Predicting Solar Power Production:

(1)

Predicting Solar Power Production:

Irradiance Forecasting Models, Applications and Future Prospects

Green MountAIn ColleGe VerMont

SolAr eleCtrIC Power ASSoCIAtIon wAShInGton, DC

SteVen letenDre, PhD

MIrIAM MAkhyoun MIke tAylor

(2)

2

Acknowledgments

The authors would like to thank the following

individuals for taking the time to share their expertise and insights on solar forecasting: Jon Black, ISO New England; John Tedesco, Green Mountain Power; Steve Steffel, Pepco Holdings; Joshua Stein, Sandia National Laboratory; Richard Perez, SUNY Albany Atmospheric Sciences Research Center; Drake Bartlett, Xcel Energy; Jim Blatchford, CAISO; Dallas Cornier, SDG&E; Jack Peterson, Southern California Edison; Jan Kleissl, University of CA San Diego; Sue Ellen Haupt, NCAR;

Manajit Sengupta, NREL; Tom Hoff, Clean Power Research; David Edelson, New York ISO; Marc Keyser, MISO; William Holmgren, University of Arizona; Mark Ahlstrom, WindLogics Inc.; and John Zack, AWS Truepower. While we appreciate the many insights provided to us, any errors or omissions in the final report are the sole responsibility of the authors. Front cover photo of control room has been provided courtesy of ISO New England.

glossAry of terms

Area Solar Forecast: A forecast of two or more solar energy plants.

DG Solar: Solar energy systems deployed within utility distribution networks.

Direct normal irradiance (DNI): The amount of solar radiation received per unit area by a surface that is always held perpendicular (or normal) to the rays that come in a straight line from the direction of the sun at its current position in the sky.

Diffuse Horizontal Radiation (DHI): The amount of radiation received per unit area by a surface (not subject to any shade or shadow) that does not arrive on a direct path from the sun, but has been scattered by molecules and particles in the atmosphere and comes equally from all directions.

Global Horizontal Irradiance (GHI): The sum of DNI and DHI.

Global NWP: Predictions of global weather patterns.

Insolation: Short for incident or incoming solar radiation, which is a measure of solar radiation energy received on a given surface area and recorded during a given time.

Mesoscale NWP: Predictions of regional weather patterns.

Nowcast: A detailed description of the current weather along with forecasts up to 30 minutes.

Numerical Weather Prediction (NWP): The use of mathematical models of the atmosphere and oceans to predict the weather based on current weather conditions.

Persistence Forecast: Forecasts based on extrapolating current conditions into the future.

Point Solar Forecast: A solar forecast for a single solar energy plant.

Variable Generation: Sources of power that vary based on weather conditions, including wind and solar.

(3)

i

ii

3.1.2 Cloud imagery

...9

3.2 Statistical/hybrid methods to forecast solar irradiance

...10

3.3 Understanding solar forecast errors and variability

...11

3.3.1 Point versus area solar energy plant predictions and PV power output variability

...14

4. Using Solar Power Production Predictions for Utility Planning and Operations

...17

4.1 Transmission-level applications of solar energy predictions

...17

4.1.1 CAISO’s approach to using solar power predictions

...21

4.2 Distribution-level applications of solar energy predictions

...24

4.2.1 Case Study: Hawaiian Electric Company

...25

4.2.2 Case Study: Tucson Electric Power

...27

5. Solar Forecast Provider Survey Results

...29

6. Future Developments and Recommendations

...33

6.1 Current solar forecasting research projects

...33

6.2 Conclusions and Recommendations

...35

(4)

ii

iiii

list of figures

Figure 1: Modern Solar Forecasting System Schematic

...iv

Figure 2: National Solar Radiation Data Base (NSRDB) stations by class

...4

Figure 3: U.S. solar resource map

...4

Figure 4: Time scale relevant to operating power systems

...5

Figure 5: Illustration of solar plant power prediction system

...7

Figure 6: Total sky imager image

...10

Figure 7: Satellite cloud images

...10

Figure 8: Annual Root Mean Squared Error (RMSE) as a function of forecast time horizon

...12

Figure 9: Surface Radiation Network (SURFRAD) Stations Locations

...12

Figure 10: Quantifying Solar Variability

...15

Figure 11: Forecast Solar vs. Actual Output, Northern California

...22

Figure 12: Forecast Solar vs. Actual Output, Southern California

...22

Figure 13: Forecast Solar vs. Actual Output, Northern California

...22

Figure 14: Forecast Solar vs. Actual Output, Southern California

...22

Figure 15: California Independent System Operator Day-Ahead Market Timeline

...23

Figure 16: California Independent System Operator Hour-Ahead Market Timeline

...23

Figure 17: California Independent System Operator Net Load 2012 - 2020

...23

Figure 18: Distribution Impacts of Distributed Generation Solar Hawaiian Electric Company

...26

Figure 19: Impact on System Load for Hawaiian Island

...26

Figure 20: Prototype of Variable Generation Display for System Operators

...26

Figure 21: Prototype Web Interface for Variable Generation Tucson Electric Power Service Territory

...27

Figure 22: Spatial Resolution Capability of Solar Forecast Services

...30

list of tAbles

Table 1: SEPA Recommendations to Advance Solar Forecast Market Development

...iii

Table 2: Annual Error Metrics Day-Ahead and Hour-Ahead Forecasts for

Surface Radiation Network (SURFRAD) Locations

...13

Table 3: Event Example—Up-Ramp or Down-Ramp of 20% Capacity in 1 Hour

...13

Table 4: Solar Forecast End-Users and Potential Applications

...18

Table 5: Solar Forecasting Service Providers Invited to Complete Online Survey

...29

Table 6: Number of Forecast Providers Active in Each Area

...30

(5)

iii

executive summary

Solar energy represents a vast, renewable resource that can be tapped to meet society’s growing demand for electrical energy. Deployment of systems that convert solar irradiance into useful electrical energy has accelerated in the past decade in the U.S. and globally, particularly for photovoltaic (PV) systems— the direct conversion of irradiance to electricity. Utility companies play an integral part in managing this growth and integrating solar generating plants into the existing grid infrastructure and utility operations. Grid integration studies of variable generation resources, including wind and solar, conclude that it is technically feasible for these resources to provide a significant portion of the nation’s energy needs at manageable

costs. These studies recommend operational strategies and market structure changes needed to address increased levels of uncertainty that high-penetration of renewable resources presents to utility companies and grid operators due to their intermittent nature. The use of advanced forecasting of variable generation is one of these essential strategies. Forecasts of future solar energy system output allows grid operators and utilities to proactively manage variable output, and thus integrating solar resources into the existing grid at lower costs to society.

This report provides a review of solar forecasting approaches and how they are being used by grid operators, utility companies, and other market

I. DeVeloPMent oF ForeCAStInG StAnDArDS/ GuIDelIneS

The solar forecasting industry lacks a set of standards or industry guidelines, which acts as a barrier to greater use and understanding of the value of solar forecasting to utility planning and operations. Guidelines or standards with regard to defining forecast time horizons relevant to utility and market operations, measuring and reporting forecast error metrics, and data and communication requirements for modern solar forecasting systems are needed.

