IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 26, NO.

2, APRIL 2011 1273

Weather Normalization of Reliability Indices

IEEE Task Force on Weather Normalization of Reliability Indices
Heidemarie C. Caswell, Member, IEEE,
Vincent J. Forte, Member, IEEE, John C. Fraser, Member, IEEE, Anil Pahwa, Fellow, IEEE,
Tom Short, Senior Member, IEEE, Mark Thatcher, Member, IEEE, and Val G. Werner, Member, IEEE

Abstract—Weather significantly influences distribution relia- II. WEATHER DATA

bility indices, especially duration benchmarks like SAIDI. We
explore correlations with various weather parameters including Weather can be quantified based on a number of data sources,
lightning-detection network data, and wind from weather stations.
including weather station measurements, lightning detection
This paper explores a number of ways to account for the vari-
ability caused by weather. Approaches include regression models networks, as well as radar and satellite maps, to name a few.
to normalize with weather data as inputs, using outage database Weather stations provide a primary source of weather data
indicators of weather, and modifications to the 2.5 beta method of for use in correlating and normalizing to reliability data.
IEEE Std. 1366. Weather station data is often collected by governments. Such
Index Terms—Lightning, power distribution reliability, relia- site measurement data is collected hourly, and daily, monthly,
bility indices, weather, wind. and yearly aggregations are often provided. The U.S. National
Climatic Data Center has a Global Surface Summary of the
Day dataset that is freely available [1]. This dataset includes
I. INTRODUCTION U.S. and world weather data. Weather parameters can include
mean and minimum and maximum temperature parameters,
precipitation amount, mean wind speed, maximum sustained
wind speed, maximum wind gust, and indicators of fog, snow,

W EATHER is the leading contributor to variability in

reliability indices. Lightning, wind, temperature, rain,
along with snow and ice can all contribute to interruptions.
and thunder. Similar data for Canada is available [2].
Data from a weather station is measured at a fixed point;
often an airport. As such, it may not reflect the weather im-
The variability of these weather parameters contributes di- pacting a power delivery system. Because they have fixed loca-
rectly to variability in interruptions and the time customers tions, weather station data can miss intense portions of storms
spend without power. Large tropical storms and ice storms that occur between sites. Such phenomena are referred to as
cause dramatic changes in the number of interruptions and mesoscale or microscale phenomena. It is also important to note
restoration times during the event compared to an average that the quality of the data collected at some sites is better than at
day. Normal temperature and weather patterns can cause more others. For example, local shielding may make wind attributes
subtle changes in reliability statistics. less accurate in one location than another. Because of these fac-
This paper reports on several approaches to measuring, corre- tors, selection of which sites to use is important.
lating, and normalizing indices for the impacts of weather. The Selection of appropriate weather parameters is important.
science and application of weather normalization are still im- Several groups have found wind parameters to correlate with
mature. Continued progress in this area is expected. reliability parameters. The maximum sustained wind speed is
defined by the World Meteorological Organization (WMO)
as the 10-min average wind speed measured at a height of 10
Manuscript received August 27, 2010; accepted September 09, 2010. Date of m (33 ft). The US National Weather Service defines it as the
publication October 28, 2010; date of current version March 25, 2011. Paper
no. TPWRD-00646-2010 one-minute average. Environment Canada’s Meteorological
H. C. Caswell is with Pacific Power, Portland, OR 97232 USA (e-mail: heide. Service defines it as a two-minute average for hourly readings
[email protected]). and a 10-min average for synoptic readings. Wind gusts exist
V. J. Forte is with National Grid, Albany, NY 12204 USA (e-mail: Vincent.
[email protected]). and are reported only if the peak speed over a 10-min period is
J. C. Fraser is with Nova Scotia Power, Halifax, NS B3J 2W5, Canada (e-mail: at least 15 knots and this speed is at least 5-knots higher than
[email protected]).
A. Pahwa is with Kansas State University, Manhattan, KS 66506 USA
the 2-minute average wind speed.
(e-mail: [email protected]). Within this task force, there was no clear advantage reported
T. Short is with EPRI, Burnt Hills, NY 12027 USA (e-mail: [email protected]; between using maximum gust and maximum sustained speed.
[email protected]).
M. Thatcher is with Kansas City Power & Light, Kansas City, MO 64141
A possible reason for this is that due to the localized nature of
USA (e-mail: [email protected]). wind gust phenomena, any given weather monitoring point is
V. G. Werner is with We Energies, Milwaukee, WI 53203 USA (e-mail: val. not likely to detect the existence of the localized gusts consis-
[email protected]). tently enough to enable a good correlation with plant damage
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. and power outages. Within the Global Surface Summary dataset,
Digital Object Identifier 10.1109/TPWRD.2010.2078839 maximum sustained wind speed is more uniformly available.
0885-8977/$26.00 © 2010 IEEE

