Exploring the Reliability of U.S. Electric Utilities. Laboratory

Transcription

1 Exploring the Reliability of U.S. Electric Utilities Peter Larsen 1,2, James Sweeney 1, Kristina Hamachi-LaCommare 2, and Joseph Eto 2 1 Management Science & Engineering Department, Stanford University 2 Energy Analysis and Environmental Impacts Department, Lawrence Berkeley National Laboratory Abstract This paper is a follow up to research conducted by Eto et al. (212) that explored the long-term reliability of the U.S. power system. This research builds on the original analysis by including (1) a more comprehensive panel dataset of U.S. electric utilities and (2) a more robust econometric analysis of the factors which may affect reliability. First, reliability events for a representative sample of U.S. electric utilities are identified and categorized by reported cause. Next, an econometric model is introduced that relates two different measures of electric reliability to a number of correlates including annual measures of weather variability, utility sales, transmission and distribution expenditures, and the presence of outage management systems. A fixed and random effects panel regression was conducted using a cross-sectional dataset containing 12 years of data for 15+ electric utilities from across the United States. It is shown that a 5% increase in average wind speed increased the frequency and duration of reliability events by ~15% and ~5%, respectively. Higher than average wind speeds and lightning are consistently correlated with increases in the duration and frequency of reliability problems regardless of whether major events are included in the calculation of the performance metric. There also appears to be a statistically significant and increasing time trend in the frequency and duration of reliability events, but results are sensitive to whether or not major events (i.e., natural disasters) are included in the panel regression. The paper concludes with a short discussion of study limitations and policy considerations, which could improve future technical analyses and, ultimately, the reliability of the U.S. electric utility sector. Lead Author Contact Information Larsen: Lawrence Berkeley National Laboratory, One Cyclotron Road, MS9-4, Berkeley CA USA 9472 / Phone: (51) / PHLarsen@lbl.gov

2 2 I. Introduction a. Study context Several studies have quantified the annual cost of U.S. power outages with estimates ranging from $28 billion to $29 billion (LaCommare and Eto 25; Swaminathan and Sen 1998; Primen 21) There are a number of factors that could affect the long-term reliability of U.S. electric utilities including, but not limited to: abnormal weather, presence of wildlife, transmission and distribution maintenance and capital expenditures, electricity sales, and the installation of outage management systems. Despite the high costs attributed to power outages, there has been little or no research conducted to determine which factors are statistically correlated with increasing frequency and duration of outages across the United States. Eto et al. (212) first developed a basic panel dataset and preliminary econometric framework to evaluate some factors that may be correlated with more frequent and lengthy service interruptions. It was shown that (1) the frequency and duration of reliability events increased ~2% annually; (2) increases in cooling degree-days (i.e., hot weather) are correlated with increased frequency of outages; and (3) outage management systems are initially correlated with longer outages, but electric utilities appear to be learning from these systems over time. The Eto et al. (212) paper also acknowledged that additional factors should be evaluated in future studies including a larger sample of utilities and more disaggregate measures of weather variability (e.g., lightning strikes and severe storms), utility characteristics (e.g., the number of rural versus urban customers, and the extent to which transmission and distribution lines are overhead versus underground), and utility spending on transmission and distribution maintenance and upgrades, including advanced technologies. b. Research questions The overall purpose of this study is to expand on research by Eto et al. (212) by systematically evaluating the factors that affect the long-term reliability of U.S. electric utilities. Furthermore, this analysis attempts to answer the following questions: Are there utility-specific differences in reported electricity reliability? Is the share of underground distribution miles to total line miles correlated with reliability? Do reliability improvements occur after the installation of an automated outage management system (OMS)? Is the total amount of electricity delivered correlated with changes in utility reliability? How does abnormal weather affect the frequency and duration of system outages? Are previous year annual transmission and distribution expenditures correlated with the frequency and duration of system outages? Are there unexplained trends in reported electricity reliability over time?

3 3 This paper is organized as follows. Section II provides background on what is known about the factors that affect the reliability of electric utilities as well as how electric utility reliability is typically measured. The empirical method, data sources, and econometric techniques are described in Section III. Section IV presents results and section V contains a short policy discussion and conclusion. II. Background This section provides background on what is known about the factors that affect the reliability of electric utilities as well as how reliability is typically measured. a. Factors that affect the reliability of bulk power systems Not surprisingly, there are a number of causes associated with increased frequency and duration of outages. This section reviews causes of reliability events as reported by a subset of the U.S. electric utilities evaluated in the broader econometric analysis. The following utility reliability reports were consulted to determine the causes of past reliability events: Florida Public Utilities Company (FPUC 213); Rocky Mountain Power (RMP 211); Interstate Power and Light Company (IPL 213); Jersey Central Power & Light (JCP&L 213); Madison Gas and Electric Company (MGE 213); Pacific Gas & Electric Company; Portland General Electric (PGE 212); PSEG Services Corporation (PSE&G 213); and AEP Southwestern (SWEPCO 212). Table 1 provides information on the range of categories used by a selected number of utilities introduced above. Causes of reliability events are not consistently reported across utilities Electric utilities report the causes of reliability events with varying levels of detail (see Table 1). For example, Portland General Electric (PGE) reports causes by individual feeder using the following 11 general categories: equipment; lightning; loss of supply (substation); loss of supply (transmission); other; planned; public; unknown; vegetation; weather; and wildlife. In addition, PGE provides more granular information on the individual causes within each general category (e.g., the types of components that failed within the PGE equipment category). Other utilities report the causes in a less-granular format. AEP Southwestern, for example, identifies seven general causes: animals and birds; people; unknown; utility-owned equipment; other; vegetation; weather (including lightning) 1. Pacific Gas and Electric (PG&E), which was not included in Table 1, did not report the cause of their outage events using general categories. Instead, PG&E qualitatively describes the ten largest reliability events for each year over the past decade. Interestingly, the ten largest outage 1 Although AEP Southwestern does not break out the frequency or duration of outages causes by lightning strikes, other utilities, including companies not represented in Table 1, do report the effect of lightning strikes on reliability.

