You are on page 1of 8

A Methodology for Quantifying Variability of Renewable Energy Sources

by Reserve Requirement Calculation

Douglas A. Halamay Ted K.A. Brekken


Student Member Member

Oregon State University


Corvallis, OR
brekken@eecs.oregonstate.edu

Abstract -- Most of the major renewable power sources, such


as solar, wind, and ocean wave, are variable in time. Variability
arises from local turbulence, weather patterns, clouds, diurnal
variations, and seasonal variations. The variable, non-
dispatchable nature of renewable power sources requires an
increase in the utility reserve requirement to maintain
reliability. Power sources with greater minute-to-minute and
hourly variation require greater reserves. This paper presents a
methodology for quantifying the variability of renewable power
sources based on calculating the utility reserve requirement over
short and medium time scales.

Index Terms—Wind power generation, solar power


generation, Markov processes, power systems, power system
reliability, power system stability, marine technology,
simulation, reserve requirements. Fig. 1. Histogram of hourly change for wind vs. ocean wave power (adapted
from [2]).
I. INTRODUCTION
The existing utility power generation infrastructure is This paper presents a methodology for quantifying
designed to accommodate significant variability in the form variability by calculating the approximate reserve
of changes in load or the loss of generation (either planned or requirement necessary for the utility to absorb large-scale
unplanned) [1]. However, as variable renewable power power production from a given variable, renewable power
generation (e.g., wind, solar, and wave power) increases to source. The advantage of this methodology is that it yields a
large levels, the additional variability can strain grid directly useful and intuitive value that reflects the cost and
operation. To accommodate the added variability, the utility difficulty of absorbing the variability.
operators carry a certain amount of generation on standby to
cover either over-production or under-production of large II. METHODOLOGY
renewable energy plants. This flexible generation is called In order to balance generation with load on a minute-by-
the “reserve.” minute, hourly, or daily basis, the variability of both the
Different renewable power sources have different generation and the load must be examined. Variable
variability characteristics. For example, wave power renewable generation appears to the utility operators as a
generally has less hourly variation than wind power, as negative load. Therefore, all reserve calculations are done on
shown in Fig. 1. This variability can be quantified as in Fig. load-minus-renewable-generation (i.e., net load).
1 by a histogram of magnitude of change from one sample to Three different timescales are currently used by the utility
the next, or using other traditional methods such as industry to calculate reserve requirements. The first,
calculating statistical measures (e.g., mean, variance, regulation, is defined as the difference between the minute-
skewness, etc.). The disadvantage of these approaches to the to-minute load-minus-renewable-generation and the 10-
quantification of variability is that the results may not have minute average load-minus-renewable-generation. The
any direct physical meaning. second timescale, following, is defined as the difference
between the 10-minute average load-minus-renewable-