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
1274 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 26, NO. 2, APRIL 2011

Lightning detection networks provide another useful set of On the other hand, the modeling process can become very
weather benchmarking data. For a review of lightning detection cumbersome if the system is split into too many small areas.
networks, see Rakov and Uman [3]. Flash density maps pro- Moreover, recorded weather data might not be available for
vide wide and uniform coverage of lightning exposure. We can those small areas. Thus, availability of appropriate data has to
overlay the geographical coverage of the lightning activity on be taken into consideration while determining the smallest area
the circuit or region of interest. for finding the spatial influence. Similarly, to study temporal
Lightning data also allows for a consistent measure of influence, the smallest time duration is the time interval of
severity. The most common measure is the ground flash density recording the weather data, which is typically one hour. In
over a given region and time period. Different time aggregation some instances, smaller time increments may exist, but may not
periods are possible (hourly, daily, weekly, and so on), and be as accurately or consistently populated. Due to the random
these can be aligned with the appropriate reliability indicators. nature of interruption events, interruption data of an hourly
Satellite and radar weather maps provide another source of basis is sporadic.
data and have the advantage of providing spatial coverage (like Randomness of interruptions can be explained by the fact
lightning maps). Research on detection of microbursts and other that severe weather increases the probability of faults, but does
local events could lead to more precise correlations between not necessarily cause interruptions. Therefore, analysis based
weather and damage. A disadvantage is that radar and satel- on time steps of small time duration will not result in any
lite maps are difficult to quantize, meaning that it is difficult to meaningful output. Hence, we need some level of aggregation
convert a time-varying map into a usable indicator for a given of data, either spatial or temporal. Time durations or regions
region. that are too small will result in too much data with very large
Another useful weather parameter is the Palmer Drought statistical spread. Substantial statistical spread yields results
Severity Index (PDSI). This quantifies the severity of a wet or with low confidence. Both temporal and spatial aggregation
dry spell. Utilities have used this to assess soil conditions re- result in data smoothing and provide more meaningful patterns.
lated to weather impacts. Soil moisture is key to soil resistivity; Too much aggregation on the other hand can result in loss
the Palmer index can account for changes in soil moisture and of information.
can be used to assess the impact on lightning-caused damage. As explained previously, spatial aggregation over a larger ter-
Tree-caused faults can also be affected by soil moisture as ritory is not appropriate due to spatial variation of weather. Tem-
more uprooted trees happen during high winds when the soil poral aggregation would require aggregating the interruption
is wet. In the continental US, historical records of the Palmer events over a certain period of time, which could be one day,
index are available in up to ten regions within each state on a one week, or one month. Aggregation need not be over consec-
monthly basis [4]. A similar situation exists in northern areas utive hours. For example, data for all hours during a year with
for frozen ground. Ground temperature readings that would similar conditions could be added to represent a certain type of
indicate the existence of frozen ground are available for a condition. Although aggregation would result in some loss of
number of weather data monitoring stations. resolution, it is necessary to smooth the data.
However, aggregation over a larger geographical area or
larger time durations is not recommended because that could
III. APPROPRIATE RELIABILITY DATA TO USE
lead to significant loss of information. Since SAIDI, SAIFI, and
Utilities typically report annual performance of the distribu- CAIDI are typically computed over a year, they are not very
tion system using commonly accepted indices, which are system suitable for weather normalization. Total numbers of interrup-
average interruption frequency index (SAIFI), system average tions, distribution faults, customer interruptions, or customer
interruption duration index (SAIDI), and customer average minutes of interruption are more likely suitable variables for
interruption duration index (CAIDI) [5]–[7]. Annual indices, weather normalization.
however, are averages that are only a snapshot of performance Utilities keep track of interruptions in their system by cause,
for a given year and do not provide a complete representation with some including weather conditions at the time of interrup-
of the system performance [8], [9]. Various environmental and tion. In many cases equipment failure can be caused by weather
weather factors influence the system performance in a complex conditions such as lightning. Similarly, some of the unknown
way and without knowing this influence it is not possible to interruptions could be due to weather, especially lightning. As
correctly normalize the reliability performance of distribution an example, Fig. 2 shows weather during unknown interrup-
systems, and the subsequent effect on annual metrics. tions for Manhattan, Kansas in 2003 and 2004. Such evaluation
All the environmental factors have both temporal and spatial requires detailed examination of the utility’s outage database
influence on the system. Specifically, weather conditions have as well as the weather database. Although this is not typical,
predominantly temporal influence, which is created by change detailed evaluation of weather conditions during interruptions
in weather from one day to the next. Weather conditions also which are identified as being caused by weather is important to
have spatial variations within the service territory, which be- identify all the weather related interruptions. Sahai and Pahwa
come larger with an increase in the size of the service territory. [10] found that most animal-related interruptions happen during
Territories tend to have complicated microclimates that create fair weather days, which are the days with temperature between
weather volatility. These spatial variations can be accounted for 40 F and 85 F and no other weather-related activity.
by splitting the service territory into smaller areas and recording Chow et al. [11] developed a classification routine to identify
the weather conditions for each of the smaller areas. animal-caused faults based on the following interruption inputs:

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
CASWELL et al.: WEATHER NORMALIZATION OF RELIABILITY INDICES 1275

Fig. 3. Detroit Edison correlations between monthly lightning flash counts and
monthly interruptions.

IV. STATISTICAL CORRELATIONS

From expert experience it can be inferred that worse weather
around the overhead feeders will result in more interruptions
in distribution systems. Examination of interruption data from
utilities reveals that it is not always true that severe weather will
definitely lead to more interruptions on specific days. The inter-
Fig. 1. Lightning ground flash density map (Courtesy of National Grid).
ruptions seem to occur in a random manner with the probability
of faults increasing under adverse conditions. To further com-
plicate the process, each system can be impacted by differing
weather variables. Thus, it is not practical to predict the number
of interruptions for a given day very accurately. However, it is
logical to expect a statistical correlation between the weather
inputs and the variables representing interruption events. Addi-
tionally, the use of T-statistics can be helpful in performing mul-
tivariate regressions for system performance against weather
variables. Linear regression models are the easiest to implement,
but they might not provide accurate results for all the variables
due to the complex nature of the problem.
In areas with appreciable lightning, several groups have found
that the lightning ground flash density correlates well with relia-
bility, particularly with interruption events. Williams [12] found
correlation coefficients of over 0.9 between monthly flashes
Fig. 2. Weather during unknown interruptions for Manhattan, KS, in 2003 and and monthly interruptions for the Progress Florida system. Also
2004. for monthly correlations, as shown in Fig. 3, McDaniel et al.
[13] showed correlation coefficients of 0.54 for Detroit Edison
and 0.81 for Progress Carolina. Correlations were stronger in
circuit ID, weather code, time of day of the event, phases af- areas with more lightning.
fected, and protective device that operated. Animal faults are In eastern Australia, Darveniza et al. [14], [15] found higher
more likely during fair weather in the morning, where only one correlations with wind gusts than lightning in an area with 25 to
phase is affected, and for a transformer or tap fuse. These same 35 thunderstorm days annually based on three years of data. An-
classification strategies can be used to estimate how many of the nual correlation coefficients between maximum wind gusts
“unknown” interruption events are actually animal-caused faults and feeder lockouts were between 0.5 and 0.7. The correlation
and could be excluded from weather correlations. Williams [12] coefficients between lightning ground flashes and feeder lock-
describes an approach on Progress Florida’s data using max- outs were in the range of 0.1 to 0.3. These are for independent
imum likelihood analysis to assign a portion of unknown events linear correlations on days with storm events, which are the days
to animal causes and to lightning causes. His analysis used the where operation staff activates emergency event procedures be-
following interruption characteristics: month, hour, device, and cause of forecast severe weather. During thunderstorms, they es-
weather. Williams successfully used this approach for unknown timate that 80% of the outages and interruptions were from wind
and “wind/storm” interruptions. Of their unknown and storm/ gusts. They also describe how they select weather stations to
wind interruptions, 58% became assigned to lightning, and 35% achieve higher correlations; specifically, they pick stations that:
became animal caused. He was able to use this to more precisely 1) are located to best match customers and 2) have reliable wind
remove animal-caused events for weather correlation and nor- speed data. Reliability of wind speed data is evaluated by com-
malization. paring station weather data and by contacting meteorologists.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
1276 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 26, NO. 2, APRIL 2011