4 4 events in 21 were directly related to adverse weather. In January, a major storm produced wind speeds in excess of 5 mph for three straight days (Jan 18-2), heavy rainfall and lightning for four days (Jan 18-21), and heavy snowfall in the Sierra Nevada mountains (Jan 2-21). This storm affected nearly 1.2 million PG&E customers and ~4, employees were dispatched to restore service. The New York Department of Public Service s electric reliability performance report (NYDPS 213) indicated most of a utility s interruptions are a result of the following categories: major storms, tree contacts, equipment failures, and accidents. It was noted that 212 was by far the worst year ever for storm effects in the 24 years of staff recordkeeping, taking that distinction from last year (NYDPS 213). Hurricane Sandy caused most, but not all storm-related system outages during 212. ~17 million hours of customer interruptions (or nearly a quarter of the total number hourly interruptions since 1989) were attributed to Hurricane Sandy (NYDPS 213). It was reported that over two million customers were affected by this specific storm, but other storms also significantly affected system reliability including a major blizzard in January and severe thunderstorms in the summer.

5 5 Table 1. Causal categories for a selected number of electric utilities Utility name Madison Gas & Electric Company (Wisconsin) Florida Public Utilities Company (Florida) Rocky Mountain Power (Wyoming) Interstate Power & Light (Iowa/Minnesota) Jersey Central Power & Light (New Jersey) Reporting year Metric Causal categories Comments 212 SAIFI Cable failures; equipment failures; stormrelated; Reported by worst performing circuit. substations; tree-related; wildlife- related; other 212 Number of Named storm; animal; vegetation; other; Reported by two geographic divisions within service outages corrosion; unknown; transformer failure; territory. 211 SAIDI (% share); SAIFI (% share) 212 % of outage minutes 212 Number of customer hours PSE&G (New Jersey) 212 Number of customer hours Portland General Electric (Oregon) AEP Southwestern Electric Power (Texas) 212 Frequency of outage; outage duration (hours) 211 % of interruptions lightning; vehicle Weather; animals; environment; equipment; interference; loss of supply; operational; other; planned; trees Earthquake; equipment; error; lightning; major event; overload; public/other; scheduled; supply; trees; unknown; weather; wildlife Animals; equipment-related; lightningrelated; other/unknown; trees (preventable); trees (not preventable); vehicle Trees; construction (underground); construction (overhead); supply and station equipment; other; lightning; outside plant equipment; external; animals; weather Equipment; lightning; loss of supply (substation); loss of supply (transmission); other; planned; public; unknown; vegetation; weather; wildlife Animals and birds; people; unknown; utilityowned equipment; other; vegetation; weather (including lightning) Percentage of outage minutes by cause was reported for Reported by entire service territory, northern region, and central region. Causes were reported from and across four divisions within service territory. Causes were broken down by feeder and with more granularity than the general categories reported in this table.

6 6 Inconsistent definitions of major events Utilities are typically allowed to exclude major events, like severe storms, from the reliability performance calculations, because they are circumstances over which the utilities have limited control (NYDPS 213). In New York state, major events are excluded when an event causes service interruptions of at least 1 percent of customers in an operating area, and/or interruptions with duration of 24 hour or more (NYDPS 213). PSE&G defines a major event as a sustained interruption of electric service resulting from conditions beyond the control of the EDC, which may include, but is not limited to thunderstorms, tornadoes, hurricanes, heat waves or snow and ice storms, which affect at least 1% of the customers in an operating area (PSE&G 211). In contrast, the California Public Utilities Commission (PG&E 211) defines a major event as an event that meet either of the two following criteria: (a) the event is caused by earthquake, fire, or storms of sufficient intensity to give rise to a state of emergency being declared by the government, or (b) any other disaster not in (a) that affects more than 15% of the system facilities or 1% of the utility s customers, whichever is less for each event. The inconsistent exclusion of major events creates challenges in the analysis and interpretation of reliability metrics across the country. For this reason, the Institute of Electrical and Electronics Engineers (IEEE) provides voluntary guidance on how utilities should define and exclude major events. IEEE ( ) defines a major event from a statistical standpoint as a day in which the daily system System Average Interruption Duration Index (SAIDI) exceeds a Major Event Day threshold value 2.activities that occur on Major Event Days should be separately analyzed and reported. Regardless of inconsistencies in the definition of major events, electric utilities report a number of general causes for reliability events including: planned outages, weather, wildlife, T&D equipment failure, human error, vegetation, and other/unknown/external. Figure 1 which is an aggregation of information provided by the utilities listed in Table 1 shows that equipment failure (25%), vegetation (21%), other/unknown/external (2%); and weather (15%) are the top factors which affect the duration of reliability events. Figure 2 shows that equipment failure (25%), vegetation (24%), other/unknown/external (17%); and weather (15%) are the top factors which affect the frequency of reliability events. Interestingly, wildlife was listed as the cause in 11% of the total number outages, but only 4% of duration of the event was attributed to this factor. This finding implies that wildlife (e.g., squirrels, birds) cause reliability events relatively frequently, but wildlife-related causes do not necessarily lead to prolonged outages. 2 See IEEE for more information about the quantitative method used to estimate the Major Event Day threshold value.

