Impacts of Variability

and Uncertainty in Solar

Photovoltaic Generation
at Multiple Timescales
E. Ela, V. Diakov, E. Ibanez, and M. Heaney
National Renewable Energy Laboratory

NREL is a national laboratory of the U.S. Department of Energy, Office of Energy

Efficiency & Renewable Energy, operated by the Alliance for Sustainable Energy, LLC.

Technical Report
NREL/TP-5500-58274
May 2013

Contract No. DE-AC36-08GO28308

 Impacts of Variability
and Uncertainty in Solar
Photovoltaic Generation
at Multiple Timescales
E. Ela, V. Diakov, E. Ibanez, and M. Heaney
National Renewable Energy Laboratory
Prepared under Task No. SS13.3030

NREL is a national laboratory of the U.S. Department of Energy, Office of Energy

Efficiency & Renewable Energy, operated by the Alliance for Sustainable Energy, LLC.

National Renewable Energy Laboratory Technical Report

15013 Denver West Parkway NREL/TP-5500-58274
Golden, Colorado 80401 May 2013
303-275-3000 • www.nrel.gov
Contract No. DE-AC36-08GO28308
 NOTICE

This report was prepared as an account of work sponsored by an agency of the United States government.
Neither the United States government nor any agency thereof, nor any of their employees, makes any warranty,
express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of
any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately
owned rights. Reference herein to any specific commercial product, process, or service by trade name,
trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation,
or favoring by the United States government or any agency thereof. The views and opinions of authors
expressed herein do not necessarily state or reflect those of the United States government or any agency thereof.

Available electronically at http://www.osti.gov/bridge

Available for a processing fee to U.S. Department of Energy

and its contractors, in paper, from:

U.S. Department of Energy

Office of Scientific and Technical Information
P.O. Box 62
Oak Ridge, TN 37831-0062
phone: 865.576.8401
fax: 865.576.5728
email: mailto:reports@adonis.osti.gov

Available for sale to the public, in paper, from:

U.S. Department of Commerce

National Technical Information Service
5285 Port Royal Road
Springfield, VA 22161
phone: 800.553.6847
fax: 703.605.6900
email: orders@ntis.fedworld.gov
online ordering: http://www.ntis.gov/help/ordermethods.aspx

Cover Photos: (left to right) PIX 16416, PIX 17423, PIX 16560, PIX 17613, PIX 17436, PIX 17721

Printed on paper containing at least 50% wastepaper, including 10% post consumer waste.
List of Acronyms
AACEE absolute ACE in energy
ACE area control error
AGC automatic generation control
APS Arizona Public Service
CPS1 and CPS2 control performance standards (NERC)
CSP concentrating solar power
DASCUC day-ahead security-constrained unit commitment
FESTIV Flexible Scheduling Tool for Integrating VG
MAE mean absolute error
NERC North American Electric Reliability Corporation
PV photovoltaic
RPU reserve pickup
RTSCUC real-time security-constrained unit commitment
RTSCED real-time security-constrained economic dispatch
SCUC security-constrained unit commitment
SCED security-constrained economic dispatch
TEPPC Transmission Expansion Planning Policy
Committee (WECC)
VG variable generation
WECC Western Electricity Coordinating Council
WWSIS-2 Western Wind and Solar Integration Study Phase 2

iii
Executive Summary
This study investigates the effects of photovoltaic (PV) solar power variability and forecast
uncertainty on electric power grid operation in the Arizona Public Service system. Both
variability and uncertainty can create challenges for grid operators when balancing electricity
generation with demand while obeying power system constraints and maintaining the lowest
possible costs. Typically, studies of wind and solar variability are restricted to one time
resolution. In this study, we analyze variability and uncertainty across several timescales, similar
to the way power system operators manage load uncertainty. Our results aim to help grid
operators prepare for increases in PV generation share and improve system reliability when
integrating PV generation. We explore different mitigation strategies and methods to more
reliably and efficiently integrate solar power.

This study is a follow on to the Western Wind and Solar Integration Study Phase 2. The APS
balancing area within the Western Interconnection is studied in a high-load, high-PV, 5-day
period during summer, as projected to the year 2020, at a 1-minute time-sampling frequency. We
use the Flexible Scheduling Tool for Integrating Variable Generation (FESTIV). FESTIV
integrates security-constrained unit commitment, security-constrained economic dispatch, and
automatic generation control in one model. This allows multiple timescales to be studied and
accounts for intertemporal coupling between them. FESTIV attempts to realistically simulate
how most systems schedule generation in actual operation.

We find the following:

• The short-term forecast uncertainty increases the imbalance while having little impact on
operating costs.
• The variability of solar power in itself without real-time uncertainty (i.e., with perfect
hourly forecast) has little impact on imbalance and can actually improve certain
imbalance effects.
• Power system operator intervention can mitigate some of the variability as well as
uncertainty impacts while increasing operating costs.
• An improved day-ahead PV forecast reduces production costs with little effect on load-
generation imbalance.
• A faster (5-minute versus 1-hour intervals) dispatch frequency has the advantage of a
shorter forecast time horizon, so that imbalances can be corrected more often through
dispatch instead of through limited automatic generation control resources. This has the
effect of reducing both the imbalance and the system operating costs.
• Increasing regulation reserves to compensate for PV generation uncertainty improves
system balance while raising production costs.

iv
Table of Contents
Introduction ................................................................................................................................................. 1
Flexible Scheduling Tool for Integrating Variable Generation ............................................................... 5
System Description..................................................................................................................................... 9
The Impacts of Photovoltaic Variability and Uncertainty ..................................................................... 12
Imbalance Reduction Strategies ............................................................................................................. 19
Operator Action .................................................................................................................................... 19
Improved Day-Ahead Forecasts ........................................................................................................... 20
Increased Dispatch Frequency .............................................................................................................. 20
Regulation Reserve Increase ................................................................................................................ 22
Conclusion ................................................................................................................................................. 24
References ................................................................................................................................................. 27
Appendix A: Load and Interchange ........................................................................................................ 29
Appendix B: Day-Ahead Photovoltaic Forecast and Actual Photovoltaic Generation ...................... 30
Appendix C: Area Control Error Charts (Case 1–Case 7 in Order) ...................................................... 31

v
List of Figures
Figure 1. Depiction of variability and uncertainty ........................................................................................ 1
Figure 2. Operating reserve categorization [4] ............................................................................................. 3
Figure 3. Example of wind’s uncertainty and variability impacts on regulating reserve.............................. 4
Figure 4. Imbalance and regulating reserve needs resulting from variability and uncertainty ..................... 4
Figure 5. Flow diagram for FESTIV. Solid lines represent process flow, and dashed lines represent data
flow. ......................................................................................................................................... 5
Figure 6. Timeline for DASCUC, RTSCUC, RTSCED, and AGC in FESTIV ........................................... 7
Figure 7. Solar and Wind placements for the High Mix Scenario. (Yellow – PV, Orange – CSP, Green –
Wind). ...................................................................................................................................... 9
Figure 8. APS load and renewable generation as projected for a 5-day period from July 24 to July 28,
2020........................................................................................................................................ 10
Figure 9. ACE 1-day simulation. Perfect PV forecasts are used in the real-time simulation. .................... 13
Figure 10. ACE averaged over rolling 10-minute periods for 1-day simulation. Perfect forecasts are used
in the real-time simulation. .................................................................................................... 14
Figure 11. Generation for 1-day perfect forecast ........................................................................................ 15
Figure 12. Total generation and load for 5-day perfect forecast base case ................................................. 16
Figure 13. Regulation reserve requirements for Cases 1–6 (1% load) and Case 7 (WWSIS-2 [18]) ......... 22
Figure 14. CPS2 scores and production costs for all PV cases ................................................................... 25