II. eConoMIC ASSeSSMent oF SolAr ForeCAStInG VAlue relAtIVe to CoStS

There is a lack of information within the literature on the economic value that solar forecasts could provide to grid operators and utilities. While the value of solar forecasting qualitatively seems quite large relative to the costs associated with producing solar forecasts, there are currently no quantitative analyses to support this notion. Solar forecasts provide value to numerous stakeholders; an attempt should be made to quantify these values to provide the economic basis for expanded use. Furthermore, these analyses could help to identify the incremental value associated with improved forecast accuracy.

III. DIStrIbuteD SolAr workInG GrouP AnD workShoPS

While there are various groups working on different aspects of solar deployment and variable generation forecasting, there is a need for a focused effort on DG solar. A working group comprised of relevant stakeholders should be convened to identify the critical issues and begin a collaborative process to address the unique challenges of DG solar. Periodic workshops should be held to share research and current practices used to understand and manage the impacts of DG solar and the role of solar forecasting.

IV. exPAnDeD enGAGeMent oF PolICyMAkerS AnD reGulAtorS

Policymakers and regulators need to gain a greater understanding of the value that solar forecasting can bring to meeting policy objectives, including renewable portfolio standards. As grid-connected solar deployment expands and forecasting needs increase, utilities, regulators and forecast service providers need to investigate the best mechanisms in terms of reporting system characteristics providing the needed data to produce accurate solar forecasts, standards in terms of allocating the costs incurred by grid operators and utilities associated with solar forecasting, and promoting market structures and scheduling timelines that leverage modern solar forecasting services.

(6)

iv

participants for planning and operations. This report is based on a survey of the literature and interviews with experts from three broad stakeholder groups (ISO/RTO managers, utility managers, and research scientists). Seventeen companies globally were identified that potentially provide solar forecasting services and asked to complete an online questionnaire. Thirteen companies completed

the questionnaire, all of whom work with utilities or related organizations. These companies use state of the art solar forecasting systems to provide customized forecasts to meet their clients’ particular needs.

The research and interviews conducted for this report suggest several initiatives that should be undertaken to accelerate the adoption of solar forecasting in response to the anticipated growth in solar energy system deployment in the coming decade. These initiatives can be grouped into the following four broad categories presented in Table 1.

The report begins with a review of the variable nature of solar and is then followed by an examination of the various ground-level solar irradiance forecasting approaches and models used for each solar production forecast time horizon. Next the various ways in which utilities and grid operators can use forecasts of variable generation are explored, including specific uses of solar forecasts at the transmission and distribution levels. Given that California is leading the nation in the adoption of solar energy, extended coverage of the California Independent System Operator’s (CAISO’s) uses of a central solar forecast is provided. Extended coverage of the Hawaiian Electric Company (HECO) and Tucson Electric Power (TEP) is also provided to highlight efforts at the distribution level to integrate solar forecasting into energy management systems. HECO has the highest PV penetration levels of any utility company and is on the leading edge of exploring how to use solar forecasts to manage high penetration of distributed PV. TEP is a power company and balancing authority in the Southwest that has experienced significant growth in utility-scale and distributed PV deployment. Next the results of the solar forecast provider survey are presented and the report

concludes with a section on future developments and recommendations.

Although the solar forecasting industry is in the early stages of market development and acceptance, several companies currently offer solar forecasting services to various end-users as part of other services they

provide to the renewable energy industry and other weather-sensitive industries. Modern solar forecasting services use sophisticated

physical numerical weather prediction models (NWP), cloud imagery analyses, and statistical manipulations to provide state of the art solar forecasting services covering a range of forecast time horizons (see Figure 1). A point solar forecast is prepared for a particular solar power facility, while an area forecast represents the predicted aggregate output of two or more solar power plants within a defined geographic region. One study finds that solar irradiance forecast error measured in root mean square error, a measure of the standard deviation of the differences between predicted values and observed values, can range from 100 watts – 200 watts per square meter depending on the forecast method and time horizon. To place this in context, the solar resource is approximately 1,000 watts per square meter for a surface perpendicular to the sun’s rays at sea level on a clear day. Irradiance forecast errors vary depending on the overall weather conditions; forecasts on clear sky days are more accurate than

SOLAR PLANT PREDICTION ENGINE (PHYSICAL & STATISICAL MODELS) SOLAR PLANT SIMULATION SOLAR PLANT POWER PREDICTION

DATA

(NWP, CLOUD MOTION, WEATHER, SOLAR PLANT) FIGure 1: MoDern SolAr ForeCAStInG SySteM SCheMAtIC

“Although the solar forecasting industry is in the early stages of market development

and acceptance, several companies currently offer solar forecasting services”

(7)

v

vv

forecasts on partly cloudy days. Furthermore, forecast errors are lower for area solar power forecasts relative to point forecasts due to the fact that aggregation of geographically disperse systems reduces the random impact clouds have on solar generation system output. There is no standardized approach, however, to calculating and reporting forecast error for solar system predictions thus making comparisons between alternative forecasts difficult.

While much can be learned from the experiences gained using wind forecasts in the past decade for utility planning and operations, the distributed nature of solar deployment and the impact that clouds have on solar resource variability create unique and different challenges. By and large distributed (DG), behind the meter solar is invisible to balancing authorities and utility companies meaning they do not have real-time production data to provide situational awareness. The demand-side nature of DG solar shows up through shifting load patterns. While utilities have decades of experience with load forecasting, DG solar results in a new element of variability and uncertainty to system load not captured by existing load forecasting methods. This is a new paradigm for utilities and much work is needed to integrate solar forecasting into existing load forecasting practices. The consistent patterns of load seen over time are changing due to DG solar and thus load forecasting techniques will need to be modified in response, including greater emphasis on intra-day and intra-hour load forecasting.

Utilities and grid operators need to understand total system load, including the load being served by DG solar, to provide situational awareness. This is critical to proactively prepare for the net load impacts of DG solar variability resulting from changing weather patterns and emerging cloud formations. Systems and models are being developed today to make DG solar “visible” to grid operators and utilities. Clean Power Research is currently providing the California System Operator (CAISO) with forecasts of the 130,000+ DG solar energy systems in the State on an experimental basis. DG solar forecasts require data on DG solar systems, most of which is being collected through the net metering approval and interconnection processes, but might not yet be available for forecasting

organizations to access.

In California the CAISO, utility companies and other market participants are using solar forecasting services for a variety of applications. Similar organizations in other regions including New Jersey, Hawaii, Arizona, and Colorado are beginning to experiment with using forecasts of solar plant production for operations and planning purposes. Grid operators and utilities in other regions are actively researching forecasting needs and potential solutions. Regional grid operators and balancing authorities can use forecasts of variable generation to determine the need for operating reserves and for scheduling utility-scale solar plant generation. An emerging use of variable generation forecasting is to predict significant ramping events caused by sudden changes in output from wind and solar generators. Utilities and energy traders can use forecasts of variable generation to develop bidding strategies for hour-ahead and day-ahead markets. Solar energy forecasts of distributed, behind the meter solar systems can be used to create more accurate load forecasts, as grid operators must prepare to meet system load

Clean Power Research (U.S.)

3 Tier (U.S.)

AWS TruePower (U.S.)

JHtech--Solar Data Warehouse (U.S.)

