Increasing Variability in SAIDI and Implications for Identifying Major Events Days

Transcription

1 Increasing Variability in SAIDI and Implications for Identifying Major Events Days IEEE Power & Energy Society General Meeting 2014 July 30, 2014 National Harbor, MD Joseph H. Eto Kristina H. LaCommare & Michael D. Sohn Lawrence Berkeley National Laboratory (510)

2 Background IEEE DRWG has begun examining trends in the number of Major Event Days identified each year using Standard 1366 IEEE DRWG provided LBNL with daily SAIDI data from past benchmarking analyses to further explore these trends and as well as possible extensions to the method outlined in Standard 1366 Today, LBNL summarizes its findings and observations for IEEE DRWG s consideration and discussion 2

3 IEEE Standard 1366 First developed in 1998 to define reliability indices Amended in 2003 to add a consistent approach for segmenting Major Event Days Uses 2.5*beta to estimate a threshold daily SAIDI, Tmed, above which a Major Event Day is identified Tmed = exp (α+2.5β) Beta = log-normal standard deviation Alpha = log-normal statistical mean Amended in 2012 (MED identification unchanged) 3

4 Evaluating how 1366 performs For a normal distribution: Multiplying beta (the standard deviation) by 2.5 covers % of the expected observations (assuming a one-sided confidence interval) For a year of daily observations, this translates to an expectation of 2.3 Major Event Days per year Using the IEEE DRWG benchmark data, we: computed the Tmed for a forthcoming year (using the prior 5 years); and counted number of MEDs identified (for each year going forward) 4

5 Questions we sought to address Are the number of MEDs identified by Std 1366 increasing over time? Are 5-year sequential groupings of daily SAIDIs normally distributed? How well do extensions to Std 1366 perform? Option 1: Use a smaller pool of years for which SAIDIs may be more normally distributed Option 2: Choose a different multiplier on beta Option 3: Compute an alternative statistical representation for the 5-year group of SAIDIs 5

6 Are the number of MEDs identified by Std 1366 increasing over time? 7.0 Number of major events per utility mean median 100% 90% Answer: Yes! % 70% % of utilities 60% 50% 40% 6 30% 20% 10% 0% 0-3 MEDs 0-2 MEDs

7 Are sequential 5-year grouping of daily SAIDIs normally distributed? 100% 95% % of Utilities with Shapiro-Wilk Test of Normality p-value < α 90% 85% 80% 75% 70% 65% 60% 55% 50% Answer: No! % SW p-value <0.5 % SW p-value <

8 In the slides that follow, we consider the effect of different options for selecting Tmed based on direct extensions of the methods and assumptions inherent in Std 1366 The intent is to evaluate options that may improve our ability to identify a consistent number of Major Event Days over time None of these options should be confused as potential recommendations or suggestions for modifications to Std

9 Option 1: Use a smaller pool of years, for which SAIDIs might be more normally distributed Reducing the # years does increase the probability the data may be normally distributed But there are no noticeable changes in the # MEDs over time 9 9

10 Option 2: Choose a different multiplier on beta 100% 90% 80% 70% 60% 50% 40% 2.5 Beta Increasing the beta multiplier increases the share of utilities experiencing 0-3 MEDs 30% 3.0 Beta 20% 10% 0% 3.5 Beta 4.0 Beta Increasing the beta multiplier can significantly reduce the # MEDs Number of major events per utility No. of 2.5 MEDs No. of 3.0 MEDs No. of 3.5 MEDs No. of 4.0 MEDs

11 Commentary on Options 1 and 2 Efforts to try and fit the data to a normal distribution do not appear to improve the statistical predictability of the data S-W test rejected for >95% utilities not normal Decreasing the historic number of years to compute Tmed doesn t reduce the number of MEDs identified not useful Increasing the beta multiplier (variance) does reduce the number of MEDs identified but doing so implies the data are not normally distributed, statistically speaking not consistent 11

12 Option 3: Compute an alternative statistical representation for the 5-year group of SAIDIs We examined a mixture model consisting of two normal distributions Mixture models are popular for representing data with subpopulations. In this instance: can we use an automatable approach to identify and separate extreme events (e.g., owing to weather or planned outages) from all other events We applied an algorithm to identify the best fitting combination of two normal distributions to the data 12

13 Determining Tmed using 2 normal distributions Step 1: Compute best fitting mixture of 2 normal distributions Select the Tmed so that the cumulative area under the left side of the curve adds to % To the left of the dashed line: A + B = % A Tmed B 13

14 % Utilities with more than 3, 5, or 7 MEDs 100% 90% 80% 70% 60% 50% % IEEE MEDs >3 % Mix MEDs >3 % IEEE MEDs >5 % Mix MEDs >5 % IEEE MEDs >7 % Mix MEDs >7 # MEDs decreases using a 2-component mixture model compared to Std % 30% 20% 10% 0%

15 Comparing the number of MEDs per year 25 IEEE Std 1366 vs. LBNL Mixture Model 20 # major events IEEE-00 Mix-00 IEEE-01 Mix-01 IEEE-02 Mix-02 IEEE-03 Mix-03 IEEE-04 Mix-04 IEEE-05 Mix-05 IEEE-06 Mix-06 IEEE-07 Mix-07 IEEE-08 Mix-08 IEEE-09 Mix-09 IEEE-10 Mix-10 IEEE-11 Mix-11 IEEE-12 Mix-12 Mixture model yields fewer MEDs and lower variance 15

16 What value would Tmed need to be in order to estimate 2.3 MEDs per year on average? Year 2006 Range of Tmed to get ~2.27 MEDs Distance from 2 nd highest (top of vertical bar) to 3 rd highest (bottom of vertical bar) daily SAIDI 16 Identifying 2-3 MEDs per year consistently is still challenging Blue dash: mixture model Tmed Red dash: IEEE Tmed 0.00 u1 u10 u104 u12 u14 U158 u16 U164 U20 U203 u21 u210 U22 u25 u41 u44 U47 u52 u53 u57 u58 u60 u61 u62 u63 u64 u66 u68 u7 u76 u80 u83 u85 U86 u87 U9 U94 16

17 Summary of our findings to date Major events appear to be increasing over recent years Daily SAIDI for IEEE DRWG benchmark utilities are not normally distributed Using fewer historic years to calculate Tmed does not help Using higher multiples of beta can reduce the average # MEDs but is ad hoc; hence, the predictive power for future years requires further study The example of a mixture model is promising, but further study is required 17

18 Next steps and food for thought Pursue efforts to identify subsets of a utility s data that follow a log-normal distribution Take explicit account of pre-arranged interruptions? Segment data within a service territory more finely according to geography? Start a discussion on the objective of methods for identifying MEDs, especially if MEDs are found to be increasing over time Consider whether design targets for distribution are changing over time 18