TABLE I
CLASSIFICATION OF VARIABLES FOR THE BAYESIAN MODEL

Fig. 4. Correlations between wind speed and interruption events for (two-hour
window) (Courtesy of Nova Scotia Power Inc.).

Fig. 6. Increasing interruption event rates with adverse weather in Manhattan,

KS, from 1998 to 2003.

In [17], the authors also present a Bayesian approach to

study effects of wind and lightning on interruption event rates
of overhead distribution feeders. Classifications used to build
the Bayesian model are shown in Table I. Expected interrup-
tion event rates based on this classification and the Bayesian
model incorporating historical system performance under these
conditions are shown in Fig. 6. This graph shows a significant
increase in interruption event rates between level 4 and level
5 of lightning. Similarly, an increase in interruption event rate
from wind level 2 to 3 is much higher than that from wind level
Fig. 5. Line interruption rate of overhead distribution feeders due to wind and 1 to 2.
lightning based on a Poisson regression model.

V. NORMALIZATION APPROACHES
Past experience has shown that certain variables, such as wind
A. Interruption Database Indicators of Weather
speed, demonstrate exponential relationships to interruption fre-
quency. Thus, noninear regression can provide better results in In order to obtain useful database information, the data must
some cases. However, in the absence of significant theoretical be collected in a consistent manner. It is best to collect this data
modeling work, the biggest challenge is to select appropriate with a list of choices from which to choose rather than a free
nonlinearities that would provide accurate results for the se- text approach. Using such consistency will also lead to more
lected problem. Exponential regression and Poisson regression accurate data collection.
are some of the nonlinear regression approaches that have been Across the electric power industry, there is a variety of infor-
used in the past for studying effects of weather on interruptions mation that could be utilized as database indicators of weather.
in distribution systems. Radmer and others [16] studied sev- Some utilities collect only a cause, and then based on the cause
eral models for predicting vegetation-related interruption rates reported, they can determine the weather condition (e.g., a cause
of overhead lines. Their models included linear-regression, ex- of lightning, wind, snow, ice or extreme heat are prime exam-
ponential, multivariable linear-regression, and neural network ples). Some of these utilities may also have the cause separated
with temperature, precipitation, and date last trimmed as the in- into primary and secondary causes, such as wind-conductor.
puts. Fig. 4 shows the plot of 2-hour rolling count of new inter- Other utilities use root cause analysis. Utilities may run into
ruption events with respect to 2-hour average wind speed, which problems if using only one cause as a weather indicator. During
clearly shows an exponential relationship. windy conditions, the cause may be reported as fallen limb or
Zhou, Pahwa and Das [17] used multivariate linear regres- as wind. It should be made clear within that utility when a wind
sion and Zhou, Pahwa, and Yang [18] used Poisson regression cause versus a tree cause classification should be used.
to study effects of lightning and wind on interruption events Some utilities collect an addition field of information in their
on overhead lines. Fig. 5 shows results based on their analysis, database based on field observations of weather conditions at
which demonstrates that interruptions vary exponentially under the time of the service interruption. This could be lightning,
very high wind and lightning conditions. wind, mist, rain, etc. If reported accurately, this would be a

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
CASWELL et al.: WEATHER NORMALIZATION OF RELIABILITY INDICES 1277

Fig. 9. Example of normalized CMI with weather causes removed.

