UPDATED FINAL Availability Evaluation Report - … · UPDATED FINAL . Availability Evaluation Report . ... PV. This extensive project experience has ... The following table provides

UPDATED FINAL Availability Evaluation Report

Prepared For: SunPower Corporation

Prepared By:

January 14, 2013

2303 Camino Ramon, Suite 220, San Ramon, CA 94583 +1 (925) 867-3330 Fax +1 (925) 867-3331 – www.bewengineering.com

BEW Engineering Independent Engineering Report Page 2

BEW Engineering - Experience and Qualifications

BEW Engineering, a wholly owned subsidiary of DNV, is a recognized leader in solar photovoltaic engineering services and our team has well over 200 years of combined experience in this area over the last three decades. BEW has merged into DNV KEMA Renewables, Inc. under the banner DNV KEMA Energy & Sustainability. BEW has provided technical services on commercial and utility scale projects totaling more than 10 GW of PV. This extensive project experience has allowed for the development of effective energy analysis, EPC selection, design review, independent engineering, system design, financial analysis, performance prediction and other important tools which provide an efficient approach to the evaluation of these systems. BEW’s experience in technology advancement with module manufacturers, inverter suppliers, balance of system component suppliers, and mounting and tracking system manufacturers provides a unique perspective toward integrating these components within a given project. Experience with national and international codes and standards development and work with the U.S. Department of Energy and National Laboratories keeps BEW at the leading edge of standards and technology developments. Forensic experience has also assisted BEW in improving PV system implementation practices that impact real world systems. This extensive experience allows BEW to provide engineering services in an efficient and cost effective manner.

BEW has a staff of 55, with specialists in the areas of transmission and distribution systems, high grid penetration of renewable energy, curtailment risk assessment, structural and wind-loading calculations for rooftop and ground-mount systems, civil engineering, electrical engineering, codes and standards, PV system design and optimization, solar resource assessments, PV system performance analysis, and power electronics design and verification testing.



Biographies of Key Contributors

Jessica Forbess, M.Eng., Senior Engineer, PV Solar

Ms. Forbess performs independent engineering reviews on behalf of clients such as third-party developers, integrators, and investment banks. She has been involved in the photovoltaic solar industry since 2008, with an emphasis on PV performance analysis, utility-scale PV system operations and maintenance, PV system design, energy management and data acquisition systems.

Prior to joining BEW, Ms. Forbess was a Performance Engineer with the EPC division of First Solar, Inc., a leading solar module manufacturer, project developer and EPC, providing performance testing, analysis and monitoring for the largest PV plants in North America and the US in 2009 and 2010. She began her career in telecommunications as a field and test engineer rolling out flexible software-based cell network infrastructure. She completed her BSEE from the Massachusetts Institute of Technology in 1997, and her M.Eng. in EECS also from MIT in 2000.

Jeffrey D. Newmiller, Principal Engineer, PV Solar

Mr. Newmiller specializes in a number of areas of PV with his prime focus on analyzing performance data, developing performance monitoring systems, troubleshooting PV systems, and reviewing system designs for photovoltaic solar power systems. His engineering know how has been applied to various PV systems from 1 to 100,000kW. Mr. Newmiller has worked with PV since 1992 and has participated in the development of IEEE1547, UL1741, California Rule 21, and updates for the National Electric Code Articles 690 and 705. He has provided technical support to the California Energy Commission by managing the CEC Eligible Equipment Lists of Inverters and PV Modules for the first five years of implementation, and provided reviews of compliance with equipment certification requirements for California Rule 21. Mr. Newmiller has analyzed and suggested improvements on the modeling capabilities of several clients’ in-house PV performance modeling tools. Mr. Newmiller received BS and MS degrees in Mechanical Engineering (with focus on engineering measurement theory and practice) from the University of California at Davis.