978-1-4244-5287-3/10/$26.00 ©2010 IEEE 666


American Electric Reliability Corporation (NERC) and
Western Electricity Coordinating Council (WECC) reliability
guidelines for balancing power supply and demand, while
eliminating those extremely rare events where required
reserves are beyond three standard deviations from the mean
[4].
The data used in this research is load power and wind
power for the BPA balancing area, solar power for 10
locations in the Pacific Northwest calculated from irradiance,
and ocean wave power for 3 locations in the Pacific
Northwest calculated from National Oceanic and
Atmospheric Administration buoys. The solar power is
modeled as 50% photovoltaic, and 50% concentrating
thermal with a 6 hour thermal time constant. All three data
sets were from corresponding 10-minute samples for the year
Fig. 2. Definition of following and imbalance reserve requirements. 2008.
Penetration is defined as the ratio of peak renewable
generation and the hourly average load-minus-renewable- generation over the year to the peak load over the year. To
generation. The final timescale, imbalance, is defined as the achieve a certain penetration level, the generation data is
difference between the hourly average load-minus- directly scaled.
renewable-generation and the forecasted load-minus- This paper builds off of [5] and [6] by directly comparing
renewable-generation for that hour. The imbalance the variability of wind, solar, and ocean wave against each
component of the reserve requirements is directly impacted other. Another significant and unique contribution of this
by the accuracy and frequency of the forecasted load-minus- paper is the presentation of the expected increase in reserve
renewable-generation [3]. requirement as a function of the level of penetration, as
Both the following and imbalance reserve requirements explained below.
are illustrated in Fig. 2. The available data for use in this
paper was at a 10-minute sample time, so the regulation III. SYNTHETIC DATA GENERATION
reserve component is not represented as it requires 1-minute In order to develop an analysis of the various renewable
sample time data. This should not affect the analysis as the energy resources under study, a considerable amount of data
regulation reserve requirement is consistently smaller than is required. While it would be ideal to have multiple years of
the following and imbalance reserve requirements due to the actual power data for all three of wind, solar, and ocean wave
generally small minute-to-minute resource variability [3]. generation, this is simply not feasible (especially in the ocean
In order to calculate imbalance reserve requirements, the wave case, given the current lack of installed generation).
scheduled or forecasted power must be determined for both Starting with the single year (2008) of real-world data
the renewable resource and the load. For this paper, a one- available for this study, it is thus desirable to formulate a
hour persistence method is used for the forecast. The one- method for synthesizing power generation data that
hour persistence method predicts that the renewable accurately represent the characteristics of the actual data.
generation for the next hour will be the same as it is when the One method that has been discussed in literature is the use of
forecast is made. The one-hour persistence method presents a Markov chain probability transition matrices [7], [8]. This
simple and adequate baseline for forecasting the power method was adapted for this study to better suit the variability
generated by a variable, renewable source. Its primary characteristics of the renewable resources.
limitation is a lag in the forecast for rapidly increasing or
decreasing power (i.e., ramps). A. Initial Markov Chain Modeling
Utilizing the one-hour persistence method, it is then A Markov chain is a stochastic process where the next
possible to calculate the following and imbalance reserve state of a system is dependent solely on the system’s current
requirements for each renewable resource. Industry-standard state. These types of random processes can thus be
practice for computing reserve requirements involves sorting characterized by a Markov state transition matrix Mij where
the following and imbalance calculated reserves and then the current state is represented by a row i, and the next state is
keeping the middle 99.5% of the data points, eliminating the represented by a column j. For a continuous-time Markov
top and bottom 0.25% of outliers. According to the process, the current state is at time t, and the next state is at
Bonneville Power Administration (BPA), this trimming of the time t+1. In the case of power generation data, the system
data still enables the controllers to meet the necessary North state corresponds to one of a discrete number of equally-sized

667
“bins.” Therefore, the probability of transitioning from one given the previous two states. To transform this into a
“bin” to another (i.e., from one power generation level to probability transition matrix, it is then necessary to sum the
another) is represented by a single entry pij in the state total number of occurrences for each row and divide each nijk
transition matrix. by this sum to determine the probability pijk of moving to
A first-order Markov chain is one in which the next state is state k given states i and j:
only dependent on the current state. However, the Markov
chain model can also be extended to higher orders, wherein = (4)