List of Tables
Table 1. Typical Timing Parameters ............................................................................................................. 7
Table 2. APS Generation Capacity (2020 Projection) .................................................................................. 9
Table 3. Assumptions Made in Simulations ............................................................................................... 11
Table 4. Forecast Performance for 5-Day Period Using Persistence (Wind) and Persistent Cloudiness (PV)
for 60-Minute RTSCED 5-Day Period .................................................................................. 17
Table 5. Comparison of Variability and Uncertainty Impacts .................................................................... 18
Table 6. Utilization of Operator Action Using RPU ................................................................................... 19
Table 7. Day-Ahead Forecast Performance for 5-Day Period for DASCUC ............................................. 20
Table 8. Day-Ahead Uncertainty Impact Improvements ............................................................................ 20
Table 9. Forecast Performance for 5-Day Period Using Persistence (Wind) and Persistent Cloudiness (PV)
for 60-Minute RTSCED 5-Day Period .................................................................................. 21
Table 10. Impact of 5-Minute Dispatch ...................................................................................................... 21
Table 11. Impact of Increased Regulation Requirements ........................................................................... 23
Table 12. Case Comparisons....................................................................................................................... 24

vi
Introduction
Renewable power generation has seen a tremendous growth in recent years because it has
environmental benefits and zero fuel costs. Unlike many conventional generation sources,
however, many renewable resources, including wind power and photovoltaic (PV) solar power
are considered variable generation (VG). They have a maximum generation limit that changes
with time (variability) and this limit is not known with perfect accuracy (uncertainty). Variability
of VG occurs at multiple timescales, from seconds to minutes to hours, and requires movement
of other resources to ensure balance of generation and load. Uncertainty also occurs at multiple
timescales, from a few minutes ahead to hours ahead to days ahead. Resources must be available
when uncertainty is present and respond when it is resolved to ensure a balance of generation and
load. Figure 1 shows an example of the variability of a wind plant and the uncertainty of a solar
plant. The impacts of variability and uncertainty differ depending on the timescale, and strategies
to meet those impacts differ as well [1].

Variability Uncertainty
25000 2500 CSP
PV Actual
CSP
PV Forecasted
20000 2000

15000 1500
MW

MW

10000 1000
Wind Actual
5000 Forecasted Wind 500

0 0
0 6 12 18 24 30 36 42 48 0 6 12 18 24 30 36 42 48
Hours Hours

Figure 1. Depiction of variability and uncertainty

The characteristics of variability and uncertainty of PV solar power have been studied
extensively; see for example [2]. These characteristics can create challenges for system
operators, who must ensure a balance between generation and demand while obeying power
system constraints at the lowest possible cost. A number of studies have looked at the impact of
wind power plants, and some recent studies have also included solar PV [3]. The simulations that
are used in these studies, however, are typically fixed to one time resolution. This makes it
difficult to analyze the variability across several timescales. The studies use either a one- or two-
stage scheduling model when determining the commitment of generation resources and their
dispatch. This refers to using either a perfect forecast of VG or simulating one single chance of
forecast error, where the scheduling is updated only once, which typically reflects the impact of
long-term day-ahead forecast error. In reality, forecasts are updated continuously as the system
approaches real time and these forecasts will have different economic and reliability impacts at
different time horizons. It is difficult to show any reliability impacts using a single time
resolution and either a one- or two-stage scheduling model. As a result, system production costs
are typically the only primary metric used.

In this study we use a simulation tool that has the ability to evaluate both the economic and
reliability impacts of PV variability and uncertainty at multiple timescales. This information

1
should help system operators better prepare for increases of PV on their systems and develop
improved mitigation strategies to better integrate PV with enhanced reliability.

Another goal of this study is to understand how different mitigation strategies and methods can
improve the integration of solar power more reliably and efficiently. This may include improving
forecasts, improved awareness, scheduling improvements, or increased operating reserves.
Methods for determining operating reserve requirements, both in practice and in recent studies,
can vary significantly [4]. Figure 2 shows different operating reserve categories, some of which
exist today. Others may be needed for systems with higher amounts of VG. Although the
categorization can be done in many different ways, this shows operating reserve types separated
by how and why they are deployed.

Operating reserves can be characterized by their response speed (ramp rate and start time),
response duration, frequency of use (continuously or only during rare events), direction of use (up
or down), and type of control (i.e., control center activation, autonomous, or automatic, among
others). Some operating reserves are used to respond to routine variability of the generation or the
load. These variations occur on different timescales, from seconds to days, and different control
strategies can be required depending on the speed of the variability. Other operating reserves are
needed to respond to rare unexpected events such as the tripping of a generator.

Another way to classify the operating reserves could be based on whether they are deployed during
normal conditions or event conditions. Normal conditions can be based on both variability and
uncertainty, but they are continuously taking place. Events occur whether they can be predicted or
not. The standby costs and the deployment costs for each reserve category differ based on how
frequently they are used. This distinction, along with the technical requirements, makes certain
technologies more suitable for different operating reserve types than others.

Both the normal and event response categories can be further subdivided based on the required
response speed. Some events are essentially instantaneous and others take time. Different qualities
of these reserves are needed for different purposes. For example, instantaneous events need
autonomous response to arrest frequency excursions. The frequency then must be corrected back to
its scheduled setting. During both instantaneous and noninstantaneous events, the system’s
imbalance must be reduced to zero. There must also be some amount of reserves that can replace
the operating reserves after they are deployed to protect the system against a second event.
Nonevent reserves are typically meant to have zero net energy for a particular time period (over
some time, they will have equal positive deployments and negative deployments in energy). For
this reason, we do not consider the need for replacing these operating reserves in our discussion.

2
 Figure 2. Operating reserve categorization [4]
For this study, we focus mostly on regulating reserve, which is the reserve that automatic
generation control (AGC) uses to continuously correct the current imbalance of the system. The
regulating reserve is the last line of defense for correcting the imbalance during normal
conditions. Market prices throughout the United States illustrate that regulating reserve is
typically the most expensive reserve type. Regulating reserves are also the type used to support
the North American Electric Reliability Corporation (NERC) control performance standards
(CPS1 and CPS2) [5]. The impacts of the imbalance, and regulating reserve requirements that
attempt to correct those impacts, are affected by both variability and uncertainty. They will
depend greatly on how the system has been scheduled before they are needed in real time. For
example, following reserve, also called flexible reserve, can improve the situation in advance of
the need of regulation reserve, potentially reducing the need for regulation reserve [6]. Figure 3
shows how variability and uncertainty can both have impacts at different timescales, and how
these factors can influence the regulating reserve need. The illustration reflects a 5-minute
dispatch that takes 5 minutes to compute the optimization and transfer both the input and output
data. Load forecasts have been reasonably accurate in this time period because of the trending
nature of load, but VG output is not always as easily predicted. The imbalance, and the
regulating reserve need, is the difference in output with the forecasted schedule, shown in Figure
4. In this case, the variability along a second-to-second timescale is paired with the uncertainty at
the 10-minute-ahead timescale to compound the imbalance and regulating reserve need. The
other categories and their interactions affect how these reserves are used and quantified.