7 7 Figure 1. Factors that increase the duration of reliability events (n=5) Figure 2. Factors that increase the frequency of reliability events (n=5) b. How is electric utility reliability typically measured? The IEEE ( ) formally defines a number of metrics to track electric utility reliability. The System Average Interruption Duration Index (SAIDI) and System Average Interruption Frequency Index (SAIFI) are two of the most commonly used metrics to assess electric utility

8 8 reliability (Eto 212). Equation 1 shows that annual SAIDI for a utility is calculated by summing all annual customer interruption durations and dividing this number by the total number of customers served. In this equation, the restoration time for each interruption event in a given year is represented by r jt, the number of customers affected by the event is Customers jt, and the total number of customers served by the utility in a given year is Customers it. For context, the IEEE recently conducted a survey of 16 utilities and found that the median 212 SAIDI value is 236 minutes (or approximately four hours) (IEEE 213). However, it was noted that data may not be directly comparable since data collection and system differences exist; and certain exclusion differences can occur (IEEE 213). In other words, it is likely that some survey respondents may have reported SAIDI values with major events included and some reported SAIDI with major events excluded. Figures 3 and 4 show the middle 5% range of SAIDI values for utilities used in this study without and with major events included, respectively. (1) 8 7 SAIDI (without major events) th Percentile Median (U.S.) 25th Percentile Figure 3. System average interruption duration index over time (all utilities in this study; without major events; minutes)

9 9 8 7 SAIDI (with major events) th Percentile 2 Median (U.S.) 1 25th Percentile Figure 4. System average interruption duration index over time (all utilities in this study; with major events included; minutes) Equation 2 shows that annual SAIFI for a utility is calculated by summing all annual customer interruptions and dividing this number by the total number of customers served. In this equation, the number of customers affected by the event is Customers jt and the total number of customers served by the utility in a given year is Customers it. The aforementioned IEEE survey of 16 utilities found that the median 212 SAIFI value is 1.5 interruption events (IEEE 213), but similar caveats apply to the comparability of these results. Figures 5 and 6 show the middle 5% range of SAIFI values for utilities used in this study without and with major events included, respectively. The pronounced effect of major events (i.e., storms) on the duration and frequency of outages can be seen in Figures 4 and 6, respectively. The figures also show a fairly flat time trend for the reliability data without major events, but a slightly increasing trend for outages with the inclusion of major events. (2)

10 SAIFI (without major events) th Percentile Median (U.S.) 25th Percentile Figure 5. System average interruption frequency index over time (all utilities in this study; without major events) SAIFI (with major events) th Percentile Median (U.S.) 25th Percentile Figure 6. System average interruption frequency index over time (all utilities in this study; with major events) III. Methods and Data

11 11 a. General empirical method Multivariate regression is the most common empirical method used to analyze unbalanced panel data (see Section 4). Multivariate regressions generate quantitative estimates of the correlation between a dependent variable (e.g., SAIDI or SAIFI) and a set of independent or explanatory variables. I formulate the following regression equation in order to analyze the effect of annual deviations in: weather observations, T&D expenditures, and sales; and the presence of outage management systems (OMS) and IEEE reporting standards on the frequency and duration of electric utility reliability events. The general model specification described above follows previous research into multivariate panel regressions (e.g., see Erdogdu 211; Eto et al. 212). In equation (3), annual utility reliability is represented by the dependent variable: Y it. Electric utility and reporting year are represented by subscript i and t, respectively. Subscript d and f are used to differentiate between observed and unobservable variables, respectively and X di and Z fi represent observed and unobservable variables. For example, variables in X may include annual T&D expenditures and variables in Y might include non-observable factors like wildlife population (e.g., number of squirrels) living near electric utility infrastructure. Finally, Ɛ it represents the model error term and t is a variable to capture a time trend. As indicated above, the array of Z fi variables are unobservable. Accordingly, I define a new term, α i, which represents the combined effect of the unobservable variables on the dependent variable, Y it. Equation 4 describes the reduced form empirical model used in this analysis. (3) The presence of the α i component within this model is crucially important (Erdogdu 211), because it assumes that not all of the explanatory variables are captured in the array of X observable variables. If all explanatory variables were captured in the array of observable variables, then the α i term may be eliminated from the model and a pooled ordinary least squares (OLS) regression technique would be appropriate (Erdogdu 211). As indicated earlier, there are a number of explanatory factors, including presence of wildlife and vegetation coverage, which have not been collected by electric utilities with any degree of precision or consistency. For this reason, it is appropriate to leave the α i term in the model and conduct the econometric analysis assuming the presence of unobservable fixed (or random) effects. b. Data (4)