the next state is dependent on the two or more immediately
preceding states (e.g., a second-order Markov chain uses the Utilizing the state transition probability matrix, it is then
last two states to determine the transition probability for the possible to generate synthetic power generation data.
next state). In the first-order case, the state transition matrix Starting with two initial data points (given the second-order
Mij has the property Markov chain, two initial points are necessary), a random
number between 0 and 1 is first generated from a uniform
=1 (1) distribution. Through the discrete inversion method, this
uniformly distributed random number is then transformed
for any row i. Correspondingly, each of the pij satisfy the into a random number distributed according to the cumulative
property distribution function represented by the corresponding row of
0≤ ≤1 (2) the state transition probability matrix given the previous two
states. Since each row of the matrix forms a probability mass
such that each row i forms a probability mass function for the function, the cumulative sum of each row forms the
transitional probability of moving from state i at time t to corresponding cumulative distribution function. It is then
state j at time t+1. These properties can be easily extended to simple to determine the specific bin where the random
the higher-order Markov chain models as well. For example, number is greater than the previous bin but less than the next
in the second-order case, with state transition matrix Mijk and bin. This bin is therefore the state for the next data point.
individual transition probabilities pijk, with the new index k Given the continuous nature of the data and the discrete
signifying the next state, j representing the current state, and i nature of the bins, the final step is to utilize another uniformly
representing the previous state, (1) becomes distributed random number to determine exactly where within
the selected bin the next data point will fall. This process is
=1 (3) repeated moving forward through time until an entire year of
data points (52704 10-minute data points for 2008) is
for any row defined by i and j. constructed.
Previous research [8] has found that the use of second-
order Markov chain models to generate synthetic wind data B. Markov Chain Adaptations
produces results that more closely match the original data Given the unique nature of the renewable generation
than with the use of first-order models. Thus, for this study, a resources in this study, the use of a simple second-order
second-order Markov chain was used as a baseline starting Markov chain model produced results that often did not
point for data generation. One of the primary variables that resemble the initial data. All three of wind, solar, and ocean
can affect the characteristics of the generated data is the wave generation generally exhibit both monthly/seasonal and
number of bins that are used to form the state transition daily (if not hourly) dependence. For example, solar
probability matrix. Generally, using more bins provides a generation is generally much greater in the summer than in
more accurate outcome, but at the cost of additional the winter given the increased solar irradiance of the
computational complexity due to the size of the matrix [7]. summertime. Similarly, in many locations around the world,
For the wind and wave data in this study, 100 bins were used, wind generation exhibits day/night dependence (although in
while the solar data utilized 60 bins (the difference is the Pacific Northwest of the US, this dependence is far
explained further below). overshadowed by temperature and frontal effects). Thus,
The formation of the state transition probability matrix is utilizing a Markov chain model that ignores these
readily achieved by processing the real-world data (wind, dependencies produces synthetic data that fails to match
solar, and wave) on a point-by-point basis. To start, each reality.
data point’s bin is determined, and then a simple count is In order to address this disparity, this study proposed the
performed moving forward in time through the data. Given addition of two more new state indices to the previously-
the previous two data points, the corresponding bin for the described second-order Markov chain model: month of the
current data point is incremented by one, thus providing a year (12 possible states), and time (either the specific hour of
count nijk of the number of occurrences for each next state the day (24 possible states) or simply a binary day/night
state). For the solar generation data, the specific hour of the