3
 100

90

80

70

60

Power (MW) 50

40 Actual Wind

30 Forecasted Wind
(schedule)
20 Average Wind

10

0
12:00:00 12:05:00 12:10:00 12:15:00 12:20:00

Figure 3. Example of wind’s uncertainty and variability impacts on regulating reserve

20

15

10

0
12:10:00 12:15:00 12:20:00

-5
Regulation resulting from variability

Regulation resulting from uncertainty

-10

Figure 4. Imbalance and regulating reserve needs resulting from variability and uncertainty
This study examines the Arizona Public Service system (APS) in a future year (2020) with a very
high solar PV penetration. Note that although we use a realistic system with data that are
representative of the APS system and the VG that could potentially be interconnected on the
system in 2020, our results are intended to determine the factors that affect system reliability
under uncertain and variable PV and wind generation. In addition, we have made a number of
assumptions that would affect the results (e.g., consideration of the transmission network).
Consequently, the comparison of results is more important than the absolute results themselves.
The study aims to quantify variability and uncertainty impacts of PV. We also examine how
various strategies might help improve system reliability and reduce the impacts that solar PV has
on system imbalances.

4
Flexible Scheduling Tool for Integrating Variable
Generation
We use the Flexible Scheduling Tool for Integrating VG (FESTIV) in this study. FESTIV
integrates security-constrained unit commitment (SCUC), security-constrained economic
dispatch (SCED), and AGC into one model. This allows multiple timescales to be studied,
accounting for intertemporal coupling between them. The model simulates how most systems
schedule the generation in practice. The three phases of scheduling include commitment,
dispatch, and control. At the finest scheduling interval (the frequency at which AGC is run),
production costs and megawatt imbalances, or ACE, are calculated. This allows for useful
metrics that can show variability and uncertainty impacts as a function of the characteristics that
can influence them. It can also show important trade-offs when objectives are conflicting (e.g.,
reliability versus costs). Few simulation models show reliability metrics and cost metrics
together, which is the unique feature of this integration study. Figure 5 shows the flow diagram
of FESTIV, and how the integrated scheduling models interact. More information is available in
[1] and [7].

Unit status and

Run DASCUC unit start-up for
all units with start
time > tRTCSTART

tRTC no
interval?

yes

Unit status and

Run RTSCUC unit start-up for
all units

t = t+tAGC no
tRTD
interval?

yes

Dispatch
Run RTSCED schedules and
reserve
schedules for
all units

AGC schedule,
realized
Run AGC
generation for
all units,
production
cost, and ACE
Notes: DASCUC, day-ahead SCUC; RTSCUC, real-time SCUC; RTSCED, real-time SCED
Figure 5. Flow diagram for FESTIV. Solid lines represent process flow, and dashed lines represent
data flow.
Each submodel is run at specific intervals defined by the user, and their outputs are used as
inputs to other submodels. Commitments for long-start units are set by the DASCUC, which are

5
then fixed commitments for the real-time submodels. The commitment of all units is then fixed
by the RTSCUC. Next, the dispatch schedules and reserve schedules are set by the RTSCED.
This information is used by the AGC, which adjusts the resources that are providing reserve to
ensure that the ACE, which is defined for the purpose of this report as the instantaneous
imbalance between generation and load, 1 is being driven to zero. Also included is the reserve
pickup (RPU) model (not shown in Figure 5), which has similar functionality to the RTSCUC,
except that it fixes commitment and dispatch for all units, and is triggered by a reliability issue
instead of at fixed time intervals.

The FESTIV model is developed in Matlab, and uses GAMS with the CPLEX solver [8] for the
DASCUC, RTSCUC, and RTSCED algorithms. The interface between the software programs is
modeled after work done in [9]. Because of the variance in timescales between FESTIV’s
submodels, it is important to carefully treat the intertemporal coupling between the submodels to
ensure the most realistic results possible. The FESTIV has four important timing parameters that
can all be configured to any value. These include the interval resolution, I; the submodel update
frequency, t; the scheduling horizon, H; and the submodel processing time, P. Figure 6 shows
how these relate to each of the submodels. As an example, the RTSCUC model is updated every
t RTC minutes, optimizes H RTC schedules at I RTC resolution during each update, and takes P RTC
minutes to solve and send its directions to all of the resources. For comparison, Table 1 shows
common values for these parameters in the production cost models used in the previous studies
and in this study, as well as the values that are typical in actual system operations. The flexibility
of FESTIV allows for any combination of values for these parameters. The submodels differ in
what each is attempting to accomplish (e.g., commitment versus dispatch versus control) and are
at different time resolutions and horizons. FESTIV ensures that constraints affecting the ultimate
outcomes are reflected consistently in all the submodels. As described in [1] the FESTIV model
has been validated and harmonization has been achieved between the submodels and their
objectives toward realistic results. Proper communication between the submodels has also been
achieved.

1
The true definition of ACE, as defined by NERC and similarly by other reliability organizations around the world,
is the combined interchange error and frequency deviation. Because FESTIV does not model frequency, the ACE is
defined by the calculated deviation of generation and load, which is effectively the same as interchange schedule
error when the system’s frequency is at its nominal level.

6
 HDA
IDA
DA
SCUC

12:00

15:00

18:00

21:00
0:00

3:00

6:00

9:00

0:00
HRTC
IRTC

RTSCUC
tRTC
PRTC
RTSCUC

HRTD …
IRTD
RT
SCED

PRTD tRTD

RT
SCED
…

10:00

10:15
9:30
8:45

9:00

9:45
9:15

IAGC = HAGC = tAGC

AGC
AGC
tAGC
AGC
AGC
…
9:00

9:01

Figure 6. Timeline for DASCUC, RTSCUC, RTSCED, and AGC in FESTIV

Table 1. Typical Timing Parameters

Production Cost This Study Actual System
Submodel Parameter Models Operations
I DA (hours) 1 1 1
DASCUC t DA (hours) 24 (once per day) 24 (once per day) 24 (once per day)
H DA (days) 1–7 1 1–2
I RTC (minutes) Typically not used 60 15–60
RTSCUC t RTC (minutes) Typically not used 60 15–60
H RTC (minutes) Typically not used 180 60–300
I RTD (minutes) 60 60 or 5 5–60
RTSCED t RTD (minutes) 60 60 or 5 5–60
H RTD (minutes) 60 60 or 5 5–60
AGC t AGC (seconds) Typically not used 60a 4–6
a
The AGC frequency in our study is limited by the frequency of PV and wind generation data available.

7
To quantify the variability and uncertainty impacts from increasing penetrations of solar at
multiple timescales, both reliability and production cost metrics must be included. As mentioned
previously, the ACE is calculated at each time interval. Variability can cause ACE when the
movement of the imbalance is too fast for the resources to correct, or the scheduling resolution is
inadequate. Uncertainty can also cause ACE when the resources are unable to correct an
unanticipated imbalance. For example, an hourly unit commitment model will eventually face
uncertainty issues because the load and VG output cannot be predicted perfectly. It will also face
variability issues because the load and VG can vary significantly within the hour, often faster
than the resources can move. In our analysis, we use the following three ACE metrics:

• Absolute ACE in Energy (AACEE), which sums the absolute value of ACE at every
interval for the study period (in megawatt-hours)
• CPS2 violations, which are NERC metrics that signify 10-minute periods where the
absolute average of ACE exceeds a threshold called L10 [5]
• σ ACE , the standard deviation of ACE for the study period (in megawatts).
The three metrics can help show the general amount of imbalance, the amount of extreme
imbalance, and the variation of that imbalance. The total production costs can be calculated at
the finest time interval as well, to obtain a realistic value representing the cost of following load,
and to compare the two often conflicting objectives, reducing imbalances and costs.