12 12 Data used in this study was collected from a number of sources and transformed to create a panel dataset for the random (fixed) effect regressions. Independent regressors were selected for (1) their potential to proxy the cause of outages (i.e., ideal variables) and (2) whether the data was readily accessible. Table 2 compares a number of ideal ( I ) and readily available ( R ) regressors against the causal categories described in Section II. Unfortunately, a number of ideal regressors (e.g., population of squirrels near overhead T&D corridors, count and duration of customer-caused interruptions) were not readily available. However, there were also a number of regressors that were both ideal and readily available (e.g., number of lightning strikes). Table 3 and the text that follows contains a high-level description of the key data sources and coverage for the dependent variables and the independent variables that were readily available. Electric utility reliability metrics Following Eto (212), utility-specific reliability data (i.e., SAIDI, SAIFI) were initially collected from state public utility commissions (PUCs). PUCs often require investor- and publicly-owned electric utilities to file their reliability reports annually and make this information available to the public. LBNL (213) supplemented the PUC-mandated reliability metrics with annual reliability data collected through web-based research and, in some cases, direct contact with utility representatives. The reliability data collected in this study builds upon Eto (212) by (1) incorporating additional years of reliability data; and (2) increasing the number and geographic coverage of utilities. Eto (212) evaluated reliability from 2 to 29 using information from 155 U.S. electric utilities. For comparison, this study evaluates reliability from 2 to 212 using information from 212 utilities. A logarithmic transformation was applied to the reliability metrics, because (1) this type of transformation makes the interpretation easier (i.e., a one unit change in a regressor leads to an X% change in SAIDI or SAIFI); and (2) annual SAIDI (SAIFI) values are lognormally distributed (see Eto et al. 212). Presence of outage management systems and IEEE reporting standards In addition to increasing the coverage of the core reliability metrics, SAIDI and SAIFI, the presence of outage management systems (OMS) and IEEE reliability data reporting standard was denoted in the data set. If a utility installed or upgraded their outage management system, the year was indicated. Next, a binary dummy variable was created for each year and each utility indicating whether the OMS was present or not. In addition, a continuous variable was created to indicate how many years had passed since the installation of the OMS system. For example, if the utility installed the OMS in 25 and had reliability metrics in 21, then the aforementioned variable in 21 would be coded with a five indicating that five years had passed since the OMS was installed. This variable was created to evaluate the possibility that utilities could learn to use their OMS more effectively over time resulting in improved reliability. If the utility adopted the IEEE standard for reporting reliability metrics, a binary variable was created indicating whether or not the utility used the standard.

13 13 Table 2. Comparison of causal categories to potential regressors Possible Regressors Average Weather Extreme Weather Reported Causes Wildlife Vegetation Human Error T&D Equip. Other/ Unknown/ External Planned Outage Heating degree-days I,R I,R Cooling degree-days I,R I,R Windspeed I,R I,R Total precipitation I,R I,R Lightning strikes I,R I,R # of days with snowfall I I # of days with rainfall I I # of days with sleet/ice I I # of days with high wind I I % share of lines buried R R R R Population of squirrels near overhead T&D corridors I Population of birds near overhead T&D corridors I Average distance from leaf-tip to T&D line I Annual vegetation maintenance budget I,R Number of professional degrees/certifications per employee I Presence of outage management systems I?,R? I?,R? Employee morale I Count and duration of customer-caused interruptions I Number of customers per line mile R Current and lagged T&D expenditures I,R Remaining lifespan of T&D equipment I Count and duration of unknown/other/external interruptions I Sales and electricity delivered per customer I,R Count and duration of planned outages I

14 14 Table 3. Key data sources and coverage (with major events included 3 ) Data Number of Years Original units Source utilities available Reliability (frequency of 147 Varies by SAIFI LBNL (213) outage with major events) utility Reliability (duration of outage 149 Varies by SAIDI LBNL (213) with major events) utility Delivered electricity, revenue, MWh, nominal EIA/Ventyx (214) and number of customers dollars, number of customers Weather (degree-days) HDDs, CDDs NCDC/Ventyx (213) Weather (lightning strikes) Number of NLDN/LBNL (213) positively and negativelycharged strikes Weather (total precipitation) Inches NCDC/Ventyx (214) Weather (average windspeed) Miles per hour NCDC/Ventyx (214) Presence of outage Binary and LBNL (213) management system installation year Adoption of IEEE Binary LBNL (213) reporting standard Transmission and distribution extent Line miles above ground, underground, FERC/EIA/RUS/Ventyx (214) Transmission and distribution expenses and other Nominal dollars of fixed and variable costs FERC/EIA/RUS/Ventyx (214) Electricity delivered, electricity sales revenue, and number of customers Utility revenue from electricity sales, electricity delivered expressed in MWh and number of customers from were collected from the U.S. Energy Information Administration (EIA 214). Utility revenue was originally reported in nominal dollars. For the purpose of this analysis, nominal utility revenue was transformed into real 212 dollars using the Bureau of Labor Statistics Consumer Price Index for electricity expenditures (BLS 214). Sales and electricity delivered were normalized by number of customers, as shown in equations (5) and (6), respectively. (5) (6) 3 Data coverage is similar for SAIDI and SAIFI metrics without major events included.