668
day is very important to determining the generated power through the sets of previous and current bins to find a usable
given the high correlation between the two, while in the wind probability distribution. Logically, the best probability
and wave cases, a simple day/night binary state provides distribution to use would be the one that is closest in two-
enough differentiation while at the same time remaining dimensional Euclidean space to the original previous, current
computationally feasible. This difference in the number of pair. Thus, the Euclidean distance between each possible pair
states for the time index accounts for the differing number of of bins and the original (null state) bins was calculated as
bins used in the state transition probability matrix as such
described above. In order to allow for 24 states in the solar
time index, the number of bins had to be reduced from 100 to = − + − (5)
60 to make the calculations feasible, given that a five- and sorted according to the magnitude of dij. Next, a search
dimensional Markov state transition probability matrix was was performed to find the first pair of i and j that yielded a
now required. non-null probability mass function. This row of the state
The addition of these new state indices, however, transition probability matrix was then utilized to determine
introduced a new problem due to the continuous time nature the next data point for the power generation time-series.
of the data. In traditional Markov chain theory, it is possible Utilizing this method, the synthetic solar generation data
to have a state where it is possible to “transition into, but not had less than 4% of its data points produced from a null state
out of, itself” [9]. These types of states are known as case (the solar resource was the worst case given the large
trapping or absorbing states. In the case of this study, a number of hour states), while the wind and wave resources
similar problem arose when the state indices were made both had less than 1% of their data points from the null state
dependent on time. It was possible to start at time t in one set case. In all three cases, the average Euclidean distance that
of given states (month, hour, previous, and current bins) and yielded a viable probability mass function was less than 1.5
move to another set of given states at time t+1 where no (i.e., the pair of previous and current bins that yielded a
probability mass function existed due to the forward passage viable distribution was generally very close to the original
of time (and the corresponding month and hour state indices). null state pair).
By having to increment the states for the month and hour at Thus, to generate the synthetic data for this study, a
the same time as moving to a new power state, there was no Markov state transition probability matrix Mmhijk was used,
longer a guarantee that the combination of the new month, where m is the month, h is the hour (or binary day/night), i is
hour, previous, and current states would yield a row of the the previous state, j is the current state, and k is the next state.
transition matrix with a viable probability mass function. Utilizing the methods described above, 100 sets of synthetic
Thus, instead of an absorbing state it was possible to arrive in power generation time-series data were produced for each of
a null state, so-named because the values for the applicable wind, solar, and ocean wave generation.
row of the state transition probability matrix were null for this
case (i.e., no probability mass function existed because no C. Data Validation
occurrences were counted in the formation of the matrix and One of the key considerations in producing synthetic
thus by (4), all nijk were zero, resulting in null values for that power generation data is ensuring that the synthetic data
row). accurately reproduces the characteristics of the real-world
Several methods were formulated to address this problem, data from which it is sourced. While there are a variety of
with varying levels of success. An initial solution was to statistical measures (e.g., mean, variance, skewness, kurtosis)
simply perform a search through the state transition matrix to that are often used to compare sets of data, it is often the case
find the next closest hour with the same previous and current that these measures do not fully capture the unique nature of
states that yielded a valid probability distribution. This, the data in question (especially in the case of time-series data)
however, was not a complete solution because it did not deal [8]. Thus, it is important to consider several different
with those cases where the month transition was the cause of measures to determine the validity of the synthetic data.
the null state. A second solution, building off the first The three methods that were used to compare the actual
attempt, utilized a more complex two-state search to find the and synthetic data for this study were normalized
combination of month and hour that provided a valid autocorrelation, probability distribution (histogram), and
distribution. This method was deemed unsatisfactory because visual inspection. The normalized autocorrelation provides
it often produced anomalies in the data (for example, insight into the periodic nature of the resources and how well
producing solar generation in the middle of the night). the synthetically generated data compare in preserving that
The final solution to the null state problem thus required time dependence. Similarly, the probability distribution plots
that the month and hour remain unmodified so as to maintain provide insight into how well the overall power generation
the time-dependent nature of the data. Therefore, it was characteristics of the renewable resources are being captured
necessary to create an algorithm that performed a search in the synthetic data. Finally, visual inspection also plays an

669
Fig. 3. Normalized autocorrelation for actual vs. sample synthetic wind Fig. 6. Probability distribution for actual vs. sample synthetic wind power
power generation data. generation data.

Fig. 4. Normalized autocorrelation for actual vs. sample synthetic solar Fig. 7. Probability distribution for actual vs. sample synthetic solar power
power generation data. generation data.