8
System Description
This study is an addendum to the Western Wind and Solar Integration Study Phase 2 (WWSIS-
2) [10]. We examine a subset of the WWSIS-2 system in a shorter time period. We choose a high
solar penetration balancing area to retrieve more detailed results on the imbalances caused by PV
variability and uncertainty. The APS balancing area of the Western Interconnection is studied in
a high-load, high-PV time period between July 24 and July 28, 2020, at 1-minute time-sampling
frequency. APS has one of the highest penetrations of solar PV using the WWSIS-2 “high-mix”
scenario (with equal amounts of solar and wind energy at 16.5% of annual load each for the U.S.
portion of the Western Interconnection, see Figure 7). Input data, including 2020 projections for
electricity generation costs, capacities, and heat rates were developed from the Western
Electricity Coordinating Council (WECC) Transmission Expansion Planning Policy Committee
(TEPPC) database [11]. Profiles for wind, PV, concentrating solar power (CSP), and load were
developed from WWSIS-2 and the NREL renewable production database [12]. Production cost
model results (computed in PLEXOS) from this study are used to define the interchange
schedules between APS and neighboring regions; the interchange is added to the APS load.
Table 2 presents the APS generating capacity and number of units by generation type.

Figure 7. Solar and Wind placements for the High Mix Scenario. (Yellow – PV, Orange – CSP,
Green – Wind).

Table 2. APS Generation Capacity (2020 Projection)

Combined Combustion
Generation Type Biomass Cycle Turbine Coal Nuclear Other CSP PV Wind
Capacity (MW) 341 4,190 1,080 3,101 4,035 622 4,402 2,925 92
Number of units 3 11 22 9 3 7 8 31 1

9
The generation mix in the APS service area is quite different from that in the rest of the Western
Interconnection. The system has a very high PV and CSP penetration but only one wind plant. It
includes large nuclear plants that are base loaded, 3,000 MW of coal that can ramp but are not
very flexible, and about 5,000 MW of open cycle and combined cycle gas turbines that are fairly
flexible in providing both dispatch capability and regulating reserve. Overall, the system has 99
individual units, including the renewable resources. Figure 8 depicts the load and renewable
profiles for the 5-day period studied. The load profile has the interchange schedules already built
in, which is why it looks different from typical load profiles. There are large nuclear and coal
plants, and APS is a significant exporter (see Appendix A) in this scenario. Note that the CSP is
fully dispatchable and therefore has no uncertainty and only variability that is based on the
scheduling direction. Wind and PV, in addition to load, are the main causes for imbalance. The
net load (load minus wind minus solar PV) and nonrenewable load (load minus wind minus solar
PV minus CSP) are shown as well.

16,000
LOAD + EXPORTS

14,000
PV

CSP
12,000
Load and Generation (MW)

WIND

10,000
NET LOAD

NON
8,000 RENEWABLE

6,000

4,000

2,000

0
0 20 40 60 80 100 120
Time (Hour)

Figure 8. APS load and renewable generation as projected for a 5-day period from July 24 to July
28, 2020
The day-ahead forecast data in the FESTIV model are used at the hourly level. These are also
imported from the WWSIS-2 day-ahead forecast database, for wind and PV plants in the APS
service area. Appendix B contains the day-ahead forecasts for PV compared to actual output. The
APS CSP generation in this study is determined by WWSIS-2/PLEXOS results, linearly
interpolated to 1-minute time intervals. In this study, because of the nature of the high time

10
resolution modeling, a number of assumptions are made that can affect the results in one
direction or the other, as seen in Table 3. However, all of these results are made consistent in
each of the case studies, so that the comparisons are the important part of this research. The
model ignores the transmission constraints within the footprint, as optimal geographic
distribution of wind and solar is outside the scope of this study. Conventional generators follow
AGC schedules perfectly throughout time, which is not realistic. The FESTIV model also does
not model frequency response. These assumptions could underestimate the imbalance results.
However, the simulation model is a model, and therefore cannot simulate the behavior of human
operators making decisions to improve upon the optimal scheduling models. The study is also
taking a lot of fixed data from previous WWSIS2 production cost model runs, with no chance for
those runs to assist in additional flexibility from case to case. These assumptions can make the
reliability impacts look worse than they may be. Finally, there are a number of assumptions that
can have either effect. These include benchmarking, the validity of data used, and the limited
time period of the study. We therefore reemphasize the interpretation of the results in subsequent
sections to be focused on comparison and understanding of changes rather than predictions of the
actual impacts that will occur.

Table 3. Assumptions Made in Simulations

Assumptions that might Assumptions that might Assumptions that can make
make reliability impact make reliability impact reliability impacts look
results look better than results look worse than better or worse than reality
reality reality
No transmission effects are There is no human There is no benchmarking on
considered. intervention or human practices and data to exactly
intelligence in scheduling the match current APS
resources. performance.
Conventional generators are Interchange schedules are Data from TEPPC database
assumed to follow their taken from 5-minute may not exactly match
schedules perfectly. WWSIS2 production cost reality and has not gone
results, not hourly schedules, through a rigorous validation
and are fixed in FESTIV. process.
They cannot change based on
the different sensitivities.
Primary frequency response Concentrating Solar plant Five-day study might not be
(governor response) is not schedules are from WWSIS2 representative for wider time
considered, nor is the production cost results and intervals.
frequency component of the are fixed in FESTIV.
ACE equation.

11
The Impacts of Photovoltaic Variability and
Uncertainty
We performed some base analysis to gain an understanding of the amount of imbalance on the
system without any advanced mitigation strategies. Two types of reserve were used in this case
with existing requirement methods. Contingency reserve is held but not deployed until it is a last
resort and regulating reserve is continuously deployed by the AGC submodel. DASCUC,
RTSCUC, and RTSCED all use co-optimization of both reserve types with energy. Contingency
reserve was required to meet 3% of the hourly load. This is similar to the current WECC
contingency reserve rule [13]. The regulating reserve was held to meet 1% of the hourly load,
both upward and downward regulation, which is similar to how many areas procure their
regulating reserve. The AGC uses the regulating reserve to correct the ACE (i.e., bring it down to
zero). The AGC uses a smoothed ACE with a proportional and integral term. In our case, the
smoothed ACE is averaged over the past 3 minutes to eliminate noise. This typical procedure is
used in today’s systems as described in [14] and [15]. The RTSCUC and RTSCED are run at
hourly resolution in the base case (i.e., t RTC , I RTC and t RTD , I RTD = 60 minutes). Many of the
balancing areas in the Western Interconnection operate in this way today. The RTSCUC can start
and stop any units that have a start-up time less than or equal to 1 hour. The DASCUC operates
at hourly resolution as well, scheduling for a day at a time (H DAC = 24 hours) and repeating once
a day at noon (t DAC = 24 hours). The DASCUC can start or stop any unit. RTSCUC schedules for
a horizon of three intervals (H RTC = 3 hours) and RTSCED schedules for one interval (H RTD = 1
hour). Load forecasts are assumed to be perfect for all submodels. Day-ahead forecasts for PV
and wind are taken from the WWSIS-2 study [10].