15 15 Degree-days, lightning strikes, precipitation, and windspeed Information on annual heating degree-days (HDDs) and cooling degree-days (CDDs) was collected from the National Oceanic and Atmospheric Administration s National Climatic Data Center and mapped to each utility s service territory by Ventyx (NCDC/Ventyx 213). Heating (cooling) degree-days are measures directly related to temperature. Heating degree-days (HDD), a measure of cooler temperatures, are calculated by subtracting the daily average temperature from 65 degrees Fahrenheit and summing for the entire year. For example, a daily average temperature of 55 degrees Fahrenheit represents 1 heating degrees for that day. Cooling degreedays (CDD), a measure of high temperature, is calculated by subtracting 65 degrees Fahrenheit from the average daily temperature and summing for the entire year. In other words, a daily average temperature of 85 degrees would represent 2 cooling degrees (85-65) (Larsen 26). Annual HDDs and CDDs were collected for every utility in this study for the years Deviations in annual HDDs and CDDs were calculated by subtracting a utility s HDDs (CDDs) in a given year from the utility s historical average for It is hypothesized that warmer and cooler than average years will be correlated with measurable reductions in utility reliability. Accordingly, a pair of abnormally cold (hot) temperature deviation variables were created to test this hypothesis (see equations 7 and 8 below). If the cooling (heating) degreedays in a given year were less than the 12 year average, the positive deviation variable was coded with a zero. (7) Annual number of lightning strikes was collected from the National Lightning Detection Network (NLDN) (NLDN/LBNL 213). The NLDN characterizes a strike as negatively- or positively-charged depending on the direction of the strike. Negatively-charged strikes occur as a cloud-to-ground strike and are more frequent. For the purposes of this study, total strikes per year were used regardless if they were negatively or positively-charged. The latitude and longitude position for each strike was mapped to each utility s service territory and aggregated for each year from 2 to 212. The resulting dataset contain total number of annual strikes for each utility from Year-to-year deviations in lightning strikes were calculated by subtracting a utility s total number of lightning strikes in a given year from the average annual number of strikes for Like the temperature variables discussed above, it is assumed that more lightning strikes than an average year will be correlated with measurable reductions in utility reliability. Accordingly, a positive deviation lightning strike variable was created to test this hypothesis (see equation 9). If the lightning strikes in a given year were less than the 12 year average, the positive deviation variable was coded with a zero. (8)

16 16 NCDC/Ventyx (214) summed daily precipitation for each station within a utility s service territory. Next, total precipitation for all stations with each utility s service territory was divided by the total number of weather stations. To estimate precipitation at the annual level, NCDC/Ventyx summed up daily precipitation in a given year and then finally calculated an average across all stations to estimate an annual total precipitation value per utility. Deviations in annual precipitation were calculated by subtracting a utility s precipitation in a given year from the average annual precipitation for The hypothesis is that wetter (or drier) years than average will be correlated with measurable reductions in utility reliability. Accordingly, a pair of abnormally wet (dry) precipitation deviation variables were created to test this hypothesis (see equations 1 and 11 below). If the total precipitation in a given year was less than the 12 year average, the positive deviation variable was coded with a zero. Conversely, if the total precipitation in a given year was greater than the 12 year average, the negative deviation variable was coded with a zero. (9) (1) For wind speed, NCDC/Ventyx (214) averaged wind speeds (in mph) across all of the weather stations assigned to a utility. Like the other weather variables discussed above, it my hypothesis that higher wind speeds than an average year will be correlated with measurable reductions in utility reliability. Accordingly, a positive deviation windspeed variable was created to test this hypothesis (see equation 12). If the average windspeed in a given year was less than the 12 year average, the positive deviation variable was coded with a zero. (11) Non-linearity assumption for all weather measures except for lightning It is likely that the relationship between abnormal weather, including temperature, precipitation, and wind and changes in system reliability, are not linear. Hitz and Smith (24) surveyed the literature on the shape of climate change-related damage curves concluding that these curves were nonlinear. Larsen et al. (28) argued that using non-linear indicators would be a more (12)