Windlogic, Inc. (U.S.)

Global Weather Corporation (U.S.)

Green Power Labs (Canada)

Iberdrola Renewables, LLC (Spain)

irSOLaV (Spain)

Meteologica SA (Spain)

NNERGIX Energy Management, S.L. (Spain)

Meteocontrol Energy and Weather Services

(Germany)

WEPROG (Germany)

Reuniwatt (France)

ENFOR A/S (Denmark)

Datameteo (Italy)

DNV-GL (Garrad Hassan, Inc.) (Norway)

(8)

vi

net of DG solar. Distribution utilities can also use area forecasts of distributed solar to gain situational awareness to address potential reliability concerns within the distribution networks they manage. In addition, forecasts of variable generation, both central utility scale and distributed systems, can be used for planning purposes. In practice today, solar forecasting is used primarily at the system level to schedule energy production from large utility-scale solar plants. Developing tools that use forecasts of DG solar is still in the research and development stage in the U.S.

Based on questionnaire responses, some forecast providers report challenges in obtaining timely and reliable data from their utility clients to prepare a solar forecast and in some cases a lack of the necessary information technologies to efficiently

integrate with their forecasting systems. In addition, communications between forecast providers and end users can be challenging in terms of clearly articulating forecasting needs and the value that solar forecasting can bring to utility planning and operations. All solar forecasting service providers responding to the online questionnaire indicated ongoing research and development efforts to enhance their solar forecasting services.

Predicting the output of solar energy plants for various forecast time horizons can be a valuable tool allowing grid operators and utilities to reduce the costs of integrating solar sources of generation into the existing grid. This report represents a snapshot in time, as we anticipate rapid developments in forecasting methods and uses in the coming years.

tyPe

Day-ahead, hour-ahead, & intra-hour Forecasts based on time horizon.

Central utility-scale vs. DG solar Large-scale vs. Small scale solar installations.

Probabilistic vs. Deterministic A range of possible solar system output based on probability or a specific output value.

Point vs. Area forecasts Forecasts for one solar plant versus a forecast for an aggregate of geographically dispersed solar energy systems.

Output vs. Rate of change Forecast of power output versus expected change in output over the forecast time step.

(9)

1

1. | Introduction

Deployment of solar energy technologies nationally and globally is accelerating. In October of 2013 solar reached an interesting milestone as being the sole new source of generation added to the nation’s power grid that month at 530 MW. Analysts project that 2013 saw a new record in solar energy deployment in the U.S., with a projected 4,300 MW-dc of PV to come online representing a 27% growth over 2012 installation totals, while cumulative PV capacity is projected to surpass 10 GW-dc (Kann et al., 2013). Furthermore, it is projected that 2013 will be a record year for concentrating solar power (CSP) as over 800 MW-ac is expected to be commissioned (Kann et al., 2013). Economies of scale in manufacturing and continuous technical improvements should result in future cost reductions further strengthening solar energy’s position as an expanding resource to meet society’s demand for electricity.

The variable nature of solar energy production must be actively managed when the level of grid penetration reaches a certain threshold, either on a particular feeder or transmission line or at the system level. As the case with wind power, utilities and grid operators in the U.S. began to use wind power forecasting to manage the variable nature of the resource as installed wind capacity expanded exponentially in the past decade. According to data from the American Wind Energy Association, today there is over 60 GW of installed wind capacity nation-wide. Although the amount of installed solar when forecasting may provide value is not known with certainty and will be different for various utilities and balancing authorities, Pelland et al. (2013) suggest that as a rule of thumb for when solar power forecasts appear to be needed is when yearly solar energy production reaches a level of 1% – 2% of annual energy demand.

Solar forecasting as a decision support tool for utilities, balancing authorities, and power market participants is in the early stage of market development and acceptance. The wind forecasting industry is more developed and many different user groups reap benefits from integrating wind power forecasts into their

decision frameworks. The CAISO was the first ISO/ RTO to start using wind power forecasting for system operations when it introduced forecasting as part of its Participating Intermittent Resource Program (PIRP) in 2004. Three other ISOs/RTOs—Midcontinent ISO (MISO), New York ISO (NYISO), and Electric Reliability Council of Texas (ERCOT)—all introduced centralized wind power forecasting in 2008; PJM Interconnect (PJM) began implementing its wind forecasting system in 2009, South West Power Pool (SPP) in 2011 and the ISO New England (ISONE) recently instituted a central wind forecasting system at the end of 2013.

Many of the power market reforms and grid

management strategies designed to leverage wind farm forecasts to address the variable nature of the resource are directly applicable to solar. However, solar energy has its own unique challenges in terms of forecasting methods and applications. Much of the variability in solar output is predictable, given that the sun changes position throughout the day and throughout the seasons. Rapid changes in ground-level irradiance are driven by clouds, which are inherently difficult systems to model; as discussed later in this report current research and development efforts are focused on improving this aspect of solar forecasting. Another significant difference with solar relative to wind is the distributed nature of solar energy deployment. Much of the installed solar to date is embedded within distribution networks and is largely “invisible” to grid operators, and therefore DG solar systems do not currently have the ability to be curtailed if necessary like larger-scale wind plants. Systems and models are currently being developed to make DG solar “visible” to grid operators, and new inverter technologies and smart meters may provide grid mangers with tools to curtail load and or solar output to address reliability concerns from DG solar variability.

This report provides a review of solar forecasting approaches and how they are being used by grid operators, utility companies, and other market participants for planning and operations. This report

(10)

2

is based on a survey of the literature and interviews with experts from three broad stakeholder groups (ISO/RTO managers, utility managers, and research scientists). Seventeen companies globally were identified that potentially provide solar forecasting services and asked to complete an online questionnaire. Thirteen companies completed the questionnaire, all of whom work with utilities or related organizations. These companies use state of the art solar forecasting systems to provide customized forecasts to meet their clients’ particular needs.

The report begins with a review of the variable nature of solar and is then followed by an examination of the various ground-level solar irradiance forecasting approaches and models used for each solar production

forecast time horizon. Next the various ways in which utilities and grid operators can use forecasts of variable generation are explored, including specific uses of solar forecasts at the transmission and distribution levels. Given that California is leading the nation in the adoption of solar energy, extended coverage of the CAISO’s uses of a central solar forecast is provided. In addition, extended coverage of Hawaiian Electric Company (HECO) and Tucson Electric Power (TEP) is presented, as they have some of the highest levels of PV penetration and are beginning to experience system impacts from solar and other sources of variable generation. As a result, they are on the leading edge of exploring how to use solar forecasts to manage high penetration levels of distributed PV.

(11)

3

2. | The solar Resource

It is widely recognized that the solar resource is vast. Each hour the sun showers the earth with enough potential energy to more than satisfy global energy demand for an entire year. While humans have developed a number of simple and highly technical systems to convert solar energy into useful forms,

solar energy still remains a small part of our overall energy economy. Here we cover the fundamentals of solar irradiance and analyze the variable nature of the resource, which is a key barrier to greater use of the vast solar resource.