Table of Contents

1. Executive Summary ......................................................................................................................... 5

2. Introduction .................................................................................................................................... 7

3. Overview ......................................................................................................................................... 9

Variability .......................................................................................................................................... 10

4. Review of SunPower Availability Algorithm .................................................................................... 12

5. Review SunPower Historical Availability ......................................................................................... 15

Time-weighted and Energy-weighted Availability ............................................................................... 15

Availability by Subset ......................................................................................................................... 16

Variability .......................................................................................................................................... 18



1. Executive Summary

SunPower has engaged BEW to conduct a review of a subset of SunPower’s utility-scale PV plant operational data with the intent of quantifying expected availability for staffed SunPower solar plants. SunPower asserts that the availability of SunPower plants with full O&M service from SunPower on-site staff will be at least 99%, and has provided supporting data and calculations. From a financial point of view, 99% availability means that the project revenue may be reduced by no more than 1% annually due to plant outages in a p(50) estimate. BEW has reviewed the data, discussed the data validation process with SunPower, and replicated the calculations. BEW has performed additional calculations to statistically compare the availability when external forced outages are included as energy lost, and quantified the variability of annual availability. BEW has confirmed SunPower’s assertion on a p(50) basis, and has further analysis on the annual and ten-year average variability. This is based on the assumption that power plant service and maintenance activities are consistently performed throughout the life of the project.

BEW confirms SunPower’s assertion of 99% availability with the data in Table 1 below. It indicates that a single year p(50) energy estimate can reasonably use 99.5% availability. The data supporting the ten-year average availability is less complete. BEW has calculated the historic ten year average availability assuming O&M procedures and manufacturer support remain constant, Years 2 through 8 are reasonably characterized in the current data, and failure rates in Years 9 and 10 are similar to Year 1. BEW underlines that this assumes that failure rates are not significantly greater in Years 9 and 10 than in Year 1. Our estimate results in a ten-year average annual availability of 99.7% at a p(50) level, and 99.3% at p(90) level including all outages. A more detailed analysis including the assumptions we made, and confidence intervals around each cumulative probability level is found in Section 5.

Table 1: Single year and ten-year cumulative probability estimates

Single Year Availability 10 Year Average Availability

Cumulative Probability

Penalized for all outages

External outages removed



P50 99.5% 99.7% 99.7% 99.8%

P75 98.8% 99.2% 99.5% 99.7%

P90 97.2% 98.3% 99.3% 99.5%

P95 95.2% 97.3% 99.2% 99.5%



The p(50) in all cases shown in Table 1 above is significantly higher than the targeted 99% availability. Because the single year p(90) reduction in energy is similar to that typically seen in solar resource variability assessment, BEW recommends that it be included in a project finance downside uncertainty analysis. The ten-year average annual availability estimate has a higher p(90) value of 99.3%.



2. Introduction

SunPower has engaged BEW to conduct a review of a subset of SunPower’s utility-scale PV plant operational data with the intent of quantifying expected availability for staffed SunPower solar plants. SunPower asserts that the availability of SunPower plants with full O&M service from SunPower on-site staff will be at least 99%, and has provided supporting data and calculations. From a financial point of view, 99% availability means that the project revenue may be reduced by no more than 1% annually due to plant outages in a p(50) estimate. BEW has reviewed the data, discussed the data validation process with SunPower, and replicated the calculations. BEW has performed additional calculations to statistically compare the availability when external forced outages are included as energy lost, and quantified the variability of annual availability. BEW has confirmed SunPower’s assertion on a p(50) basis, and has further analysis on the annual and ten-year average variability. This is based on the assumption that power plant service and maintenance activities are consistently performed throughout the life of the project.

This report analysis covers data from sixteen solar plants in six countries, and includes all utility-scale PV plants monitored by SunPower and operated by on-site staff. The following table provides a detailed list of the solar projects, which ranged in size from 6.1 MWDC to 45 MWDC. The earliest Commercial Operation Date was January 1, 2005, and the most recent plant began operating commercially in January 2012. The data used in the availability analysis was recorded from January 1, 2008 to June 6, 2012. Data from plants becomes available at different times, with the oldest data being from Nellis in January 2008, data from Bavaria starting in September 2009, and the rest following after.