17 17 appropriate choice for estimating damage to infrastructure. I hypothesize that a 1% increase in average annual wind speed will impact reliability differently than a 5% increase in average annual wind speed. However, it was assumed that increases in lightning strikes had a linear relationship with reliability (i.e., lightning strikes affect reliability in a similar fashion regardless of whether there was a 1% increase in the number of strikes or a 5% increase). Transmission and distribution (T&D) line miles Annual reported mileage of each utility s transmission and distribution system from 2 to 212 was originally collected from the Federal Energy Regulatory Commission (FERC) and Rural Utilities Service of the U.S. Department of Agriculture and processed by Ventyx (FERC/RUS/EIA/Ventyx 214). Where available, T&D line characteristics were reported as overhead miles, underground miles, and other miles. The number of customers was divided by total T&D line miles to create the number of customers per line mile (see equation 13). For this analysis, the share of line miles buried underground was calculated by dividing the total mileage of the T&D system from the total number of miles that were reported as being installed underground (see equation 14). (13) Annual transmission and distribution expenses Annual fixed and variable expenses for each utility s transmission and distribution system from 1999 to 212 were collected from a number of sources including FERC for independentlyowned utilities; the EIA for publicly-owned utilities; and the Rural Utilities Service for generation, transmission, and distribution cooperatives (FERC/RUS/EIA/Ventyx 214). Fixed and variable expenses were originally reported in nominal dollars. For this analysis, each utility s nominal fixed and variable expenses were transformed into real 212 dollars using the appropriate regional Handy-Whitman index of T&D construction costs (213) (see equations 15 and 16). In addition, fixed and variable costs were lagged by one year to reflect the idea that expenditures would not have an effect on reliability measurements until the following year. Finally, annual fixed and variable T&D expenses were combined into total annual transmission and distribution expenses and normalized by number of customers (see equation 17). (14) (15) (16)

18 18 As originally discussed in Eto et al. (212), this type of panel data is considered short (Cameron and Trivendi 29) because the data structure has 15-2 utilities, but a relatively few number of time periods (i.e., 13). In addition, the dataset is also unbalanced (Wooldridge 22) because there are missing reliability metrics and other regressors for a number of utility-year combinations. In summary, the analysis that follows is based on a short, unbalanced panel data set. c. Specific econometric method The completeness and structure of the dataset influenced the choice of tests used to determine the appropriate statistical models to carry out the parameter estimation (Eto et al. 212). A sequence of statistical testing was carried out to determine the specific econometric method for each regression. An initial test was conducted to detect for the presence of general cross-sectional effects. If cross-sectional effects were detected, additional testing was conducted to determine whether a fixed or random effects model specification was more appropriate. Testing for the presence of no cross-sectional effects First, a simple F-test was conducted to test for the presence of general cross-sectional effects. In this case, the F-test examines whether, under the null hypothesis, there are no utility-level effects present in the data. If the null hypothesis is not rejected, a pooled regression is more appropriate. If the null hypothesis is rejected, additional testing is required to determine whether a random or fixed effects model specification will result in smaller standard errors during the estimation process. Testing for the presence of random effects As described earlier, the SAIDI and SAIFI reliability data collected consists of up to 12 years of annual information for 15+ electric utilities. Eto et al. (212) first describe the method used to evaluate this type of information: Fixed and random effects models are particularly useful for this type of analysis because they enable the regressions to explicitly account for differences in the outcomes (i.e., SAIDI and SAIFI) that are correlated with differences in the sources of the data for these outcomes (i.e., the utilities).utilities follow different practices in reporting reliability (e.g., whether or not they use IEEE Standard ). Fixed and random effects models can explicitly account for these correlations and thereby remove the influence of these differences from the other correlations under consideration (Eto et al. 212). If general cross-sectional effects were detected in the previous step, a Hausman (1978) test was undertaken in order to determine whether a fixed or random effects is the appropriate model choice. The Hausman test examines whether, under the null hypothesis, the individual utility effects are uncorrelated with the other regressors in the model (Hausman 1978; Greene 2; Eto et al. 212). If the null is not rejected, both effects models are consistent, but only the random effects model is efficient. In this case, the fixed and random effects models will have the same expected values, but the random effects model will have relatively smaller standard errors. (17)

19 19 Greene (2) notes that using a fixed effects model when the random effects model is consistent may lead to an erroneous interpretation of the statistical significance of coefficients. IV. Empirical Analysis and Discussion This section contains a discussion of the (1) summary statistics for the key variables used in this analysis; (2) results testing for cross-sectional and random effects; and (3) results of the panel regressions. a. Summary statistics of key variables Tables 4 and 5 contain summary statistics for the without and with major events panel datasets, respectively. For example, these tables show that the average annual duration of customer outages (SAIDI) is ~14 minutes (~2 hours) when major events are excluded from the metric, but ~372 minutes (~6 hours) when major events are included. Table 4. Summary statistics for without major events variables Variable (units) Number of observations Min Mean Median Max Standard Deviation SAIDI 2, , SAIFI 2, HDD (#) 2, ,87.1 5,2.7 9,697. 2,23.7 CDD (#) 2,21 1, ,26. 4, Lightning strikes (strikes per customer) 2, Precipitation (inches) 2, Windspeed (mph) 2, T&D lines (customers per line miles) 2, , Share of underground (%) 84.1% 22.2% 2.4% 89.8% 15.3% Delivered electricity (MWh per 2, customer) T&D expenditures ($212 per customer) 2,84 $4.4 $883. $239.8 $52,261. $2,328.4 These tables also show that the average annual frequency of customer outages (SAIFI) is 1.4 events when major events are excluded from the metric and 1.8 events when major events are included. Interestingly, these two tables show that inclusion of major events in the calculation of reliability increases the frequency of outages by ~3% and the duration of outages by more than 25%. Table 5. Summary statistics for with major events variables Variable (units) Number of observations Min Mean Median Max Standard Deviation SAIDI 1, , SAIFI 1, HDD (#) 1, ,16.8 5,329. 9,136. 2,.6 CDD (#) 1,794 1, , Lightning strikes (strikes per customer) 1, Precipitation (inches) 1,