2.1 | fundAmentAls of solAr irrAdiAnce

The quantity of fossil fuel sources of energy are assessed

based on sophisticated geologic surveys that quantify the stock of recoverable reserves. Solar energy, in contrast, is quantified based on an assessment of energy flows over various time frames and at particular locations. Typically this flow is quantified as kWh per square meter per time period (day, month, or year) for a particular location on earth. Solar irradiance (irradiance is the power of electromagnetic radiation per unit area incident on a surface) consists of two primary components: direct and diffuse. Direct normal irradiance (DNI) is the amount of solar radiation received per unit area by a surface that is always held perpendicular (or normal) to the rays that come in a straight line from the direction of the sun at its current position in the sky. Diffuse Horizontal Radiation (DHI) is the amount of radiation received per unit area by a surface (not subject to any shade or shadow) that does not arrive on a direct path from the sun, but has been scattered by molecules and particles in the atmosphere and comes equally from all directions. Global Horizontal Irradiance (GHI), also referred to as total solar radiation, is the sum of DNI and DHI. The relevant irradiance metric for concentrating solar technologies is DNI, while GHI is relevant for all non-concentrating solar technologies and applications.1

To properly analyze the potential of solar power, understanding the sun’s trajectory across the earth’s sky is critical. The amount of potential solar energy striking

1 This report focuses primarily on GHI forecasting as DNI forecasting is notoriously difficult and is less critical as many concentrating solar power plants integrate energy storage or natural gas combustion backup to address the variability challenge. Forecasting GHI is an important tool to address the variability of

utili-a collector surfutili-ace vutili-aries utili-as the solutili-ar utili-azimuth utili-angle changes throughout the day. The accepted convention for the solar azimuth angle for solar energy applications is clockwise from due north, so east is 90°, south is 180° and west is 270°. In addition, the solar zenith angle, or solar elevation, defines how high the sun is from the horizon, which changes throughout the year with the seasons. Solar energy collectors can be designed to track the sun seasonally based on changes in the solar zenith angle or daily based on the changing solar azimuth angle (single axis tracker) or both (double axis tracker). Concentrating solar energy technologies utilize duel-tracking systems to position the solar collector devices perpendicular to the sun’s rays to maximize DNI exposure.

The National Renewable Energy Laboratory (NREL) maintains the National Solar Radiation Database, (NSRDB). The updated 1991-2010 NSRDB contains hourly solar and meteorological data for 1,454 locations in the U.S. and its territories. Figure 2 identifies

the 1,454 different NSRDB locations by station classification. The station classification is based on data quality and completeness. Class I and II sites have no missing data, with the former having less than 25 percent of the data for the period of record exceeding an uncertainty of 11 percent (Wilcox, 2012).

NREL produces typical meteorological year (TMY) datasets derived from the 1961-1990 and 1991-2005 NSRDB archives, with the latest version referred to as TMY3. The TMY3 data sets contain hourly values of solar radiation and meteorological elements for a one-year period; these hourly insolation values are

(12)

4

based on averaging data from multiple years and thus represent typical conditions. Their intended use is for computer simulations of solar energy conversion systems and building systems to simplify performance comparisons of different system types, configurations, and locations. Given that they represent typical rather than extreme conditions, NREL suggest that they are not suited for designing systems to meet the worst-case conditions occurring at a particular location. Figure 3 presents the solar resource map for the U.S. derived from TMY3 data files. The annual average daily solar energy striking a flat-plate fixed collector in the U.S.

ranges from an annual average of over 6.5 kWh/m2_/day

in the Southwest, to approximately 3.5 kWh/m2_/day

in the Pacific Northwest. Solar is an abundant energy resource, but not as concentrated relative to fossil fuel energy sources.

In the early 1990s solar resource assessment techniques were developed using satellite-derived cloud cover

data. Through well-tested algorithms cloud cover data from geostationary weather satellites can be used to estimate ground-level irradiance at a particular location at any given time; this method was adopted as part of the 2010 NSRDB update. Satellite-based resource assessment techniques have been rigorously validated using ground-level irradiance measurements (Perez and Seals, 1997). The availability of satellite-derived solar resource data providing time and location specific irradiance data freed analysts from the limitations of average solar resource data, thus allowing detailed analyses of solar energy’s potential contribution as a grid-connected resource (Letendre and Perez, 1996). These techniques have been further refined and developed by NREL to create simulated solar production datasets for the continental U.S. for use in large-scale solar integration studies. As discussed below, the use of satellite cloud images for irradiance forecasting is a fundamental technique used today for solar power predictions up to 6 hours ahead.

2.2 | the vAriAble nAture of solAr energy

As described above the solar resource striking a collector

surface, either fixed or tracking, is variable throughout the day and throughout the year. To a large degree the variability associated with the position of the sun in the sky impacts solar power generating systems in a relatively uniform manner and can be readily

anticipated. The stochastic, short-term variability component of irradiance striking the earth driven by clouds and weather conditions creates uncertainty in output from solar energy systems, which is of greatest concern to grid operators and utility companies.

Changes in the sun’s position during the day can result in

FIGure 3: u.S. SolAr reSourCe MAP

(Source: http://www.nrel.gov/gis/images/eere_pv/ national_photovoltaic_2012-01.jpg)

FIGure 2: nSrDb StAtIonS by ClASS

(13)

5

a 10% - 30% change in output over a 15-minute interval for single-axis tracking photovoltaic (PV) plants. In contrast, changes in insolation at a particular point can exceed 60% of the peak insolation in a matter of seconds (Mills et. al, 2009). The impact on a PV plant’s output from passing clouds, however, depends on the size of the system, cloud speed, cloud height, and other factors (Mills et al., 2009). Larger arrays will experience less variability from passing clouds as compared to smaller arrays. For PV arrays with a rated capacity of 100 MW, the time it takes to shade the system will occur over minutes not seconds.

Electricity is a unique form of energy that must be both generated and consumed simultaneously. Although electricity can be stored for later use, the current grid has very little capacity to store electricity and round-trip efficiency losses associated with energy storage are significant. Thus, grid operators have implemented systems and procedures that consistently match power supply with power consumption to achieve reliability standards established by national and regional electric power reliability organizations.

Variability is an inherent characteristic of the electric power system given that the demand for power varies throughout the day and the seasons. Although there is uncertainty associated with load forecasting, experience has afforded the industry a level of comfort around load forecast errors and their trends. Furthermore, system operators must also manage the uncertainty associated with generators and transmission assets that may not perform as anticipated due to technical malfunctions or weather-related factors. System operators and planners rely primarily on operating reserves, which are additional generating units scheduled beyond what is anticipated to meet forecast loads, to address the variability and uncertainty of power systems (Mills et al., 2009). Figure 4 illustrates the time scales relevant to operating power systems and the strategies used to manage the variability and uncertainty associated with each distinct time interval.