Fig. 5. Normalized autocorrelation for actual vs. sample synthetic ocean Fig. 8. Probability distribution for actual vs. sample synthetic ocean wave
wave power generation data. power generation data.

important role in verifying that the synthetic data match a 7, and 8, respectively. It is important to note that the
variety of characteristics of the actual data that are difficult to synthetic data shown in these plots is a sample of one of the
capture using purely statistical methods (e.g., wind ramps). 100 sets of generated time-series data. While the
The plots for normalized autocorrelation for wind, solar, autocorrelation plots for solar and ocean wave are almost
and ocean wave generation are shown in Fig. 3, 4, and 5, perfect matches, the wind autocorrelation has some
respectively. In addition, the probability distribution plots for deviations between the synthetic and actual data, particularly
wind, solar, and ocean wave generation are shown in Fig. 6, for the small amount of diurnal periodicity. However,

670
Fig. 9. Percentage change in following reserve requirement from base load Fig. 12. Percentage change in imbalance reserve requirement from base load
for wind as a function of penetration percentage. for wind as a function of penetration percentage.

Fig. 10. Percentage change in following reserve requirement from base load Fig. 13. Percentage change in imbalance reserve requirement from base load
for solar as a function of penetration percentage. for solar as a function of penetration percentage.

Fig. 11. Percentage change in following reserve requirement from base load Fig. 14. Percentage change in imbalance reserve requirement from base load
for ocean wave as a function of penetration percentage. for ocean wave as a function of penetration percentage.

compared to previous results [8], the synthetic data generated reasonably well with the real-world data.
with the additional state indices for month and hour
demonstrates a marked improvement. Nevertheless, this is an IV. RESULTS
area that could benefit from additional research in the future. Utilizing each of the 100 sets of data scaled at 1 percent
For the probability distribution plots in Fig. 6, 7, and 8, the intervals between 1 and 15 percent penetration for each of
match between actual and synthetic data is clearly quite good. wind, solar, and ocean wave generation, the following and
Thus, it was determined that the synthetic data validates

671
Fig. 15. Comparison of change in following reserve requirement from base Fig. 16. Comparison of change in imbalance reserve requirement from base
load between wind, solar, and ocean wave generation as a function of load between wind, solar, and ocean wave generation as a function of
penetration percentage. penetration percentage.
imbalance reserve requirements were calculated. In order to requirements as penetration percentage increases than the
minimize the effect of any irregularities caused by the wind resource (and even compared to the solar resource in the
synthetic data, the average following and imbalance reserve imbalance case). This is an important result because it
requirements over the 100 sets were utilized. In order to demonstrates that the use of diversified renewable resources
provide clarity, the following and imbalance reserve will allow an increase in penetration percentage while
requirements were plotted separately for wind, solar, and lessening the impact of increasing reserve requirements [5].
ocean wave. Figs. 9, 10, and 11 demonstrate the following For clarity, Fig. 15 and Fig. 16 directly illustrate a
reserve requirements as a percentage increase over the comparison between wind, solar, and ocean wave generation
reserve requirement for base load as the penetration for following and imbalance reserves (without the error bars
percentage increases for wind, solar, and ocean wave, and at the same scale). In the imbalance case, as penetration
respectively. Similarly, Figs. 12, 13, and 14 demonstrate the approaches 15 percent, reserve requirements for wind are
imbalance reserve requirements as a percentage increase over increased by almost 15 percent over the base load, while
the base load reserve for increasing penetration percentage reserve requirements for solar are only increased by roughly
for wind, solar, and ocean wave, respectively. Given the use 7 percent and for ocean wave by less than 4.2 percent. Thus,
of averaging over the 100 synthetic data sets, the error bars the results demonstrate how the variability of each of these
illustrate a one standard deviation error on either side of the resources (and the impact of that variability on a power
average values for each penetration percentage on all of the system) can be quantified through reserve requirements.
plots.
The quadratic best-fit line was found and is presented for V. CONCLUSIONS
each plot, along with the value of the coefficient of The paper presents several key conclusions.
determination that demonstrates the quality of the fit.  A method for generating synthetic renewable
Clearly, the quadratic functions are a good fit to the data. It resource power data that closely replicates the
is important to note that for each of wind, solar, and ocean salient statistical properties of the real data is
wave, the scaling for the following and imbalance plots was presented and analyzed.
kept consistent to make it clear how the imbalance reserve  Ocean wave power and solar power generally have
requirement is generally larger in each case than the lower reserve requirements than wind power. The
following reserve requirement. Similarly, when comparing mathematical relationship between penetration and
the different renewable resources, it is important to take note reserve requirements for each renewable power
of the scale differences between the three sets of plots. source is also presented.
Comparing the wind and solar resources (Fig. 9-10, 12-
 Quantifying variability of renewable energy sources
13), it is clear that as penetration percentage increases, the
by reserve requirements is valid and meaningful,
variability of wind causes a much larger increase in reserve
and in some ways more immediately meaningful in
requirement (for both following and imbalance) than in the
understanding the real-world implications of
solar case. However, the comparison between wind and
renewable energy integration than traditional
ocean wave is even more impressive (Fig. 9 compared to Fig.
statistical measures of variability.
11 and Fig. 12 compared to Fig. 14). The ocean wave
resource requires a significantly smaller increase in reserve