Fig. 7. Example of normalized CAIDI with the removal of events during pe-
riods with at least seven interruptions per hour due to weather causes over a
five-hour continuous period.
B. Target Adjustment
Another method for normalizing the year-to-year weather
condition’s effect upon reliability metrics is to create a model
to adjust year-end reliability targets. Thatcher [19] describes
how KCP&L used storm statistics to adjust targets for SAIDI.
This approach requires historical storm statistic analysis to
determine the relative number of storms per year, as well as
the SAIDI impact per storm. Only those storms that would
be included in the reliability metrics would be considered
(those excluded by IEEE 1366-2003 would not be considered).
KCP&L developed a sliding scale based on the number of
storms, and storms were classified based on the number of
Fig. 8. Example of normalized SAIFI with storms removed where a storm is customers out of service.
defined as more than 10 000 customers affected by weather conditions within a
24-hour period. The yearly target for SAIDI could then be set, but coupled
with the historical average number of storms. For every storm
above the average number of storms, the yearly SAIDI target
great source to use for weather normalization. Other utilities would shift up in the amount of the historical SAIDI impact per
may employ more sophisticated means of weather data collec- storm. For every storm below the average number of storms,
tion in conjunction with a service interruption. Some may col- the yearly SAIDI target would shift down in the amount of the
lect, analyze, and store lighting data with the interruption data. historical SAIDI impact per storm.
Others may link local weather reporting station observations Although SAIDI itself is not altered, the effect of the addi-
with their interruption database. No matter what source is uti- tional storm activity is averaged out.
lized for weather database indicators, as long as it is consistent
and accurate, the database can be normalized for weather. C. Modified Beta Methods
The application of the weather normalization of the database IEEE Std. 1366-2003 [5] recommends the 2.5 beta method-
by excluding interruption indicators of weather begins by the ology for discriminating Major Event Days (MED) in perfor-
selection of the weather indicators that are to be excluded. This mance data. The premise of the methodology is that over a suf-
may include all indicators of weather within a database, or more ficiently long period of time, the distribution of the natural log-
likely, it will be a subset of all available indicators. Two similar arithm of the daily SAIDI approximates a Gaussian distribution
ways to do this include: curve. While any threshold level might be adopted, the IEEE
1) Rate—exclude interruption events by weather cause or Std. 1366-2003 recommends that a threshold of 2.5 standard de-
weather condition per hour for at least continuous hours viations from the mean of the normalized distribution be used
where is decided upon by the user. See Fig. 7 for an as the cutoff.
example. Nova Scotia Power Inc. (NSPI) has used a modified beta
2) Persistent activity—use specified weather-related cause(s) threshold to further suppress variability. The 2.5 beta threshold
and/or weather condition(s) over a predetermined contin- identifies approximately 2 days annually, and a 2.0 beta
uous time period combined with a number of interruption threshold identifies on average approximately 6 more days
events threshold or the percentage of total customers-im- annually. For NSPI, the approach of setting aside the data
pacted threshold. See the example in Fig. 8. that exceeds the 2.0 beta threshold eliminated most of the
Fig. 9 shows an example for a northeastern U.S. utility that year-to-year variability in the performance statistics, which
keeps weather codes along with cause codes. With interrup- was the primary objective. The approach also offers the ben-
tions that have weather codes of significant weather removed, efits of being non-subjective, reasonably easy to implement,
the trend line is much flatter. The data are shown for customer understandable, and easily communicated.
minutes of interruption scaled by the mean yearly value with The use of both 2.5 and 2.0 beta thresholds allows perfor-
weather events included. mance data to be stratified for analysis into “Base Operations,”