20 2 Variable (units) Number of observations Min Mean Median Max Standard Deviation Windspeed (mph) 1, T&D lines (customers per line miles) 1, , Share of underground (%) 648.6% 24.6% 23.4% 89.8% 16.1% Delivered electricity (MWh per 1, customer) T&D expenditures ($212 per customer) 1,499 $4.4 $734.6 $235.1 $11,76. $1,659.2 b. Test results for the presence of cross-sectional and random effects As discussed, an F-test was conducted to detect for the presence of cross-sectional effects. The results of the F-test (see Table 6) indicate that the null hypothesis of no utility effects should be rejected for all four regressions (i.e., there are cross-sectional effects present in the data). Table 6. Test results for the presence of no utility effects (F-test) One-way fixed effect (utility) Reliability metric F-value Degrees of freedom Reject null of Prob. > F (numerator/denominator) no effects? Log of SAIDI without major events /48 <.1 Yes Log of SAIDI with major events /326 <.1 Yes Log of SAIFI without major events /48 <.1 Yes Log of SAIFI with major events /33 <.1 Yes Accordingly, the following regression specification was used (see equation 18) to analyze the effect of the following regressors on the frequency and duration of electric utility reliability: heating and cooling degree-days; lightning strikes; precipitation; windspeed; T&D extent and expenditures; delivered electricity; the share of T&D line miles underground; and the presence of outage management systems (OMS). Table 7 shows that the Hausman test failed to reject the null hypothesis of random effects for the SAIDI regressions and the SAIFI regression with major events, but rejected the null for the SAIFI regression without major events. I concluded that the random effects model was a more appropriate choice for interpreting the results from the SAIDI regressions and SAIFI (with major events), but that the fixed effects model was preferred for the SAIFI regression without major events. I estimated standard errors for the one-way effects models using a similar method described in Eto et al.(212), which was originally based on methods proposed by Baltagi and Chang (1994) and Wansbeek and Kapteyn (1989). (18)

21 21 Table 7. Test results for the presence of random effects (Hausman 1978) One-way random effect (utility) Reliability metric m-value Degrees of freedom Prob. > m Reject null of random effects at p.1? Log of SAIDI without major events No Log of SAIDI with major events No Log of SAIFI without major events Yes Log of SAIFI with major events No c. Discussion of regression results Table 8 and Table 9 show results for the SAIDI and SAIFI panel regressions with and without the effect of major events. Appendix B contains regression fit diagnostics. General findings In general, reliability events are increasing and lasting longer when major events are included in the performance metric calculation. I found that the frequency and duration of reliability events has increased ~2% (Table 9) and ~8% (Table 8) annually since 2, respectively. This statistically significant time trend associated with major events suggests that severe weatherrelated impacts are becoming slightly more frequent, but this increase in weather-related events is also correlated with relatively longer lasting outages. If major events are excluded from reliability performance metric calculations, there was a statistically significant reduction in the frequency of outages (1% reduction per year), but not a statistically significant change in the duration of outages over time. Higher than average wind speeds and lightning are consistently correlated with increases in the duration and frequency of reliability problems regardless of whether major events are included in the performance metric. What factors are correlated with the duration of reliability events? If major events are excluded, I observe the following statistically significant results for the random effects model: +1% increase in the annual number of lightning strikes is correlated with a +1.2% increase in the duration of reliability events; +5% increase in annual average wind speed is correlated with a +6% increase in the duration of reliability events (see Figure 7); +1% increase in annual average wind speed is correlated with a -2% decrease in the duration of reliability events; and +1 increase in the number of customers per T&D line mile is correlated with a ~7% decrease in the duration of reliability events. Above average wind and lightning but not wind and lighting associated with major events is still correlated with longer duration outages. Customers per T&D line mile a measure of population density is associated with shorter duration outages. This finding implies that urban

22 22 and suburban-based electric utilities may be able to generally restore service more quickly than their rural counterparts. If major events are included, I observe the following for the random effects model: +1% increase in total annual precipitation is correlated with a +17% increase in the duration of reliability events; +5% increase in annual average wind speed is correlated with a +49% increase in the duration of reliability events (see Figure 7); +1% increase in annual average wind speed is correlated with a +69% increase in the duration of reliability events Above average precipitation and lightning is correlated with longer duration outages, but no other potential factors, except for the time trend, are statistically significant in the random effects model (when major events are included) SAIDI (without major events) SAIDI (with major events) Outage Duration (Minutes) % Duration + 69% Duration + 6% Duration -2% Duration 1 5 +% +5% +1% % Above Average Windspeed (2-212) Figure 7. Above average wind and duration of outages (SAIDI)