Table 2: SunPower staffed Utility Scale Projects

Site Project Country Commercial Date of Operation

kWp

M0190 Bavaria Solar 1 / Muehlhausen DE 1/1/2005 6,269 M0311 GE Energy Financial Services / Serpa PT 1/17/2007 10,980 M0407 United States Air Force / Nellis Air Force Base US 10/15/2007 14,056 M0505 360 Corporate / Olivenza ES 11/24/2008 18,000 M0672 Florida Power and Light / FPL DeSoto US 9/15/2009 27,604 M0693 Sunshire S.r.l / Tolentino IT 7/1/2010 5,013 M0695 Cassiopea PV Srl (SunRay Renewable Energy) IT 6/1/2010 24,008 M0696 Florida Power and Light / FPL Space Coast US 9/28/2010 10,000 M0726 Exelon Generation Company, LLC / Cook County US 3/23/2010 10,000 M0815 Montalto Centauro IT 9/18/2010 8,800 M0818 Xcel - Greater Sandhill I US 9/23/2010 20,072 M0827 Andromeda PV Srl (SunRay Renewable Energy) IT 3/31/2011 45,000 M0837 Amherstburg Solar Farm CA 7/1/2011 20,000 M0853 Montalto Andromeda IT 3/31/2011 6,100 M0865 Iberdrola Alamosa US 1/6/2012 35,100 M0869 Iberdrola Copper Crossing US 9/1/2011 23,000

SunPower provided the data in Staffed_Availability_Study_(Energy-Weighted)_2012_07_25.xlsx. Additional information was provided in discussions with SunPower including interactive sessions with their SCADA management system, OSI PI.



3. Overview

At a portfolio level, staffed SunPower plants have achieved at least 99% availability both on a time-weighted basis and an energy-weighted basis when forced grid outages are excluded. When grid outages are included, the portfolio maintains 99% availability on an energy-weighted basis, though not on a time-weighted basis as shown in Figure 1.

Figure 1: Portfolio Availability

While the figure above indicates there is minimal difference between time-weighted availability and energy-weighted availability, a more detailed review by plant shows that there is a difference that exceeds 1% in specific instances. Because the primary goal of tracking availability in most contexts is to quantify and minimize the energy lost through operational failure, BEW recommends using estimated energy to weight the operating time, which more closely represents energy lost. If estimated energy isn’t available, BEW considers weighting by irradiance to be sufficiently accurate for the purposes of improving the calculation of availability over a flat time-weighted calculation.

The energy-weighting process in this report only considers the energy on a daily basis, and therefore does not consider the difference in an outage at the end of the day compared to an outage at noon. The analysis in this report does take into account the difference between an outage on a cloudy winter day and a sunny summer day. BEW expects that daily weighting is sufficient for this analysis, because we expect outages to be either unbiased as to time of day, or slightly biased to the ends of the day rather than the peak. Two very typical inverter failures are failure to start up in the morning, and failures due to high temperatures. Start-up failures are typically resolved by on-site staff before peak irradiance, and high temperature failures are slightly more likely to occur a few hours after peak irradiance.

98.82%

99.19% 99.03%

99.28%

98.0%

98.5%

99.0%

99.5%

100.0%

Penalized for All Energy Lost Not Penalized for Grid Outages

Portfolio Availability

Time Weighted Energy Weighted Target



Variability An individual solar plant may operate at a much lower availability than the portfolio average. It may operate at a higher availability as well, but as availability cannot exceed 100%, the distribution of annual availability is not symmetric as a typical normal distribution would be.

To calculate the variability of availability, BEW used a logarithmic transform of the data, a set of all of the annual availabilities for all plants, with a total of 37 data points. We did not assume that the data was perfectly log-normal, and therefore a kernel estimate was used to curve-fit the probability distribution that best reflects the available information.

BEW confirms SunPower’s assertion of 99% availability with the data in Table 1 below. It indicates that a single year p(50) energy estimate can reasonably use 99.5% availability. The data supporting the ten-year average availability is less complete. BEW has calculated the historic ten year average availability assuming O&M procedures and manufacturer support remain constant, Years 2 through 8 are reasonably characterized in the current data, and failure rates in Years 9 and 10 are similar to Year 1. BEW underlines that this assumes that failure rates are not significantly greater in Years 9 and 10 than in Year 1. Our estimate results in a ten-year average annual availability of 99.7% at a p(50) level, and 99.3% at p(90) level including all outages. A more detailed analysis including the assumptions we made, and confidence intervals around each cumulative probability level is found in Section 4.

