
PV generation. An investigation into extreme ramp events revealed that these were likely not
caused by cloud transients but rather operational issues.
Further studies should be conducted to deepen our understanding of this variability and its
interaction with both utility PV plant dispatch targets and overall power system frequency
management. This study has analysed historical data to assess the potential impacts of PV
variability on system operation. Modelling of PV at higher penetrations will be necessary for
a more comprehensive understanding of the impact of PV variability and uncertainty.
References
AEMO (2016) Ancillary Services Market Causer Pays Data. Available at:
https://www.aemo.com.au/Electricity/National-Electricity-Market-NEM/Data/Ancillary-
Services/Ancillary-Services-Market-Causer-Pays-Data.
ARENA (2015) Investigating the Impact of Solar Variability on Grid Stability.
Bucciarelli, M. et al. (2015) ‘PV Power Output Uncertainty in Australia’, Renewable Energy,
89(July 2013), pp. 9–13. doi: 10.1016/j.renene.2015.11.080.
Curtright, A. E. and Apt, J. (2008) ‘The character of power output from utility-scale
photovoltaic systems’, Progress in Photovoltaics: Research and Applications, 16(3), pp. 241–
247. doi: 10.1002/pip.786.
Ela, E. et al. (2013) ‘Impacts of Variability and Uncertainty in Solar Photovoltaic Generation
at Multiple Timescales’, National Renewable Energy Laboratory, (May). doi: NREL/TP-
5500-58274.
Fattori, F. and Anglani, N. (2015) ‘Review on the stochastic nature of photovoltaic generation
and its impact on the energy systems: Why it matters’, in 2015 IEEE Energy Conversion
Congress and Exposition (ECCE). IEEE, pp. 1122–1129. doi: 10.1109/ECCE.2015.7309816.
Gibescu, M., Nijhuis, M. and Rawn, B. (2014) ‘Prediction of power fluctuation classes for
photovoltaic installations and potential benefits of dynamic reserve allocation’, IET
Renewable Power Generation, 8(3), pp. 314–323. doi: 10.1049/iet-rpg.2013.0098.
van Haaren, R., Morjaria, M. and Fthenakis, V. (2014) ‘Empirical assessment of short-term
variability from utility-scale solar PV plants’, Progress in Photovoltaics: Research and
Applications, 22(5), pp. 548–559. doi: 10.1002/pip.2302.
Haghdadi, N. et al. (2017) ‘A method to estimate the location and orientation of distributed
photovoltaic systems from their generation output data’, Renewable Energy. Pergamon, 108,
pp. 390–400. doi: 10.1016/J.RENENE.2017.02.080.
Heslop, S. and MacGill, I. (2014) ‘Comparative analysis of the variability of fixed and
tracking photovoltaic systems’, Solar Energy. Elsevier Ltd, 107, pp. 351–364. doi:
10.1016/j.solener.2014.05.015.
Ibanez, E. et al. (2012) ‘A Solar Reserve Methodology for Renewable Energy Integration
Studies Based on Sub-Hourly Variability Analysis Preprint’, in Integration of Solar Power
into Power Systems Conference. Lisbon, Portugal. Available at:
https://www.nrel.gov/docs/fy12osti/56169.pdf (Accessed: 5 October 2017).
Marcos, J. et al. (2012) ‘Smoothing of PV power fluctuations by geographical dispersion’,
Progress in Photovoltaics: Research and Applications, 20(2), pp. 226–237. doi:
10.1002/pip.1127.
Mills, A. et al. (2011) ‘Dark Shadows’, IEEE Power and Energy Magazine, 9(3), pp. 33–41.
doi: 10.1109/MPE.2011.940575.
Murata, A., Yamaguchi, H. and Otani, K. (2009) ‘A method of estimating the output
fluctuation of many photovoltaic power generation systems dispersed in a wide area’,
Electrical Engineering in Japan (English translation of Denki Gakkai Ronbunshi), 166(4), pp.
9–19. doi: 10.1002/eej.20723.
Riesz, J. et al. (2011) Impact of the LRET on the costs of FCAS , NCAS and Transmission
augmentation. Available at: http://www.aemc.gov.au/Media/docs/ROAM Report-566727a5-
bfb8-4c7d-a88a-0b9469e5d19c-0.pdf.
Riesz, J. and Milligan, M. (2015) ‘Designing electricity markets for a high penetration of
variable renewables’, Wiley Interdisciplinary Reviews: Energy and Environment. Wiley
Periodicals, Inc., 4(3), pp. 279–289. doi: 10.1002/wene.137.
Sayeef, S. et al. (2012) ‘Characterisation and integration of high penetration solar power in
Australia: a solar intermittency study’, Publications.Csiro.Au, p. 60. Available at:
https://publications.csiro.au/rpr/download?pid=csiro:EP116553&dsid=DS2.
Shedd, S. et al. (2012) ‘A Statistical Characterization of Solar Photovoltaic Power Variability
at Small Timescales Preprint’, in The 2nd Annual International Workshop on Integration of
Solar Power into Power Systems Conference. Lisbon, Portugal.
Summers, K. (2017) Fast frequency service - treating the symptom not the cause. Available
at: http://www.escosa.sa.gov.au/ArticleDocuments/1047/20170208-Inquiry-
LicensingInverterConnectedGeneratorsIssuesPaperSubmission-
KSummers.pdf.aspx?Embed=Y.