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
1278 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 26, NO. 2, APRIL 2011

bad weather. For example, a hazard tree-removal program may

be most effective under high winds.
In normalizing, it is critical to identify the “normal” parame-
ters. These drive the expected normal system performance and
can produce a normal system model. This can be used by se-
lecting the worst in years for a given variable, selecting a
mean or median and range of performance. Obviously, this can
result in a complicated normal model. Thereafter, this model
can be adjusted based upon the correlation coefficients of each
weather variable and produce an expected range of system per-
formance.
Choosing appropriate time and spatial conditions is impor-
Fig. 10. NPSI customer hour trend with 2.0 and 2.5-beta cutoffs.
tant. The best correlations may occur for tighter time and spatial
windows, but we normally want reliability measures at wider
windows, such as system SAIDI annually.
“Storm Days,” and “Major Event Days” (see Fig. 10). For NSPI, Williams [12] describes a normalization method for Progress
the Base Operations extend over approximately 357 days of the Florida for lightning data based on a predicted and
year, so interruption statistics in this category accurately rep- . These normalized parameters are estimated indices
resent the normal plant and operational response. Storm Days based on the given lightning level. They use the linear regression
occur with sufficient frequency each year that system design, between the monthly correlation between flashes and customer
plant condition, and operational capabilities are tested to rea- interruptions (with animal-caused interruptions removed). For
sonable levels, and the performance results are meaningful mea- a given month, the customer interruptions are predicted from
sures of business performance for this area of operations. the lightning flash count for that month, and the actual animal-
The scope of the stress imposed on the plant and business caused interruptions are added. If this predicted value (
processes on Major Event Days is open-ended and while the for example) is better than the actual value (SAIDI), the system
performance on these days is of interest, it is a less meaningful performed better relative to the weather that it is subjected to.
measure than for normal and storm operations. Despite that limi- Darveniza et al. [14] found correlation between daily circuit
tation, the performance data for MEDs can be employed to iden- lockouts and maximum wind gust speed for a utility in South-
tify appropriate storm-proofing options. east Queensland, Australia. The correlation emerges when the
An additional benefit was identified when it was realized that maximum gust exceeds 40 km/hr. To help normalize to wind,
by stratifying performance in this way there was very good they suggested the concept of a wind severity index (WSI) de-
alignment with the levels of storm response under the NSPI fined as the sum of the portion of the maximum wind gust that
Emergency Service Restoration Plan (ESRP). Crews were oper- exceeds 40 km/hr
ating in a storm response mode more often than the approximate
2 days annually identified by a 2.5 beta cutoff. However, for the
NSPI system, the 6 “Storm” days and 2 MEDs comprise almost
all of the days each year on which Field Crews are operating in where mWGSmax is the maximum wind gusts in km/hr aver-
one of the ESRP storm response modes. aged over a set of monitoring locations.

D. Normalizing Based on Statistical Correlations

VI. SUMMARY
As addressed previously, a variety of correlation approaches
can be taken. These models will largely be based upon the data Using weather data to directly normalize reliability indices is
available, the temporal parameters used, and the weather vari- challenging. In moderate to high lightning areas, lightning can
ables that appear to drive performance. Expert input may help correlate strongly with reliability, and it is straightforward to
determine the statistical model that is chosen. The exponen- use for normalization. Wind has a nonlinear relationship, and
tial nature of wind forces likely leads to a nonlinear correlation correlations are more difficult to obtain.
model for this variable. By the same token, a linear model using Indirect normalization methods, including the modified beta
lightning flash frequency would be logical. In the situation that method and target adjustment, are easier to apply. The indirect
several variables impact performance, (which is generally true methods generally exclude data, leaving the benchmark seg-
for all utilities) the use of a multivariate regression model will mented to exclude performance during poor weather.
likely deliver the most statistically significant results. Improvement is expected as new approaches for normalizing
If one or more weather parameters significantly correlate with reliability indices are developed, weather data improves, and the
reliability benchmarks, one may attempt to normalize the reli- industry gains more experience with various approaches.
ability data to adjust for weather variations. An advantage of
directly correlating to weather is that it is possible to judge the REFERENCES
performance of the system to weather. If weather parameters
[1] U.S. National Climatic Data Center, Global surface summary of the
are just used to remove the impacts of weather, one might not day—GSOD. Asheville, NC. [Online]. Available: http://www1.ncdc.
see the impacts of improvements that become prominent during noaa.gov/pub/data/gsod