Table 3: Single year and ten-year cumulative probability estimates

Single Year Availability 10 Year Average Availability






P50 99.5% 99.7% 99.7% 99.8%

P75 98.8% 99.2% 99.5% 99.7%

P90 97.2% 98.3% 99.3% 99.5%

P95 95.2% 97.3% 99.2% 99.5%

The p(50) in all cases shown in Table 1 above is significantly higher than the targeted 99% availability. Because the single year p(90) reduction in energy is similar to that typically seen in solar resource variability assessment, BEW recommends that it be included in a project finance downside uncertainty



analysis. The ten-year average annual availability estimate has a higher p(90) value of 99.3%. The data in Table 1 is shown graphically in Figure 2.

Figure 2: Estimated single year and ten year availability probability distribution

0%10%20%30%40%50%60%70%80%90%

100%

84% 86% 88% 90% 92% 94% 96% 98% 100%

Cum

ulat

ive

Prob

ailit

y

Availability

Availability Variability

Penalized for all outages (1 year) External outages removed (1 year)

Penalized for all outages (10 year) External outages removed (10 year)



4. Review of SunPower Availability Algorithm

The power industry typically calculates availability on a time-weighted basis, as most power plants run at a fairly consistent capacity factor. Annual availability is calculated by dividing the hours the plant was generating energy by the number of hours in a year. Exceptions are allowed for planned maintenance, force majeure conditions, and other contingencies.

Contracts in the solar industry commonly also use a time-weighted basis to calculate availability. Minimal data from utility-scale solar plants has been available to calculate whether a time-weighted basis would tend to over- or under-estimate the energy lost through outages, since the lost energy at any given point in time is highly variable. Annual availability is calculated on an energy-weighted basis by dividing the generated energy by the total possible generated energy based on the available irradiance at the site. SunPower has provided data to enable BEW to calculate availability for each basis.

SunPower has described their availability algorithm in discussions with BEW, including demonstrating specific examples directly from the SCADA management system, OSI PI. We have not validated the actual software implementation of the availability calculation in the SunPower data collection system.

An Availability data point is calculated automatically in the OSI PI system used to monitor the plants. It is calculated on a per-inverter basis at 15 minute increments when the average POA irradiance is greater than 100 W/m2over the 15 minute interval. If an inverter generates any energy in a 15 minute period, its Availability is marked as 1 for that period. If an inverter generates no energy in a 15 minute period, its Availability is marked as 0 for that period. BEW finds this algorithm to be generally reasonable for utility-scale PV plants, but notes the following limitations to the algorithm:

1. 100 W/m2 is higher than ideal for a threshold of inverter operation. A typical PVSim model of a single axis tracking system showed 0.5% of the annual generation occurred during hours with less than 100 Wh/m2. While it is unlikely that the inverter would be configured such that it would fail to generate under those conditions while operating correctly in all others, it is an upper bound on the potentially missing energy for a typical single axis tracking system. BEW has not performed an in-depth review of the inverter performance at different levels of detection, but verified that operational data from a single inverter showed only a 0.06% lower annual time-weighted availability when the threshold of inverter operation was 50 W/m2 instead of 100 W/m2 and a 0.14% lower annual time-weighted availability when the threshold of inverter operation was 25 W/m2 instead of 100 W/m2. The comparisons would show a lower variation in energy-weighted availability, given the small amount of energy involved at the edges of the day.

2. Inverter outages lasting less than 15 minutes are not detected. This includes inverters that trip off line and recover within the 15 minute period. BEW expects that these outages are monitored through alarms, and that the actual energy lost is minimal, given their short duration.

3. If the POA irradiance does not exceed 100 W/m2, no Availability is calculated. BEW has observed 6,054 out of 326,423 inverter-days in the data where this was the case for the entire day, which



is 1.9% of the total data. The low POA values are likely due to missing or bad data through communication errors, or very cloudy winter days. These days were not included in the analysis.

4. No DC or tracker motor availability is included. This analysis does not consider whether the inverters are operating with reduced capacity. BEW does not regard this lack of data as a deficiency, as it is not typically used as a metric in contractual agreements, and we expect significant DC capacity or tracker failures would be detected through more typical contractual metrics like performance index. We note these issues as metrics that should be tracked in the future for improved O&M cost and availability forecasting. A few causes of reduced capacity are:

a. Open fuses at the DC feeder or string level b. Poor alignment of modules through tracker failures. No information was available on

the percentage of time the modules were aligned appropriately. The Availability data was averaged over each day to create a Daily Availability between 0 and 1 for each inverter on each day, and exported. OSI PI also calculates the Daily Expected kWh for each inverter using environmental data. The Expected kWh was exported to calculate an Energy-Weighted Availability along with a Time-Weighted Availability. The rest of the analysis to quantify expected availability of a staffed SunPower solar plant was performed outside of OSI PI.