23 23 Table 8. Results for SAIDI regressions Log of SAIDI (without major events) Log of SAIDI (with major events) Regression Method: Pooled Fixed Effects Random Effects Pooled Fixed Random Effects Effects Intercept *** *** *** (15.384) ( ) (12.525) (38.99) ( ) ( ) Electricity delivered (MWh per customer) -.16*** (.6).128 (.9).1 (.21).2 (.77) (.429).28 (.18) Lagged T&D expenditures ($212 per customer) -.516** (.2) (.379) -.89 (.253) (.313) (.5268).212 (.762) Years since outage management system installation.36 (.85).92 (.9).14 (.79) (.185) -.15 (.32) (.244) Outage management system? -.49 (.63).819* (.491).747 (.488).774 (.1233) (.1527) -.6 (.1363) Abnormally cold weather (% above average HDDs).12 (.43) -.14 (.3) -.12 (.28) (.256) (.268) -.27 (.247) Abnormally cold weather squared () () ().19 (.23).32 (.24).31 (.22) Abnormally warm weather (% above average CDDs).5 (.58).44 (.39).52 (.38).21 (.16) -.14 (.97) -.31 (.96) Abnormally warm weather squared (.2) -.2 (.1) -.2* (.1) -.3 (.2) -.2 (.2) -.2 (.2) Abnormally high # of lightning strikes (% above average strikes).8 (.8).12** (.5).12*** (.5).9 (.15).16 (.15).1 (.14) Abnormally windy (% above average wind speed).184 (.149).224** (.91).237*** (.91).994*** (.36).166*** (.294).176*** (.277) Abnormally windy squared -.3 (.13) -.26*** (.9) -.25*** (.9) -.45** (.21) -.58*** (.2) -.55*** (.18) Abnormally wet (% above average total precipitation) -.74* (.38) -.31 (.28) -.34 (.28).158 (.97).177** (.81).173** (.85) Abnormally wet squared.1* (.1).1 (.1).1 (.1) -.2 (.1) -.2 (.1) -.1 (.1) Abnormally dry (% below average total precipitation) -.6 (.5).3 (.39).25 (.37) (.114) -.86 (.17) -.13 (.15) Abnormally dry squared.1 (.1) (.1) (.1).5** (.2).2 (.2).2 (.2) Year.1 (.75).5 (.67).82 (.62).753*** (.19).786*** (.279).886*** (.237)

24 24 Log of SAIDI (without major events) Log of SAIDI (with major events) Regression Method: Pooled Fixed Effects Random Effects Pooled Fixed Random Effects Effects Number of customers per line mile -.98*** (.13) -.2 (.39) -.68*** (.25) -.34 (.37).354** (.169).85 (.85) Share of underground T&D miles to total T&D miles -.65*** (.15).12 (.52) -.27 (.35) -.149*** (.34) -.47 (.117) (.79) Notes: (1) Standard errors are presented in parentheses underneath coefficient. (2) *** represents coefficients that are significant at the 1% level. (3) ** represents coefficients that are significant at the 5% level. (4) * represents coefficients that are significant at the 1% level. (5) represents preferred model specification.

25 25 What factors are correlated with the frequency of reliability events? If major events are excluded, I observe the following statistically significant results for the fixed effects model 4 : +1% increase in the annual number of lightning strikes is correlated with a +.7% increase in the frequency of reliability events; +5% increase in annual average wind speed is correlated with a +3% increase in the frequency of reliability events (see Figure 8); +1% increase in annual average wind speed is correlated with a -2% decrease in the frequency of reliability events; Frequency of outages increases by 1% for every year that an outage management system has been operational; and +1% increase in share of T&D miles buried underground is correlated with a ~5% increase in the frequency of reliability events. Above average wind and lightning but not wind and lighting associated with major events is correlated with more frequent outages. The finding which implies that outages increase every year after an OMS is installed is counter-intuitive and contrary to findings by Eto et al. (212). Eto et al. (212) found that the installation of an OMS was correlated with more frequent outages, but that an OMS-related learning effect may have reduced the frequency of outages over time. The increase in outages associated with a larger share of underground miles is somewhat counter-intuitive, but may be related to human error (e.g., customer digging that leads to unplanned outages). The benefits of underground miles may be minimal when major weatherrelated events are not included in the reliability performance calculations. Perhaps most interestingly, the frequency of outage events is decreasing about 1% per year (i.e., reliability is improving over time) if major events are not included. If major events are included, I observe the following for the random effects model: +1% increase in annual lightning strikes is correlated with a +1% increase in the frequency of reliability events; +5% increase in annual average wind speed is correlated with a +15% increase in the frequency of outages (see Figure 8); +1% increase in annual average wind speed is correlated with a +18% increase in the frequency of outages Above average wind and lightning is correlated with more frequent outages, but no other potential factors, except for the time trend, are statistically significant in the random effects model (when major events are included). 4 It is important to note that the electricity delivered per customer was marginally significant in the random effects model. In this specific model, 1 additional MWh delivered per customer is correlated with a 3% increase in the frequency of outages.