For illustrative purposes, perfect hourly forecasts are used in both RTSCUC and RTSCED for a
1-day simulation to give an example of the ACE, illustrated in Figure 9, with the L10 limit
shown for comparison. The total ACE is a result of the subhourly variability of the load, PV, and
wind, because there is no uncertainty. Note that the system is still being scheduled at an hourly
time resolution for the dispatch, so whenever the regulation is not able to keep up with the
variability within the hour, significant imbalance can result. In practice, a system operator will
take actions to correct large imbalances instead of waiting the full 60 minutes to correct the
issue. Because this is not done here, the imbalance may be higher than in reality. This operator
action is explored later in the sensitivity cases. Although the ACE has swings that reach up to
100 MW, the swings are short and therefore not as significant as persistent excessive ACE.
Generally, the ACE stays close to zero.

12
 Figure 9. ACE 1-day simulation. Perfect PV forecasts are used in the real-time simulation.
To get a better idea of the imbalance, Figure 10 shows the ACE averaged over rolling 10-minute
periods for every 1-minute interval to filter out noise. The NERC CPS2 standard requires 90% of
all 10-minute averaged ACE levels to be below the L10 limit for a month. For reference, the
current L10 for APS is 48 MW, and it should be increased to account for the increased 2020 load
in this study. Figure 10 illustrates two instances where the ACE is persistently greater than 50
MW. These instances happen to occur at evening load drop-off, where, along with the morning
load pickup, ACE and frequency issues regularly occur today (see, for example, [16] and [17]).
For the most part, however, the ACE is kept to within –20 and 20 MW.

13
 150

100 Current L10 bounds

Rolling 10-minute average ACE (MW)

50

-50

-100

-150
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time (Hour)

Figure 10. ACE averaged over rolling 10-minute periods for 1-day simulation. Perfect forecasts are
used in the real-time simulation.
Figure 11 shows the generation by category for this 1-day perfect forecasting simulation.
Selected groups of units are shown for their 1-minute generation for the 1-day period. The high
frequency noise seen for the conventional generating units, show the resources being given
signals by AGC to regulate the load, wind, and PV variability that occurs within the hour. The
CSP is used to follow the load pickup and drop-off as well as the interchange scheduling
changes.

14
 5,000

4,500
PV
4,000 CSP
Wind
3,500
Four Corners
Generation (MW)

3,000 Palo Verde

Cholla
2,500
Black Mountain
2,000 Gila River
Ocotil
1,500 SaguarAZ
Sundance
1,000
RedHawk
500 West Phoenix
Yucca
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time (Hour)

Notes: Palo Verde, nuclear; Ocotil, steam; Cholla and Four Corners, coal; Sundance, gas turbine type generators
Figure 11. Generation for 1-day perfect forecast
For our analysis of the PV effects, we now compare two full 5-day simulations. Case 1 does not
include any PV on the system (wind and CSP are kept). PV is added in Case 2 assuming perfect
real-time PV forecasts (day-ahead forecasts still have uncertainty). Figure 12 shows the total
generation and load plus their interchange for the study period for Case 2. Overall, the generation
follows the load quite well. In some instances, the variability is excessive and cannot be
compensated as closely. For example, during hours 40 and 41 (highlighted rectangle), some
excessive variability is seen in the load. The perfect forecasts for those hours were based on the
hourly averaged load, which did not capture the subhourly variation. Similar instances occurred
with the PV variability that was significantly above or below the hourly average. Even with
perfect certainty in the hourly forecast, excessive imbalance can be caused by the resolution of
the scheduling interval. Some areas and studies have been proposing additional flexibility
reserve (e.g., following reserve from Figure 2) that can be used to mitigate this variability [6].

15
 Figure 12. Total generation and load for 5-day perfect forecast base case
We set the L10 level for CPS2 to 65 MW in this study to account for projected load growth in
2020. Case 1 results in 36 violations while the perfect hourly PV forecast case results in 29
violations of L10, leaving a CPS2 score of about 95.2% and 96.0%, respectively. Both scores
meet the NERC criteria. Adding the perfectly forecasted PV reduced the amount of CPS2
violations, which is counterintuitive. There are instances where the load was increasing
significantly within the hour during the morning load pickup. The PV was also ramping up
which acted as additional regulating reserve meeting the ramp requirement and avoided some 10
minute intervals that initially had violated CPS2. While it reduced a few occasions of significant
imbalance, the perfectly forecasted PV did increase the total imbalance from 2,690 MWh to
2,976 MWh of AACEE, or about 22.4 to 24.8 MW of ACE per hour. The PV also increased the
σ ACE from 49.8 to 51.2 MW. As expected, the total production cost of electricity decreased from
$15.040 million to $12.061 million with the zero-cost PV added, or from $12.21/MWh to
$9.79/MWh of load.

The comparisons made when adding PV to the system with perfect hourly forecasts show the
impact of minute-to-minute variability of PV on the system when scheduling at hourly time
frames. Next, we study the real-time uncertainty impacts that occur from solar PV and wind by
adding realistic forecast errors to the real-time scheduling models RTSCUC and RTSCED. The
case study now uses a persistence forecast for wind power and a persistent cloudiness forecast
for solar PV. This means that for wind, the actual output when the RTSCED or RTSCUC model
started is the forecast for 1 hour ahead. For PV, this means that the current cloudiness will
remain constant for the predicted output, but the typical rise and fall based on the sun during
clear sky conditions are incorporated into the forecast. This is described in the following
equation, taken from [18]. P F is the forecast, P is the output, P CS is the clear sky output, and SPI
is the solar power index (P/P CS ). Note that in the base case simulation, ∆t is 60 minutes plus the

16
simulated time it takes to run the model (e.g., I RTD + P RTD ). In total, this results in 65 minutes for
RTSCED and 75 minutes for RTSCUC.

𝑃𝐹 (𝑡 + ∆𝑡) = 𝑃(𝑡) + 𝑆𝑃𝐼(𝑡) ∗ [𝑃𝐶𝑆 (𝑡 + ∆𝑡) − 𝑃𝐶𝑆 (𝑡)]

Table 4 shows the performance characteristics for PV and wind when using these forecast
methods and the 60-minute RTSCED. The wind persistence forecast is not great because of the
long, 60-minute look-ahead horizon. The PV forecast is more accurate (in terms of percentage of
nameplate), but because of the higher installed PV capacity, the uncertainty in its forecast has a
significant effect. The PV forecast is also slightly overforecasting on average.

Table 4. Forecast Performance for 5-Day Period Using Persistence (Wind) and Persistent
Cloudiness (PV) for 60-Minute RTSCED 5-Day Period
Bias MAE (MW / % Standard Deviation
(MW / % of nameplate) nameplate) (MW / % of nameplate)
PV 23.0 / 0.8 55.4 / 1.9 88.4 / 3.0
Wind –1.0 / –0.3 11.6 / 3.9 18.7 / 6.0
Note: MAE, mean absolute error

The forecast uncertainty increases the imbalance. The simulation results show 55 CPS2
violations, resulting in a 92.4% score, which is closer to the 90% limit of CPS2 compliance, but
still above. Overall, there is 3,530 MW of AACEE, resulting in about 29.4 MW of ACE per
hour. The σ ACE results in 61.2 MW, a 20% increase than Case 2. Total production costs are
$12.036 million, slightly lower than those from the perfect hourly forecast case. 2 The system
cost less because it did not spend extra when regulation reserve ran out. This mirrors the
conclusion that although they may have great impacts on reliability, real-time forecast errors
typically do not have significant impacts on production costs when everything else remains equal
[1]. It is important to reiterate that the same load and VG output is realized in Case 2 and Case 3,
and every remaining case, other than Case 1 which uses the same data except without PV.