The Time-Weighted Availability was evaluated by calculating the average of the Daily Availability over different categories, including age of plant, country of plant, and each plant on an annual and lifetime basis.

The Energy-Weighted Availability was evaluated by calculating the Lost kWh by subtracting the Daily Availability from 1 and multiplying by the Daily Expected kWh. Then Energy-Weighted Availability was evaluated over the same set of categories by dividing the total Lost kWh by the total Daily Expected kWh and subtracting from 1 for each category. This calculation is expressed in the following set of equations.

𝐴𝑣𝑎𝑖𝑙𝐸 = 1− ∑𝐿𝑜𝑠𝑡 𝑘𝑊ℎ𝐷𝑎𝑖𝑙𝑦

∑𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑘𝑊ℎ𝐷𝑎𝑖𝑙𝑦

𝐴𝑣𝑎𝑖𝑙𝐸 = 1− ∑ [�1− 𝐴𝑣𝑎𝑖𝑙𝐷𝑎𝑖𝑙𝑦� × 𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑘𝑊ℎ𝐷𝑎𝑖𝑙𝑦]

∑𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑘𝑊ℎ𝐷𝑎𝑖𝑙𝑦

In addition, specific outages were excluded in the analysis outside of OSI PI. The exclusions fell into two general categories, Forced and Communication Error. BEW has reviewed the causes identified by SunPower and in general agrees that they are reasonable and typical exclusions in the industry.

It is typical to remove outages requested or caused by the utility or the grid operator from Availability calculations for O&M contracts as well as force majeure events. However, BEW recommends reviewing Availability from an energy generation estimation stand-point, no matter the cause, as when energy is not generated it may affect the project revenue stream. Power purchase agreements and interconnection agreements may include compensation for lost revenue due to outages caused by the utility, but the circumstances that are covered vary between agreements.



BEW believes that most of the grid outages in the excluded data are not typical annual events, and might be considered more likely in the first years or in a subset of 20 years on a given distribution or transmission line. Because only three plants are more than four years old, BEW is not able to quantify the likelihood of grid outages in later years of plant operation versus early years.

Some of the Communication Errors were due to no communication, and SunPower has demonstrated that they verified that the inverters were generating energy though the inverters were not reporting directly to OSI PI. Other Communication Errors were due to erroneous readings of irradiance at night. BEW has no concerns about excluding Communication Error events from calculation.

BEW has performed the availability analysis for two different sets of exclusions:

1) The best case, where all outages identified by SunPower were removed from the data set 2) The worst case, where only the Communication Errors were removed, and therefore SunPower

was penalized for all energy lost



5. Review SunPower Historical Availability

Time-weighted and Energy-weighted Availability BEW reviewed both the time-weighted and energy-weighted availability for the overall portfolio, with and without the exclusions. The time-weighted availability was similar to the energy-weighted availability at a portfolio level.

Figure 3: Comparison of time-weighted and energy-weighted availability over the portfolio

However, when the difference in the two techniques was examined on a per plant level, larger differences are found. Because the primary goal of tracking availability is to quantify and minimize the energy lost through operational failure, BEW recommends using energy-weighted availability. Discussion of availability below focuses on the energy-weighted availability.

98.82%

99.19% 99.03%

99.28%

98.0%

98.5%

99.0%

99.5%

100.0%

Penalized for All Energy Lost Not Penalized for Grid Outages

Portfolio Availability

Time Weighted Energy Weighted Target



Figure 4: Comparison of time-weighted and energy-weighted availability at the plant level

Availability by Subset BEW has reviewed the availability data over a number of different categories. BEW has plotted the energy-weighted availability for the worst case availability, which includes energy lost from all external outages, and for the best case availability, where days with externally caused outages were removed.