672
ACKNOWLEDGMENT
This material is based upon work supported by the
National Science Foundation under Grant No. 0846533. The
authors would also like to thank Dr. Tuba Ozkan-Haller of
Oregon State University for her assistance with the wave
power generation code.

REFERENCES
[1] M. Milligan, K. Porter, E. DeMeo, P. Denholm, H. Holttinen, B. Kirby,
N. Miller, A. Mills, M. O’Malley, M. Schuerger, and L. Soder, “Wind
power myths debunked,” Power and Energy Magazine, IEEE, vol. 7,
November-December 2009.
[2] G. Sinden, “Renewable resource characteristics and network
integration,” in Renewable Electricity and the Grid: The Challenge of
Variability (G. Boyle, ed.), Earthscan, 2007.
[3] “2010 BPA Rate Case Wholesale Power Rate Final Proposal
Generation Inputs Study,” Bonneville Power Administration, Tech.
Rep., July 2009. [Online]. Available:
http://www.bpa.gov/corporate/ratecase/2008/2010_BPA_Rate_Case/do
cs/WP-10-FS-BPA-08_Web.pdf
[4] “Reserve Capacity Forecast For Wind Generation Within-Hour
Balancing Service,” Bonneville Power Administration, Tech. Rep.,
February 2008. [Online]. Available:
https://secure.bpa.gov/ratecase/openfile.aspx?fileName=WI-09-E-
BPA-02_Testimony.pdf&contentType=application%2fpdf
[5] D. Halamay, T.K.A. Brekken, A. Simmons, S. McArthur, “Reserve
Requirement Impacts of Large-Scale Integration of Wind, Solar, and
Ocean Wave Power Generation,” Power and Energy Society General
Meeting, Minneapolis, July 2010.
[6] S. McArthur, T.K.A. Brekken, “Ocean Wave Power Data Generation
for Grid Integration Studies,” Power and Energy Society General
Meeting, Minneapolis, July 2010.
[7] F. O. Hocaoglu, O. N. Gerek, and M. Kurban, “The effect of Markov
chain state size for synthetic wind speed generation,” in Probabilistic
Methods Applied to Power Systems, 2008. PMAPS ‘08. Proceedings of
the 10th International Conference on, 2008.
[8] A. Shamshad, M. A. Bawadi, W. M. A. W. Hussin, T. A. Majid, and S.
A. M. Sanusi, “First and second order Markov chain models for
synthetic generation of wind speed time series,” Energy, vol. 30, no. 5,
2005. [Online]. Available:
http://www.sciencedirect.com/science/article/B6V2S-4CTCT5C-
2/2/245b1b41e386f78fb969757f26e3b3a8
[9] H. Stark and J. W. Woods, Probability and random processes with
applications to signal processing, 3rd ed. Upper Saddle River, NJ:
Prentice Hall, 2002, pp. 366-372.

673

You might also like