Table 5 compares these three cases. Appendix C gives the full ACE values for all case studies.
Comparing Case 2 with Case 1 shows the variability impacts of the added solar PV. Comparing
Case 3 with Case 2 shows the uncertainty impacts of PV and wind. For this period, and this
system, the imbalance impacts of forecasting imperfectly are far greater than the effects of
variability of the PV alone. With this base understanding of how variability and uncertainty of
PV affect the system imbalance, we now focus on outlining different strategies to attempt to
reduce the imbalance of these variability and uncertainty impacts, while also attempting to limit
significant production cost increase.

2
In all cases, the total amount of ACE for the study period, must be paid back (if negative) or sold (if positive) at the
average LMP, to ensure total costs are comparable between cases. This is similar to paying back inadvertent
interchange.

17
 Table 5. Comparison of Variability and Uncertainty Impacts
Case CPS2 Score AACEE σ ACE (MW) Production
(MWh) Costs
($million)
Case 1: No PV 35 violations 2,690 49.8 $15.040
95.2%
Case 2: Perfect real-time forecast at 29 violations 2,976 51.2 $12.061
60-minute intervals (variability 96.0%
impacts)
Case 3: Imperfect real-time forecast 55 violations 3,530 61.2 $12.036
at 60-minute intervals (variability and 92.4%
real-time uncertainty impacts)

18
Imbalance Reduction Strategies
Operator Action
In the first three case studies, dispatch and commitment in real time were performed at 60-minute
intervals. Excessive ACE could occur within the dispatch interval, but would not be corrected
until the next hour. In reality, when the system is not in balance, operators would intervene and
direct resources to change their dispatch or commitment to correct the imbalance. This is similar
to operators deploying contingency reserve during contingencies. To simulate operator action,
we use the RPU option built within FESTIV. The commitment and dispatch will continue to run
at t RTC = t RTD = 60 minute intervals; however, if excessive ACE occurs within the hour, the
operator would run the RPU to correct the imbalance. For this, if the ACE exceeds the L10 (65
MW) in either direction, the operators would run the RPU to correct the imbalance immediately
instead of waiting for the next automatic RTSCED or RTSCUC run. The RPU uses a single 10-
minute interval in which commitment and dispatch decisions are made. Table 6 shows the
results.

Table 6. Utilization of Operator Action Using RPU

Case CPS2 Score AACEE σ ACE (MW) Production
(MWh) Costs
($million)
Case 1: No PV 35 violations 2,690 49.8 $15.040
95.2%
Case 2: Perfect real-time forecast at 29 violations 2,976 51.2 $12.061
60-minute intervals (variability 96.0%
impacts)
Case 3: Imperfect real-time forecast at 55 violations 3,530 61.2 $12.036
60-minute intervals (variability and 92.4%
real-time uncertainty impacts)
Case 4: Imperfect real-time forecasts 28 violations 2808 45.1 $12.577
at 60-minute intervals with RPU 96.1%

The operator action reduced the amount of CPS2 violations significantly, down to the level of
perfect real-time forecast case (Case 2). The RPU run also reduced the AACEE to 2,808 MWh
and σ ACE to 45.1 MW, with the σ ACE being lower than that of the perfect real-time forecast case
(Case 2) and no PV case (Case 1). This highlights the fact that the use of operator action to
reduce ACE off the nominal scheduling intervals will mitigate some of the variability impacts as
well as the uncertainty impacts. Since the standard deviation and CPS2 violations were reduced
below Case 1, but AACEE was not, this can suggest that there was a greater focus on reducing
significant imbalances than reducing the overall amount of ACE. In other words, when the
operator noticed the large imbalance he would correct it, but the operator would have ignored
imbalances when ACE was occurring that was not leading to a reliability issue. The use of the
operator action increased the total production costs by a fair amount. Case 2 and Case 3 were
very close in total production costs, but the RPU case resulted in over a $500,000 (greater than
4%) increase, or about $0.50 per MWh. So, even though the use of operator action with the use
of RPU to respond to the excessive imbalances allowed the system to greatly reduce the ACE
impacts, doing so was costly. These off-nominal-time actions occurred very rarely when PV was
not producing, and occurred on average about 30 times a day, or a little more than once an hour.

19
Improved Day-Ahead Forecasts
The case studies so far have focused on the real-time uncertainty impacts of PV and wind. In the
next case study, we fix the day-ahead forecasts to perfectly match the hourly average of PV and
wind generation, but keep the real-time forecasts imperfect as in Case 3 and Case 4. Studies have
shown that perfect day-ahead wind and PV forecasts reduce the production costs of the system
by improving the efficiency of the unit commitment decision [19]. The impact of day-ahead PV
forecasts on imbalance and ACE metrics, however, has not been studied sufficiently. Table 7
shows the day-ahead forecast performance and Table 8 shows the results for the perfect day-
ahead forecast case.

Table 7. Day-Ahead Forecast Performance for 5-Day Period for DASCUC

Bias MAE (MW / % Standard Deviation
(MW / % of nameplate) nameplate) (MW / % of nameplate)
PV 53.6 / 1.8 94.2 / 3.2 149.7 / 5.1
Wind 15.8 / 5.2 37.7 / 12.6 60.8 / 20.3

Table 8. Day-Ahead Uncertainty Impact Improvements

Case CPS2 Score AACEE σ ACE Production Costs
(MWh) (MW) ($million)
Case 1: No PV 35 violations 2,690 49.8 $15.040
95.2%
Case 2: Perfect real-time forecast at 29 violations 2,976 51.2 $12.061
60-minute intervals (variability 96.0%
impacts)
Case 3: Imperfect real-time forecast 55 violations 3,530 61.2 $12.036
at 60-minute intervals (variability and 92.4%
real-time uncertainty impacts)
Case 5: Imperfect real-time forecast 54 violations 3,574 61.2 $11.938
at 60-minute intervals with perfect 92.5%
DA forecasts

The imbalance results are not changed significantly from Case 3, since the real-time forecasts are
identical. The total production costs are reduced to $11.938 million, the lowest total cost yet, but
at less than 1%, perhaps not as much as is normally the case for wind power day-ahead forecast
improvements [19]. This likely results from the fact that day-ahead forecast errors for PV are not
as significant as day-ahead forecast errors for wind. This can be seen with a comparison of Table
7 with Table 4. Although the day-ahead wind forecast error is significantly greater for the day-
ahead forecasts compared to the real-time forecast error (MAE is increased by about 3 times), the
day-ahead forecast error for PV is not as significant (MAE is increased by about 1.5 times). This
could show that the improvements in real-time or short-term solar PV forecasts are more
important than improvements to day-ahead solar PV forecasts.

Increased Dispatch Frequency

Many studies and areas are showing the cost benefits of increasing the dispatch frequency to
faster scheduling intervals, for example [20]. Few, however, have shown the reliability benefits
for reducing imbalance. Faster dispatch frequency has the advantage of having the information at

20
a more granular scale, so that imbalance can be corrected more often through dispatch instead of
limited AGC resources. This has the effect of reducing the subhourly variability impacts. It also
has the advantage of improving real-time forecasts because the look-ahead interval is shortened,
reducing the real-time uncertainty impacts. For this case study, we operate the RTSCED at t RTD
= I RTD = 5 minutes, with the total forecast horizon being 10 minutes (i.e., I RTD + P RTD ). Table 9
shows the RTSCED forecast errors, which are significantly reduced for both the persistent
cloudiness PV forecast and for the persistence wind forecast. In addition, the forecast bias is now
essentially eliminated.