In Figure 5 below, the availability of each plant was calculated for the period from September 29, 2009 onward. The majority of the plants achieved greater than 99% availability even when external outages were included in the calculation. Only four plants did not achieve 99% availability with external outages included. Only one plant did not achieve 98% availability with external outages included, and it achieved 98% availability when the external outages were excluded.

97.0%

97.5%

98.0%

98.5%

99.0%

99.5%

100.0%

Availability by Site

Time-Weighted Energy-Weighted Target



Figure 5: Energy-weighted availability by plant

In Figure 6 below, the availability of the plants in each country was calculated for the period from September 29, 2009 onward. Only one country did not achieve 99% availability with external outages included. The single poor performing plant in Figure 5 is the only plant in Portugal, so the poor performance should not necessarily be regarded as a general problem with availability in Portugal.

Figure 6: Energy-weighted availability by country

90%

92%

94%

96%

98%

100%

Availability by Plant

w/ External Outages w/o External Outages Target

90%

92%

94%

96%

98%

100%

CA DE ES IT PT US

Availability by Country




In Figure 7 below, the availability of the plants at each year of their life was calculated for the period from September 29, 2009 onward. The majority of the years achieved greater than 99% availability even when external outages were included in the calculation. Year 5 did not achieve 99% availability with external outages included because the single poor performing plant in Figure 5 was one of only three plants with data at year 5.

Figure 7: Energy-weighted availability by plant age

Variability As is evident in the figures above, an individual solar plant may operate at a much lower availability than the portfolio average. It may operate at a higher availability as well, but as availability may not exceed 100%, the distribution of annual availability is not symmetric as a typical normal distribution would be.

For the variability analysis, it is assumed that the service and maintenance activities are consistent throughout the period of deployment of the PV power plant. This is a critical assumption and if the level of maintenance were to be reduced then the availability would be expected to drop and the application of this analysis would be inappropriate.

To calculate the variability of availability, BEW used a logit1 transform of the data, a set of all of the annual availabilities for all plants, with a total of 37 data points. Data where availability was recorded as

1 Logit is defined as log of odds ratio, or logit(p) = ln[p/(1-p)], with corresponding inverse expit(q) = eq/(1+eq).

90%

92%

94%

96%

98%

100%

1 2 3 4 5 6 7 8

Availability by Plant Age




100% was re-coded as 99.999% (a value larger than all non-100% values) to permit use of the transform. This invalidates probability analysis at the high extremes of availability, but provides appropriate weighting to support analysis at low levels of availability. We did not assume that the data was perfectly logit-normal, and therefore a kernel estimate was used to curve-fit the probability distribution that best reflects the available information. Because of the logit transform, the cumulative probability graph shown in Figure 8 cannot be translated directly to read that p(50) is approximately 95% of the mean availability. Rather, the p(50) is approximately 95% of the mean of the logit transform of availability. For example, to find p(90) or the availability for which 90% of cases will be greater, the intersection of 100% - 90% = 10% cumulative probability with the red kernel estimate curve is 66.6% of the mean of the logit of the data (5.760), so

𝑝(90) = 𝑒0.617× 5.760

1 + 𝑒0.617 × 5.760 = 97.2%

Figure 8 and Figure 9 show the translated probabilities, with corresponding means of logit availability, 5.760 and 6.082.

Figure 8: Variability of availability including all outages

It is good statistical practice in any industry to only use a normal distribution fit when the data can be shown to meet the criteria of a normal distribution. A normal distribution fit is plotted with the kernel estimation fit in Figure 8 and Figure 9. There is a significant difference in the two fits shown in both figures. In both figures, the kernel estimation fit is more conservative at the p(90) level, but at

0%

20%

40%

60%

80%

100%

0% 50% 100% 150% 200% 250%

Cum

ulat

ive

Prob

abili

ty

Percent of Mean of Logit(Availability)

Penalized for All Outages SunPower annual availability variability for staffed plants

years 2009 Q4 through 2012 Q2

Sample Kernel Estimate Normal



approximately the p(70) level, the normal distribution is the more conservative. The kernel estimation fit is more conservative again starting at approximately the p(10) level. When all external outages are removed, the bottom of the distribution is pulled in closer to the mean, as we would expect. Figure 9 below shows that this tighter distribution skews the distribution further from the normal fit than when all of the outages are included. Therefore it is important to use the kernel estimation fit to calculate the cumulative probability function.