Grid operators have decades of experience operating power systems with conventional, dispatchable power plants fueled with coal, natural gas, nuclear or hydro resources. Dispatchable resources tend to use fuels

that can be stored or otherwise predictably relied upon over relatively short time frames, and technologies that have relatively known, quantifiable outage rates associated with them. Undoubtedly, greater use of intermittent generation including wind and solar will add a new dimension of variability to the power system. In response to rapid growth in wind and solar generation on regional grids and state-level renewable energy mandates, the U.S. DOE and the NREL teamed-up to analyze the impacts of increasing wind and solar energy on the western and eastern electric grids. Initiated in 2007, the Eastern Wind Integration and Transmission Study (EWITS) evaluated the power system impacts associated with 20% - 30% wind energy penetration on the Eastern Interconnection. The Western Wind and Solar Integration Study (WWSIS) analyzed the operational implications associated with up to 35% wind and solar energy penetration on the Western Interconnection. The high renewable

SYSTEM LOAD (MW)

TIME (HOUR OF DAY)

0 4 8 12 16 20 24 SECONDS TO MINUTES REGULATION TENS OF MINUTES TO HOURS LOAD FOLLOWING DAY SCHEDULING C C B A B A

FIGure 4: tIMe SCAle releVAnt to oPerAtInG Power SySteMS

(14)

6

energy penetration scenarios in these studies were evaluated in terms of integration costs as well as potential transmission constraints and operational changes that might arise. Phase II of the WWSIS was completed in September of 2013.

Bird and Lew (2012) provide a summary of the lessons learned from the Phase I EWITS and WWSIS studies, and others, on integrating variable generation into the U.S. bulk power system. The key findings are:

_{Higher penetrations of variable renewable} generation are manageable: The Eastern and Western studies found that by 2025, with operational changes and expanded transmission access, high penetrations—up to 30% or 35%—of variable renewable generation are technically feasible to integrate so that load and generation can be balanced for all hours of the year.

_{Larger balancing areas and geographic diversity} of renewable resources are preferred to minimize impacts: Both the Eastern and Western studies as well as other recent studies show that large operating areas and sufficient transmission access are effective measures for managing wind energy. Larger operating areas provide benefits of reducing the variability of the wind and solar when it is spread over a larger geographic area and providing access to more resources for balancing the system. _{Sub-hourly scheduling and planning are preferred:}

Both studies reinforce the idea found by other integration studies that sub-hourly scheduling and dispatch can aide in the integration of higher penetrations of wind and solar on the grid. _{Using advanced forecasting can reduce costs:}

Forecasting is an important tool for reducing the uncertainty and managing the variability associated with wind and solar generation. While forecast accuracy is increased closer to the time of actual generation, incorporating day-ahead forecasting into the process of committing generation units can help operators mitigate the uncertainty of wind generation.

_{Demand response can be cost-effective:}_Demand response is one mechanism for providing systems with operational flexibility. It can be useful for

assisting with the integration of variable renewable generation by helping to address large loss of generation events.

_{Integration costs are manageable:} Systems can incur costs due to increased balancing resources needed to manage variable renewable energy generation on the system. The costs can include the need for additional operating reserves, changes in dispatch, and new transmission.

_{Renewables can play a role in resource adequacy –} capacity value: While wind and solar are primarily energy resources, they can contribute to resource adequacy of the system. However, in contrast to conventional generators, only a portion of the capacity of wind and solar facilities can be relied on to be available to meet peak demand.

(Bird and Lew, 2012, pages 4 – 8)

Phase II of the WWSIS evaluated the increased costs associated with increased cycling and ramping of conventional generation resources in response to high penetration of wind and solar, and the associated emissions impacts. The study concluded that increased cycling costs assuming 33% wind and solar penetration are small relative to the reduced fuel costs, which were estimated to be $0.14 – $0.67/MWh and $28 – $29/ MWh respectively. In terms of emissions impacts the study found that cycling had very little (<5%) impact on the CO2, NOX, and SO2 emissions reductions from

wind and solar and that wind- and solar-induced cycling can have a positive or negative impact on emissions rates depending on the generation mix and wind and solar penetration level.

The high renewables penetration studies to date highlight the importance of advanced forecasting of variable generation, which allows grid operators to anticipate and actively manage the uncertain nature of wind and solar plants. Experience to date in the wind sector suggests that significant economic benefits are realized from integrating forecasting into scheduling systems and market operations. The economic benefits of forecasting are a result of reduced imbalance charges and penalties, competitive knowledge advantage in real-time and day-ahead energy market trading, and more efficient project construction, operations, and maintenance planning.

(15)

7

3. | Forecasting solar Irradiance

Solar irradiance forecasting provides a critical input to predicting a solar power plants’ output at various points in the future. As discussed above, solar plant predictions provide grid operators, utilities, and market participants data for use in decision support tools, including scheduling reserve capacity or developing bidding strategies for hour-ahead and day-ahead wholesale power markets. These different uses of solar power predictions require different types of forecasts. For example, a forecast may apply to a single PV system (point forecast) or refer to the aggregation of large numbers of distributed PV systems spread over an extended geographic area (area forecast). A forecast that focuses on the rate of change in solar power output may be needed for decision support tools designed to predict significant ramp events on regional grids. Solar power prediction methods are generally characterized as physical or statistical, however in practice the lines between these approaches is blurred. Physical approaches explicitly model physical atmospheric phenomenon as part of the irradiance prediction using numerical weather prediction (NWP) models or sky images. Statistical approaches predict irradiance from training and statistically-derived values. For example, a physical approach

may use developing cloud vector-based forecasts via interpolation of recent consecutive sky images and a statistical approach may use current and historical power output alone to predict future output. NWP model outputs can also be fed into statistical models to improve the forecast, but often it is the case where physical NWP models have to be developed and run at higher spatial and temporal resolutions than the general NWP models. The solar plant simulation component of a forecasting system is often a separate module that uses the required meteorological forecasts (e.g., irradiance, ambient temperature, and wind speed) as inputs, and can also be in the form of a physical model with known specifications that can be modeled or a statistically-derived one. Solar forecast providers in practice draw from multiple forecasting methods to tailor solar plant power predictions to end-user needs. A persistence forecast is often used as the benchmark for assessing the skill of various solar plant power prediction models. A persistence forecast simply extrapolates current plant output into the future, with perhaps adjustments based on changing sun angles. Figure 5 illustrates the key components of an advanced solar plant power prediction system. The solar plant simulation component is required to predict solar

SOLAR PLANT PREDICTION ENGINE

NWP DATA CLOUD MOTION DATA WEATHER DATA SOLAR PLANT DATA

SOLAR PLANT SIMULATION PHYSICAL MODELS (IRRADIENCE/MODULE TEMPERATURE) STATISTICAL MODELS SOLAR PLANT POWER PREDICTION NORMALIZED POWER TIME 0 2 4 6 8 10 12 14 16 18 20 22 24 FIGure 5: IlluStrAtIon oF SolAr PlAnt Power PreDICtIon SySteM

(16)

8

plant power production using a set of mathematical equations that combines solar resource information with solar system specifications to produce power output predictions. The majority of solar energy plants receive incoming irradiance on a tilted plane. As a result, the forecasted GHI has to be converted according to the orientation and tilt of the solar plant collectors to model the irradiance actually utilized by the solar energy plant. Solar energy power predictions can be provided to end-users either in a deterministic or probabilistic format, with the former being a specific

value for solar energy production at a particular time step (15-minutes or one hour) and the latter being a range of values based on probability theory. While still in the research phase, forecasts could also provide information on the expected variability of output for the time period for which energy output is being forecasted. For example, an hour ahead forecast may predict 350 MW on average for the next hour for a 500 MW plant; this does not indicate how variable that output will be during the forecast hour.