Table 9. Forecast Performance for 5-Day Period Using Persistence (Wind) and Persistent
Cloudiness (PV) for 60-Minute RTSCED 5-Day Period
Bias Mean Absolute Error Standard Deviation
(MW/% of nameplate) (MW/% nameplate) (MW/% of nameplate)
PV (5 minutes) 0.3 / 0.0 20.3 / 0.7 34.6 / 1.1
PV (60 minutes) 23.0 / 0.8 55.4 / 1.9 88.4 / 3.0
Wind (5 minutes) 0.0 / 0.0 7.8 / 2.6 12.3 / 4.1
Wind (60 minutes) –1.0 / 0.3 11.6 / 3.9 18.7 / 6.0

Because the RTSCUC is kept constant at 60-minute resolution and update frequency, additional
commitments of quick-start resources do not help in this case. The results are influenced only by
improving the persistence wind forecast and persistent cloudiness PV forecasts, and by better
dispatching of 5-minute variability. Table 10 shows the results.

Table 10. Impact of 5-Minute Dispatch

Case CPS2 Score AACEE σ ACE (MW) Production
(MWh) Costs
($million)
Case 1: No PV 35 violations 2,690 49.8 $15.040
95.2%
Case 2: Perfect real-time forecast at 60- 29 violations 2,976 51.2 $12.061
minute intervals (variability impacts) 96.0%
Case 3: Imperfect real-time forecast at 55 violations 3,530 61.2 $12.036
60-minute intervals (variability and 92.4%
real-time uncertainty impacts)
Case 6: Imperfect real-time forecasts at 24 violations 2,305 46.2 $11.938
5-minute intervals 96.7%

Using the 5-minute RTSCED operation has significantly improved the imbalance results. There
are only 24 CPS2 violations throughout the 5-day simulation, resulting in a 96.7% score. The
AACEE and σ ACE are also reduced to 2,305 MWh (19.2 MW of ACE per hour) and 46.2 MW,
respectively. The three ACE metrics have been improved the greatest yet. In fact, all metrics fall
to significantly below those seen in Case 1, the case without PV. The likely reason is that the
load variability impacts in Case 1 were reduced by the 5-minute dispatch in addition to the
reduction of the PV and wind variability and real-time uncertainty impacts. The production costs
are also decreased, showing that reducing those impacts with the 5-minute dispatch does not
have high costs.

21
Regulation Reserve Increase
As discussed earlier, regulation reserve is the reserve the AGC uses to correct the minute-to-
minute imbalance in between the dispatch intervals. In the following example, we increase the
regulation reserve for the 5-minute case (Case 6) to see how the increased regulation reserve
affects the imbalance. For further information on how studies have been proposing increases in
regulation reserve, see [4]. In this case study, we apply the methodology used in the WWSIS-2
study, detailed in [18]. It accounts for uncertainty present in the wind and PV by assessing the
95th percentile of annual 10-minute-ahead forecast errors. The 10-minute time frame was again
based on the 5-minute dispatch and process time (as in the previous case, I RTD + P RTD = 10
minutes). The regulation requirement was then determined based on the expected error with
explanatory variables of the solar power index and clear sky ramp. In other words, the extent of
cloudiness and the anticipated typical solar ramp are used to predict the regulation reserve
requirement each hour. The wind component was based on the wind output alone. These
components were added geometrically with each other and 1% of the load from the previous
requirement to arrive at the new hourly regulation reserve requirement. Figure 13 compares the
two reserve methods in both upward and downward directions.

250

200
Regulation Requirement (MW)

150

100

50

0
0 20 40 60 80 100 120
-50

-100

-150

-200 1% Regulation Up
1% Regulation Down
-250 WWSIS-2 Regulation Up
WWSIS-2 Regulation Down
Figure 13. Regulation reserve requirements for Cases 1–6 (1% load) and Case 7 (WWSIS-2 [18])

22
Now Case 7 is run with identical properties to Case 6, but for the increased regulation reserve
requirements. Table 11 gives the results.

Table 11. Impact of Increased Regulation Requirements

Case CPS2 Score AACEE σ ACE (MW) Production
(MWh) Costs
($million)
Case 1: No PV 35 violations 2,690 49.8 $15.040
95.2%
Case 2: Perfect real-time forecast at 60- 29 violations 2,976 51.2 $12.061
minute intervals (variability impacts) 96.0%
Case 3: Imperfect real-time forecast at 55 violations 3,530 61.2 $12.036
60-minute intervals (variability and 92.4%
real-time uncertainty impacts)
Case 6: Imperfect real-time forecasts at 24 violations 2,305 46.2 $11.938
5-minute intervals 96.7%
Case 7: Imperfect real-time forecasts at 22 violations 2055 41.7 $12.005M
5-minute intervals, with increased 96.9%
WWSIS2 regulation reserves

With the increased regulation reserve requirements, the CPS2 violations are reduced slightly
from 24 to 22 violations, resulting in a 96.9% CPS2 score. The AACEE and σ ACE are also
reduced to the lowest levels yet at 2055 MWh (17 MW of ACE per hour) and 41.7 MW,
respectively. The increased regulation reserve requirements raise the production costs by
$67,000. This cost increase is not trivial (0.6%), but perhaps justified if a better CPS2 score is
needed. The intelligent manner of only increasing the requirements during instances when the
PV and wind might have higher variability and uncertainty, has likely led to the minimal increase
in production costs. Then again, if the CPS2 score is already well above the NERC criteria,
perhaps the increased requirements might not be necessary.

23
Conclusion
All case study results are shown in Table 12. The first three cases show the base case impacts of
PV variability and uncertainty, and the next four cases illustrate the effects of the mitigation
strategies that can be used to reduce these impacts. Figure 14 shows the CPS2 scores along with
production costs for all PV cases. The goal is to have high CPS2 scores and low production
costs. For comparison of full 5-day ACE results for all cases, see Appendix C.

Table 12. Case Comparisons

Case VG Forecast Mitigation CPS2 Score AACEE σ ACE Production
Strategy (MWh) (MW) Costs
($million)
Case 1 N/A None (60-minute 35 violations 2,690 49.8 $15.040
(No PV) schedule) 95.2%
Case 2 Perfect real-time None (60-minute 29 violations 2,976 51.2 $12.061
Imperfect day- schedule) 96.0%
ahead
Case 3 Imperfect None (60-minute 55 violations 3,530 61.2 $12.036
real-time and day- schedule) 92.4%
ahead
Case 4 Imperfect RPU (operator 28 violations 2808 45.1 $12.577
real-time and day- action) 96.1%
ahead
Case 5 Imperfect real- Perfect day-ahead 54 violations 3,574 61.2 $11.938
time forecast 92.5%
Perfect day-
ahead
Case 6 Imperfect 5-minute schedule 24 violations 2,305 46.2 $11.938
real-time and day- 96.7%
ahead
Case 7 Imperfect 5-minute 22 violations 2,055 41.7 $12.005
real-time and day- schedule, 96.9%
ahead increased
regulation
reserves