Figure 9: Variability of availability without external grid outages

Table 4 and Table 5 below show the transformed cumulative probabilities. Since a limited number of samples were used to estimate the variability, estimates for P-levels far from the center of the distribution are less well-defined. Confidence intervals obtained from bootstrapping the available data are computed to illustrate that conclusions from this data regarding conservative downside analyses (e.g. P95) are considerably more uncertain than conclusions regarding more likely outcomes (P50). In general, larger sample sizes which provide more information on the shape of the tails of the distribution can reduce the uncertainty of the downside estimates.

0%10%20%30%40%50%60%70%80%90%

100%

0% 50% 100% 150% 200% 250%

Cum

ulat

ive

Prob

abili

ty


All External Outages Removed SunPower annual availability variability for staffed plants

years 2009 Q4 through 2012 Q2 Sample Kernel Estimate Normal



Table 4: Single year availability including all outages


Availability Upper Confidence Interval (p95)

Lower Confidence Interval (p95)

P50 99.5% 99.3% 99.7%

P75 98.8% 98.1% 99.3%

P90 97.2% 94.1% 98.5%

P95 95.2% 88.1% 97.8%

Table 5: Single year availability without external grid outages




P50 99.7% 99.5% 99.8%

P75 99.2% 98.8% 99.5%

P90 98.3% 97.2% 99.0%

P95 97.3% 95.8% 98.6%

BEW expects that a well-maintained solar plant will have a higher 10 year average annual availability than the single year variability analysis shows. However, SunPower does not have enough data collected to calculate the 10 year average annual availability without making some assumptions, as the majority of their utility plants have been operating for less than five years.

In order to calculate the 10 year average annual variability, we have assumed that Years 9 and 10 in a plant’s life will be similar to Year 1, which had the lowest average availability2, as might be expected due to infant mortality of components. Assuming SunPower’s O&M planned maintenance and response times remain at the current levels, or improve, and assuming that inverter component parts are

2 The exception is Year 5, which had too few plants to average to be considered statistically significant as an individual Year.



available as quickly as they have been in Year 1, the time to repair after any failure should be similar to that in Year 1. We also assume that the failure rates in Year 9 and 10 are similar to those in Year 1. We assume that Years 2 through 8 will have a different average failure rate, lower on average, and interchangeable through those years.

Figure 10and Figure 11 show the translated probabilities, with corresponding means of logit availability, 5.760 and 6.082.

Figure 10: Ten-year average annual availability variability including all outages

0%10%20%30%40%50%60%70%80%90%

100%

60% 70% 80% 90% 100% 110% 120% 130% 140%

Cum

ulat

ive

Prob

abili

ty


Penalized for All Outages SunPower 10-year average annual availability variability for staffed plants


Kernel Estimate Normal



Figure 11: Ten-year average annual availability variability without external outages

Table 6 and Table 7 below show the transformed cumulative probabilities for a ten year average annual availability. The confidence levels reported below don’t capture the uncertainty in the variability calculation, as the uncertainty comes from the assumptions used to characterize the availability in Years 4 through 10, rather than widely spread data and low sample sizes. In general, larger sample sizes which provide more information on the shape of the tails of the distribution can reduce the uncertainty of the downside estimates.

0%10%20%30%40%50%60%70%80%90%

100%

60% 70% 80% 90% 100% 110% 120% 130% 140%

Cum

ulat

ive

Prob

abili

ty

Percent of Mean of Logit(Availablity)

All External Outages Removed SunPower 10-year average annual availability variability for staffed plants


Kernel Estimate Normal



Table 6: Ten-year availability including all outages




P50 99.7% 99.7% 99.7%

P75 99.5% 99.5% 99.5%

P90 99.3% 99.3% 99.4%

P95 99.2% 99.1% 99.2%

Table 7: Ten-year availability without external outages




P50 99.8% 99.7% 99.8%

P75 99.7% 99.6% 99.7%

P90 99.5% 99.5% 99.6%

P95 99.5% 99.4% 99.5%

Documents

UPDATED FINAL Availability Evaluation Report - … · UPDATED FINAL . Availability Evaluation Report . ... PV. This extensive project experience has ... The following table provides