3.1 | PhysicAl methods to forecAst solAr irrAdiAnce

The numerical weather prediction (NWP) method

draws from one or more of the several existing global weather models to forecast plane of array irradiance and provide the necessary meteorological data to estimate back of module temperature, and is generally accepted as the most accurate method for forecast horizons of 6 hours and beyond. Cloud imagery techniques, using either satellite cloud images or images from terrestrial-based sky imaging devices, predict cloud motion into the future and then apply established algorithms to predict plane of array irradiance; these

methods are generally understood to produce the most accurate forecasts in the 0 – 6 hours ahead time horizon. Solar forecast service providers integrate various physical and statistical methods (dictated by the forecast time horizon) to forecast plane of array irradiance. Post processing of physical data is used to eliminate systematic errors and “tune” the solar plant simulation model based on a variety of relevant data including historic solar plant production data; these post-processing methods are referred to as model output statistics (MOS).

3.1.1 | numericAl weAther Prediction (nwP)

NWP uses the power of computers to make weather

forecasts using complex computer programs run on supercomputers that generate predictions of many atmospheric variables such as temperature, pressure, wind, and rainfall. Three-dimensional models of the atmosphere and oceans are used to predict the weather based on current weather conditions; initialization is the term used to describe the process of entering observation data into the model to generate initial conditions. A number of global and regional forecast models are run in different countries worldwide, using current weather observations relayed from devices known as radiosondes on weather balloons, radar, weather satellites and ground station measurements as inputs. NWP model runs are typically initiated two to four times per day, for example at 0, 6, 12 and 18

UTC (Universal Time Coordinated). In order to limit computational requirements, global NWP models have relatively coarse resolutions with grid spacing on the order of 40 km to 90 km. Mesoscale NWP models cover a limited geographical area with higher resolution, which attempt to account for local terrain and weather phenomena in more detail than global models. The weather forecasts that we rely on daily are the results of one or more of these NWP models. Perez et al. (2013) combined and evaluated three independent validations of GHI multi-day NWP forecast models that were conducted in the U.S., Canada, and Europe. Two NWP models were used in each of the three validation efforts analyzed— the European Centre for Medium-Range Weather Forecasts (ECMWF) global model and the Weather

(17)

9

Research Forecasting (WRF) mesoscale model

initialized with Global Forecast System (GFS) forecasts. They concluded that the GFS-based WRF model forecasts do not perform as well as global forecast-based approaches such as the ECMWF model and that the simple averaging of the various models’ outputs perform better than individual models. Combining multiple NWP forecasts is a standard approach for improving forecast performance, which is referred to as an ensemble forecast. Ensemble forecasts can

also be produced by varying the initial conditions or physical parameterizations within a single NWP model to yield a probabilistic forecast. Perez et al. (2013) also found that post-processing using MOS techniques significantly improve forecast performance for mesoscale models in the U.S. Statistical tools applied to forecasts generated using the Canadian Global Environmental Multiscale Model (GEM) based on GFS forecasts achieved similar performance as the ECMWF forecasts in North America and in Central Europe.

3.1.2 | cloud imAgery

The second broad category of physical solar power prediction approaches uses cloud images to predict cloud movement, and the associated impacts on plane of array irradiance and back of module temperature into the future at a particular geographic site. This approach reduces forecast error over NWP forecasts in the 0 – 6 hour-ahead (intra-day) timeframe. Higher spatial resolution (10-100 meters) and shorter sampling rates (30 seconds) of cloud images are achieved using total sky imaging (TSI) devices relative to satellite cloud images (1 km and 15 minutes respectively).

TSI-based forecasting is used to predict solar plant power output in real time (nowcast) up to 10 - 30 minutes ahead (intra-hour) through advanced image processing and cloud tracking techniques. Due to the limited view capabilities, TSI devices are unable to detect clouds that will impact a site beyond the 30 minute time horizon. TSI devices take an overhead image of the surrounding sky (Figure 6), which provides the images that are then processed to generate a forecast. In general this approach assumes the opacity, direction, and velocity of movement of the clouds in the future is consistent with the initial conditions observed through the TSI device. Irradiance is predicted based on current cloud shadows and then the cloud shadows are relocated forward in time based on cloud velocity and direction to generate the forecast (Pellend et. al, 2013). One total sky imager, depending on cloud height, can deliver an image for 5 - 10 square miles under cloudy sky conditions and 15 square miles in totally clear sky conditions. Thus, one imager can be used for a multi-MW solar farm, but several

TSI devices would be needed for a multi-hundred MW farm. Solar power predictions using a TSI-based system are the best technique to predict short term ramps for individual solar power plants. However, the need for managing ramps on a large grid system decreases with geographic distribution (discussed in greater detail below). Thus, total sky imagers may have limited applicability to island regions, small balancing authorities, and perhaps to independent solar power producers that must meet certain ramp rate parameters. Given that TSI devices are not ubiquitous, cloud images obtained from geostationary satellites (Figure 7) provide an alternative way to obtain cloud images, which are available for the entire globe albeit with lower spatial resolution and longer sampling rates than TSI images. Clouds reflect light into the satellite leading to detection and the ability to calculate the amount of light transmitted through the cloud, thus generating cloud images with the opacity characterized that are processed in a similar manner to cloud images generated using TSI devices. The direction and speed of clouds can be assessed using subsequent satellite images to predict future impacts on plane of array irradiance for a particular solar plant site. The lower spatial and higher temporal resolution for satellite images causes satellite forecasts to be less accurate than sky imagery on intra-hour time scales, but extensive comparisons or combinations of the two approaches have not been conducted (Pellend et al., 2013). The satellite imagery technique for solar power predictions is considered the best forecasting technique from 1 to 6 hours in advance of the forecast time horizon.

(18)

10

3.2 | stAtisticAl/hybrid methods to forecAst

solAr irrAdiAnce

A persistence forecast is the most basic statistical approach to predict the future output from a solar plant; past time series data is used to forecast solar plant power output in the future, with minor adjustments based on the sun’s position in the sky. Predicting solar plant production using purely statistical methods is not generally part of modern solar power prediction systems. However, hybrid approaches make use of advanced statistical techniques to correct for known deficiencies associated with different forecasting methods through adjustments for model biases or automated learning techniques.

As mentioned above, MOS uses statistical correlations between observed weather elements and climatological data, satellite retrievals, or modeled parameters to obtain localized statistical correction functions. This allows for the enhancement of low-resolution data by considering local effects including topographic shading, or for correcting systematic deviations of a numerical model, satellite retrievals, or ground sensors. One of the major disadvantages of statistical methods is the requirement for accurate historic datasets used to develop statistical correlations separately for each location. This means that MOS-based forecasts are not immediately available for larger areas or for locations

without prior measurements, such as most non-urban solar power plants (Kleissl, 2010).