24
 Figure 14. CPS2 scores and production costs for all PV cases
This study, which compares imbalance results with those of production costs, analyzed the
impacts of solar photovoltaic, along with impacts of other sources like load and wind, at multiple
operational timescales. Although many of the previous studies focused primarily on production
costs, this study aimed to show how imbalance and active power control standards were changed
with variability and uncertainty of these sources along with production costs. We used FESTIV
to quantify the variability and uncertainty impacts of solar PV, and then to determine how
numerous mitigation strategies perform in trying to reduce those impacts. The real-time
uncertainty impacts certainly had a greater impact on imbalance than did the PV variability. We
even showed that perfectly predicted PV, when spread over large geographic areas as they were
in this study, could reduce imbalance impacts when the PV was providing additional ramping
capability coinciding with load variability. We also determined that the real-time uncertainty of
PV had little impact on production costs. Day-ahead PV forecasts, however, when improved, had
little impact on variability and uncertainty imbalance impact reduction. They were able to reduce
production costs, but not as significantly as has been seen with wind forecasts. Operator action
seemed to improve the imbalance impacts with a focus on reducing the larger noticeable
imbalances. However, it involved significant incremental production costs. Moving from a 60-
minute dispatch to a 5-minute dispatch appeared to be a good option. This greatly reduced the
variability and uncertainty impacts, even from the case without PV, and did not increase
production costs. Increasing the regulation reserve requirements had a positive impact on
reducing the imbalance impacts even further. The question would be whether or not the

25
improvements in reducing imbalance are worth the marginal increase in production cost, if the
reliability criteria have already been exceeded.

This study is a snapshot into analyzing imbalance impacts with costs. Our work demonstrates
how multiple metrics can be used to measure the reliability impacts and the production cost
impacts together. Each of these metrics may have different meaning and different value to
different balancing areas, and it is important to create strategies that fit the goals of that area. Full
reliability analysis would require significantly more analysis on other aspects such as frequency
response, adequacy, voltage stability, and transient stability. This study considers only the active
power imbalance on a particular balancing area, ignoring the rest of the interconnection (other
than the modeling of the interchange schedules). Further studies can evaluate more scheduling
strategies, including more optimal operating reserve requirement methodologies. As the industry
continues to evolve and integrates higher quantities of renewable resources on the power system,
these methods could be adopted if they prove superior to current methods.

26
References
[1] Ela, E.; O’Malley, M. “Studying the Variability and Uncertainty of Variable Generation
at Multiple Timescales.” IEEE Transactions on Power Systems (27:3), pp. 1324-1333,
August 2012.
[2] Mills, A. et al. Understanding Variability and Uncertainty of Photovoltaics for
Integration with the Electric Power System. Technical Report LBNL-2855E. Berkeley,
CA: Lawrence Berkeley National Laboratory, 2009.
[3] Western Wind and Solar Integration Study. NREL/SR-550-47434. Work performed by
GE Energy, Schenectady, NY. Golden, CO: National Renewable Energy Laboratory,
2010.
[4] Ela, E.; Milligan, M.; Kirby, B. Operating Reserves and Variable Generation.
NREL/TP 5500-51978. Golden, CO: National Renewable Energy Laboratory, August
2011.
[5] North American Electric Reliability Corporation. “Reliability Standards for the Bulk
Electric Systems of North America,” Available: www.nerc.com, June 2012.
[6] Navid, N.; Rosenwald, G.; Chatterjee, D. Ramp Capability for Load Following in the
MISO Markets. Version 1.0. Carmel, IN: MISO, July 2011. Accessed March 18, 2013:
https://www.midwestiso.org/Library/Repository/Communication%20Material/Key%20
Presentations%20and%20Whitepapers/Ramp%20Capability%20for%20Load%20Follo
wing%20in%20MISO%20Markets%20White%20Paper.pdf.
[7] Ela, E.; Milligan, M.; O’Malley, M. “A Flexible Power System Operations Simulation
Model for Assessing Wind Integration.” 2011 IEEE Power & Energy Society General
Meeting, “The Electrification of Transportation & The Grid of the Future. July 24–28,
2011, Detroit, Michigan.”.
[8] GAMS Development Corporation. GAMS: The Solver Manuals. Version 23.9.
Washington, DC: GAMS, 2011.
[9] Ferris, M. MATLAB and GAMS: Interfacing Optimization and Visualization Software.
Mathematical Programming Technical Report 98-10. Madison, WI: University of
Wisconsin Computer Sciences Department, 2005.
[10] D. Lew et al., The Western Wind and Solar Integration Study Phase 2. Forthcoming.
[11] WECC TEPPC. 2009 Study Program Results Report, YEAR. Accessed Sept. 2012:
www.wecc.biz/committees/BOD/TEPPC/Shared Documents/TEPPC Annual Reports/.
2009
[12] National Renewable Energy Laboratory, Wind Integration
Datasets. http://www.nrel.gov/wind/integrationdatasets.
[13] WECC. Western Electricity Coordinating Council Minimum Operating Reliability
Criteria, April 2005. Accessed Sept., 2012:
http://www.wecc.biz/Standards/Development/WECC%200099/Shared%20Documents/
WECC%20Reliability%20Criteria%20MORC%202005.pdf.
[14] Jaleeli, N.; VanSlyck, L.S.; Ewart, D.N.; Fink, L.H.; Hoffmann, A.G. “Understanding
Automatic Generation Control.” IEEE Transactions on Power Systems (7:3), August
1992; pp. 1106–1122.
[15] Report: ENTSO-E, UCTE Operational Handbook Policy 1, Load-Frequency Control
and Performance, March 2009.
[16] NERC. 22:00 Frequency Excursions (Final Report). NERC Frequency Excursion Task
Force, August 28, 2002.

27
[17] Weissbach, T.; Welfonder, E. “High Frequency Deviations within the European Power
System: Origins and Proposals for Improvement.” Proceedings of the Power Systems
Conference and Exposition, March 15–18, 2009, Seattle, Washington.
[18] Ibanez, E.; Brinkman, G.; Hummon, M.; Lew, D. “A Solar Reserve Methodology for
Renewable Energy Integration Studies Based on Sub-Hourly Variability Analysis” 2nd
International Workshop on Integration of Solar Power in Power Systems, November
12–13. 2012, Lisbon, Portugal.
[19] Lew, D.; Milligan, M.; Jordan, G.; Piwko, R. “The Value of Wind Power Forecasting.”
Proceedings of 91st American Meteorological Society Annual Meeting, January 23–27,
2011, Washington, D.C.
[20] Milligan, M.; Kirby, B.; Beuning, S. “Combining Balancing Areas’ Variability: Impacts
on Wind Integration in the Western Interconnection.” AWEA WindPower 2010,
Golden, CO: National Renewable Energy Laboratory, July 2010.

28
Appendix A: Load and Interchange
16,000
14,000
12,000
10,000
MW

8,000
6,000
4,000
2,000
APS Load+Export APS Load Exports
0

29
Appendix B: Day-Ahead Photovoltaic Forecast and
Actual Photovoltaic Generation

30
Appendix C: Area Control Error Charts (Case 1–Case 7
in Order)

Case 1: No PV

31
 Case 2: Perfect real-time forecast at 60-minute intervals (variability impacts)

Case 3: Imperfect real-time forecast at 60-minute intervals (variability and real-time

uncertainty impacts)

32
 Case 4: Imperfect real-time forecasts at 60-minute intervals with RPU

Case 5: Imperfect real-time forecast at 60-minute intervals with perfect DA forecasts

33
 Case 6: Imperfect real-time forecasts at 5-minute intervals

Case 7: Imperfect real-time forecasts at 5-minute intervals, with increased WWSIS2

regulation reserves

34