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.
CASWELL et al.: WEATHER NORMALIZATION OF RELIABILITY INDICES 1279

[2] Environment Canada, National climate data and information [12] C. W. Williams, “Weather normalization of power system reliability
archive. [Online]. Available: http://www.climate.weatherof- indices,” presented at the IEEE Power Eng. Soc. Transmission and Dis-
fice.ec.gc.ca/prods_servs/index_e.html tribution Conf. Expo., Dallas, TX, 2003.
[3] V. Rakov and M. Uman, Lightning: Physics and Effects. Cambridge, [13] J. McDaniel, C. Williams, and A. Vestal, “Lightning and reliability—A
U.K.: Cambridge Univ. Press, 2003. comparison of three utilities,” presented at the Rural Electric Power
[4] U.S. National Climatic Data Center, Time bias corrected divi- Conf., Raleigh-Durham, NC, 2003.
sional temperature-precipitation-drought index. [Online]. Available: [14] M. Darveniza, C. Arnold, B. Holcombe, and P. Rainbird, “The rela-
http://www1.ncdc.noaa.gov/pub/data/cirs/ tionships between weather variables and reliability indices for a distri-
[5] IEEE Guide for Electric Power Distribution Reliability Indices, IEEE bution system in South-East Queensland,” presented at the Int. Conf.
Std. 1366-2003, 2003. Electricity Distribution, Vienna, Austria, 2007.
[6] C. Warren, R. Ammon, and G. Welch, “A survey of distribution relia- [15] M. Darveniza, C. Dunn, and B. Holcombe, “A comparison of faults
bility measurement practices in the US,” IEEE Trans. Power Del., vol. and outages on an electrical distribution system caused by lightning
14, no. 1, pp. 250–257, Jan. 1999. and wind gusts,” presented at the 29th Int. Conf. Lightning Protection,
[7] C. Warren, P. Inc, and N. Schenectady, “Distribution reliability: What Uppsala, Sweden, 2008.
is it?,” IEEE Ind. Appl. Mag., vol. 2, no. 4, pp. 32–37, Jul./Aug. 1996. [16] D. Radmer, P. Kuntz, R. Christie, S. Venkata, and R. Fletcher,
[8] R. E. Brown, “Identifying worst performing feeders,” presented at the “Predicting vegetation-related failure rates for overhead distribution
8th Int. Conf. Probabilistic Methods Applied to Power Systems, Ames, feeders,” IEEE Trans. Power Del., vol. 17, no. 4, pp. 1170–1175, Oct.
IA, 2004, Iowa State Univ. 2002.
[9] N. Balijepalli, S. Venkata, and R. Christie, “Predicting distribution [17] Y. Zhou, A. Pahwa, and S. Das, “Prediction of weather-related fail-
system performance against regulatory reliability standards,” IEEE ures of overhead distribution feeders,” Prob. Eng. Inf. Sci., vol. 20, pp.
Trans. Power Del., vol. 19, no. 1, pp. 350–356, Jan. 2004. 117–125, 2005.
[10] S. Sahai and A. Pahwa, “A probabilistic approach for animal-caused [18] Y. Zhou, A. Pahwa, and S. Yang, “Modeling weather-related failures
outages in overhead distribution systems,” presented at the Int. Conf. of overhead distribution lines,” IEEE Trans. Power Syst., vol. 21, no. 4,
Probabilistic Methods Applied to Power Systems, Stockholm, Sweden, pp. 1683–1690, Nov. 2006.
2006. [19] M. A. Thatcher, “Weather normalization of yearly reliability targets,”
[11] M. Y. Chow, S. O. Yee, and L. S. Taylor, “Recognizing animal-caused presented at the IEEE Power Eng. Soc. Gen. Meeting, Tampa, FL,
faults in power distribution systems using artificial neural networks,” 2007.
IEEE Trans. Power Del., vol. 8, no. 3, pp. 1268–1274, Jul. 1993.

Authorized licensed use limited to: University of Saskatchewan. Downloaded on April 01,2023 at 05:00:00 UTC from IEEE Xplore. Restrictions apply.