Purely statistical approaches to solar power predictions do not attempt to simulate future production using irradiance and module temperature; their starting point is a training dataset that contains a variety of inputs that could include NWP outputs, ground station or satellite data, historic solar plant production data etc. The dataset is used to train models—such as autoregressive or artificial intelligence models—that produce a forecast of solar plant output over a specified future timeframe based on past inputs available at the time when the model is run (Pelland et al., 2013). The line between physical and statistical methods is blurred in today’s modern solar power prediction systems. Best practices adopt a hybrid approach whereby physical models produce irradiance and module temperature forecasts used in a solar plant simulation model, the output of which is then subject to statistical post-processing to improve accuracy. Past solar plant output is an important source of data to train MOS post-processing models thereby improving future forecast skill. The best statistical approaches make use of the knowledge from physical models to select input variables used to train a statistical

FIGure 7: SAtellIte ClouD IMAGeS

(Source: Perez et al., 2010a) FIGure 6: totAl Sky IMAGer IMAGe

(19)

11

forecasting model. As discussed later in this report, IBM was awarded a grant from the U.S. DOD in 2012 to advance the state of the art in statistical applications to improve solar forecasting. The ability to leverage new computational techniques including machine learning

and the rapid statistical processing of large amounts of data using super computers will likely have a significant impact on the quality of solar power prediction models in the coming decade.

3.3 | understAnding solAr forecAst errors And vAriAbility

Ultimately, a forecast’s accuracy is judged by the

difference between the solar plant output prediction and the actual plant output over some time interval (hourly) and some specified time period (year). Much research has been done to characterize the error of various forecasting approaches by comparing solar irradiance forecast results with ground-measured irradiance to assess the accuracy of various solar forecasting approaches over multiple time horizons. The three standard error metrics reported in the literature include the root mean square error (RMSE) measuring a model’s overall performance (especially where extreme events are a concern), the mean absolute error (MAE), another measure of overall performance, and the mean bias error (MBE) a measure of a forecast’s bias, meaning its tendency to over or under predict. These error metrics for solar irradiance are given in watts per square meter, but can also be presented in relative (%) terms; PV forecasts are commonly normalized by dividing by the rated capacity of the PV system for which the forecast was generated.

In terms of irradiance forecasting, the RMSE represents the square root of the sum of the squares of the

difference between hourly forecasted irradiance and hourly measured irradiance over a one year period. The MAE is defined in the context of irradiance forecasting to be the sum of the absolute values of the differences between hourly forecasted irradiance values and hourly measured irradiance values over a one year period divided by the number of observations (8,760). The MBA is calculated similarly to the MAE with the exception that the differences between hourly forecasted irradiance values and hourly measured irradiance values are simply summed and not

converted to absolute values, thus this error metric will indicate whether forecast tends to over or under predict the solar resource.

Climate conditions and forecast horizons have been shown to directly impact the accuracy of irradiance forecasts. Clear skies are relatively easier to forecast resulting in lower forecast errors than partly cloudy and overcast weather conditions. Forecast accuracies also decrease as the forecast time horizon increases. The designation of the best solar forecasting approach given forecast time horizon identified earlier in the report is based on research that characterizes the error of different forecasting approaches for various forecast time horizons. Figure 8 is an important graph produced by Perez et al. (2010b) demonstrating that satellite cloud motion forecasts perform better than NWP forecasts in the 1 – 6 hour time horizon, beyond which NWP produces a superior forecast. In addition, Figure 8 illustrates how forecast error increases as the forecast time horizon extends into the future.

The forecast algorithms evaluated in Figure 8 were used to forecast hourly GHI, which were validated against ground measurements for seven climatically distinct locations in the U.S. for one year. The short term forecasts that extend up to 6-hours ahead are based upon cloud motion derived from consecutive geostationary satellite images. The medium term NWP forecasts extend up to 6-days ahead and are modeled from gridded cloud cover forecasts from the U.S.

“the three standard error metrics reported in the literature include the root mean square error (rMSe) measuring a model’s overall performance (especially where extreme events are a concern), the mean absolute error (MAe), another measure of overall performance, and the mean bias error

(Mbe) a measure of a forecast’s bias, meaning its tendency to over or under predict.”

(20)

12

National Digital Forecast Database (NDFD). These two methods are compared against two benchmark forecasts, including a simple persistence forecast and a satellite nowcast (irradiance estimates at the time the satellite images were taken). Hourly forecasts were tested against irradiance data from the seven Surface Radiation Budget Network (SURFRAD) stations across the U.S. as depicted in Figure 9, which cover several distinct climatic environments.

Table 2 contains RMSE metrics for hour-ahead and day-ahead solar forecasts for each of the seven SURFRAD stations (Perez et al., 2010b). The RMSE

statistics and the mean observed GHI are presented in watts/m2_{. The table also includes a clearness index in}

percentages for each location, with higher percentages indicating relatively clear skies in comparison to lower values. The data presented here is based on the entire validation period of one year. As observed from the data presented in Table 2, hour-ahead solar forecasts have lower errors than day-ahead forecasts.

The standard error metrics used to assess solar forecasts provide different perspectives on forecast quality and their published values depend on the time span of the validation analysis, use of sunlight hours only, and data aggregation techniques which could make comparisons between forecasts meaningless (Kostylev and Pavlovski, 2011). Hoff et al. (2012) evaluate different approaches to reporting relative error metrics for irradiance forecasting offering recommendations for the most appropriate approach using a subjective grading scale including whether:

The method is commonly accepted;

_{Is simple to understand;}

Depends on the 24-hr. versus daytime only

distinction;

Depends on the data selection threshold (the

value below at which data is excluded for night time hours); and

_{Is affected by outliers.}

Based on these criteria, they conclude that the relative error metric of MAE divided by the average measured irradiance values over the validation period provides the best practical measure of relative dispersion error for solar resource forecasts. They also suggest that the RMSE divided by solar plant capacity is desirable because it is commonly accepted and is simple to understand; furthermore the wind industry has already adopted this method.

An alternative approach to assessing the performance of a forecasting system is to use categorical statistical measures, which assess whether or not the forecast accurately predicts specific events. Table 3 was taken from a recent DOE presentation outlining such a framework. Here the event under consideration would be ramps (up or down) equal to 20% of the rated solar capacity occur in a given hour. Simply

1 2 3 4 5 6 2 3 4 5 6 7 HR HR HR HR HR HR DAY DAY DAY DAY DAY DAY

RSME (W/M

-2)

FORECAST TIME HORIZON

MEASURED PERSISTENCE CLOUD-MOTION FORECAST NDFD SATELLITE REFERENCE 7-SITE COMPOSITE 300 250 200 150 100 50

FIGure 8: AnnuAl rMSe AS A FunCtIon oF ForeCASt tIMe horIzon

(Source: Perez et al., 2010b)

DESERT ROCK, NV FORT PECK, MT SIOUX FALLS, SD BOULDER, CO GOODWIN CREEK, MS PENN STATE, PA BONDVILLE, IL

FIGure 9: SurFrAD StAtIon loCAtIonS

(21)

13

adding up the number of times the forecast predicted the correct outcome and divide by the total number of predictions provides an assessment of the forecasting system in terms of the percentage of times the forecast predicted the outcome correctly.

Ultimately, the appropriate error metric must meet the needs of the end-user and be relevant for the particular use of the solar power prediction. Grid operators need metrics that accurately reflect the costs of forecast errors, while solar resource research scientists require indicators of the relative performance of different forecast models, and of a

single model under different conditions. Solar forecast user groups should engage in a collaborative process to provide guidance to forecast providers in terms of what they need to know relative to the quality of the forecast services being provided.

loCAtIon

MeAn obSerVeD

GhI CleArneSS InDex

rMSe hour-AheAD ForeCASt rMSe DAy-AheAD ForeCASt DESERT ROCK, NV 498 w/m2 _90% _{80 w/m}2 _{125 w/m}2 FORT PECK, MT 357 w/m2 _75% _{94 w/m}2 _{148 w/m}2 BOULDER, CO 369 w/m2 _71% _{120 w/m}2 _{188 w/m}2 SIOUX FALLS, SD 364 w/m2 _76% _{68 w/m}2 _{140 w/m}2 BONDVILLE, IL 349 w/m2 _69% _{85 w/m}2 _{151 w/m}2 GOODWIN CREEK, MS 397 w/m2 _76% _{80 w/m}2 _{149 w/m}2 PENN STATE, PA 323 w/m2 _66% _{86 w/m}2 _{141 w/m}2 tAble 2: AnnuAl error MetrICS DAy-AheAD AnD hour-AheAD ForeCAStS For SurFrAD loCAtIonS

(Sources: Perez et al., 2010b)

ContInGenCy tAble obSerVAtIon Fore CAS t yeS no totAl

yeS Hit False Alarm Forecast Yes

no Miss Correct Null Forecast No

totAl Observation Yes Observation No Total Sample Size

Example: Accuracy = (Hits + Correct Negs)/Total

(22)

14

3.3.1 | Point versus AreA solAr energy PlAnt Predictions

And Pv Power outPut vAriAbility

As discussed in the next section, there are a variety of solar forecast end-users and applications. Certain end-user groups, in particular balancing authorities, are concerned with the aggregate impact that the fleet of solar power systems within their geographic footprint has on system reliability and market efficiency. The diminished variability of the output from multiple wind turbines within a wind farm relative to a single wind turbine is well understood and documented. Variability in power production further declines when a broader geographic perspective is adopted including multiple geographically dispersed wind farms. The underlying weather changes that impact wind speeds do not impact geographically dispersed wind farms uniformly or at the same time, thus reducing the aggregate impact on variability of concern to grid operators.

Similar to wind plant output, the variability of multiple solar power plants when viewed in aggregate is much less than a single solar power plant. There is currently no standard for characterizing PV power output variability. Perez et al. (2011) developed a set of empirical models capable of extracting metrics quantifying the short-term variability of the solar resource based upon site/time specific satellite-derived hourly irradiance data. The model is satellite-derived from over 92,000 experimental hourly data points at twenty-four sites in the U.S. The model returns four metrics characterizing intra-hourly variability; these include the standard deviation of the global irradiance clear sky index (ratio of actual insolation to the clear sky insolation), and the mean index change from one time interval to the next, as well as the maximum and standard deviation of the latter. The variability time scales addressed in the model are 20 seconds, 1-minute, 5-minutes and 15 minutes.

The variability within a given time interval (e.g., one-minute changes over one hour) is reduced when multiple geographically dispersed solar energy systems are aggregated. Hoff and Perez (2010) find that the power output variability (one minute

time interval) from 1,000 MW of dispersed 4-kW residential PV systems corresponds to the output variability associated with 0.2 percent of what it would be if the PV capacity was concentrated in a single location.

Hoff and Perez (2011) develop a method to analyze irradiance correlation coefficients between two locations as a function of distance, time interval, and other parameters. The left side of Figure 10 illustrates measured 10-second irradiance data, while the right side of the chart illustrates the change in irradiance using a 10-second time interval for a network of 25 weather monitoring stations in a 400-meter grid located in Cordelia Junction, CA on November 7, 2010. The light gray lines represent irradiance and variability for an individual station, while the red lines correspond to average irradiance and variability across 25 locations. Based on this analysis, variability is reduced by more than 70% when comparing the single-site measurements with the average of the measurements from the 25 sites. Correlation analyses from 70,000 station pair combinations demonstrates that irradiance correlation coefficients between two sites decreases predictably with distance and that the decrease occurs more slowly with longer time intervals (Hoff and Perez, 2011). This provides statistical evidence that variability of solar power predictions between plants becomes increasing unrelated with increased distance and

thus aggregation reduces the overall variability of multiple PV systems’ output.

Several studies demonstrate that error metrics for single point irradiance forecasts are much higher than an average of forecasts for multiple points, commonly referred to as an area forecast. For example, Lorenz et al. (2009) demonstrate that the day-ahead irradiance forecasts for single stations in Germany show a RMSE of 36%, where as an area forecast for the entire country results in a RMSE of 13%. In addition, Pelland et al. (2011) found that the RMSE for the forecast of the average irradiance of 10 ground stations across Canada and the U.S. was approximately

(23)

15

67% lower than the RMSEs of the irradiance at each of the individual ground stations.

The research to date on solar plant variability demonstrates that smoothing (reduced variability) occurs both within PV plants and for an aggregate of solar power plants within a larger geographic region. Furthermore, forecast error for one individual irradiance sensor will be higher than that for an individual solar energy plant occupying some defined area depending on system size. Forecast error will likewise be reduced for an aggregation of several geographically dispersed solar plants. The relationship between PV power output variability and forecast error is through the impacts that clouds have on ground level insolation. Clouds lead to PV power output variability and similarly make forecasting irradiance more challenging. Improved characterization of these dynamics will help grid operators anticipate and better manage the variable nature of solar energy production.

FIGure 10: QuAntIFyInG SolAr VArIAbIlIty

(Source Hoff and Perez, 2011)

IRRADIANCE

6:00 12:00 18:00

IRRADIANCE (W/M

^2)

PACIFIC STANDARD TIME

1000 750 500 250 0 1 LOCATION 25 LOCATIONS CLEAR SKY

10 SEC. DATA FROM 400M X 400M GRID AT CORDELIA JUNCTION, CA (NOV. 7, 2010)

10-SECOND CHANGE IN IRRADIANCE

6:00 12:00 18:00

IRRADIANCE CHANGE (W/M

^2)

PACIFIC STANDARD TIME

500 250 0 -250 -500 1 LOCATION (α = 31) 25 LOCATIONS (α = 9)

10 SEC. DATA FROM 400M X 400M GRID AT CORDELIA JUNCTION, CA (NOV. 7, 2010)

(24)

16

(25)

17

4. | Using solar Power Production Predictions

for Utility Planning and operations

Table 4 lists the various solar forecast end-users and potential applications of area forecasts and point forecasts. Here we define area forecasts as a composite of two or more geographically dispersed solar plant power predictions, whereas a point solar forecast is a prediction of future output for one specific solar plant. While solar plant predictions will be used in many of the same ways that wind forecasts are being used today, the distributed nature of much of the solar being deployed will result in new end-users and applications. Area forecasts of solar energy power predictions can be produced for entire balancing areas, nodes within wholesale markets, distribution systems, or even feeder-level forecasts within distribution systems depending on end-user requirements.

The end-users and the applications of solar forecasting services will be different for each region depending on the mix of central utility-scale solar versus distributed solar, and whether or not electricity markets are

unregulated or regulated. In regions with competitive wholesale power markets, predicting intra-hour, intra-day, and day-ahead power production from single or multiple solar power plants may be used for reliability planning and efficient market operations. In those regions with vertically integra

Predicting Solar Power Production: