![](bg2.png)
Section 6 presents conclusions and future work that
authors are pursuing in this field.
2 DATA ANALYSIS APPROACH
Utilities face difficult challenges regarding how to
use available data for storm planning. First, the
current available data are used primarily for tracking
purposes and not for proactive storm planning.
Second, the sources of data and data are
heterogeneous in nature. Third, relying on data
collected on past storms is challenging as no two
storms are the same. These present a challenge when
comparing storm-restoration performance of the past
and present (Johnson, 2004). Another major data
challenge is that utilities do not have a standardized
method for collecting data on storm-restoration.
In spite of these challenges the authors believe it
is possible to demonstrate the potential predictive
capabilities that machine learning models can
provide with current data sources, imperfect as they
may be. These data originates from heterogeneous
sources and geographically dispersed environments.
Primary data types available can be classified as
static structured and unstructured historical data and
dynamic real-time structured and unstructured data.
Static data can be used to develop machine learning
models while dynamic data can be used by trained
models to analyse storm situations in near real-time.
Static structured historical data includes GIS and
grid topology, storm data, grid damage, OMS data,
work management systems, work flow management
with power restoration actions, grid intelligent
electronic device (IED) data, vegetation and terrain,
and transmission and generation data. Static
unstructured historical data includes on-line social
networks historical data, historical multi-media
storm damage data, and unstructured damage
reports. Dynamic real-time structured data includes
weather feeds, grid sensor feeds, real-time OMS
data, emergency response data, SCADA data, phasor
measurement unit data, 61850 GOOSE data,
network management and fault data, meter data, and
IED data. Dynamic real-time unstructured data
includes real-time OMS data, drones or robotic
systems data, multi-media data, and repair crew
report.
3 THE DATA
Storm damage projection refers to the use of
prediction methods to project the severity and
locations of damages, resource needs and time for
power restoration after a storm has hit the power
grid. Storm damage measurements include peak
number of customers without power, outage
duration, peak number of line restoration personnel,
and equipment damage. Based on the projection,
plans are made for positioning restoration resources,
prioritizing repairs, and minimizing disruptions.
3.1 Data Sources
To develop machine learning models for storm
damage projection we looked into several data
sources public, proprietary, structured, unstructured,
and acquired historical data related from an
electrical Utility in the US, referred as Public Utility
due to proprietary constraints. The sections below
provide some details on the data used.
3.1.1 Weather Data
As a big source of public data, National Weather
Service (NWS) has a large collection of historical
weather data that can be downloaded through its
website. The following two weather data sets are
used in this study.
Severe weather event database for the United
States from 1950 to 2011. A severe weather event is
identified by timestamps of event type, begin date
and time and end date time, begin and end locations
of latitude and longitude, and a magnitude of
severity. Typical events for the selected region
include hail, thunderstorm wind, flash flood, and
tornado. The database contains over 900,000 records
with a total of 1.1 GB. This data is used to identify
severe storms in this study.
Hourly weather data from over 10,000 weather
stations all over the world from 2000 to 2012. The
hourly weather include location of observing
weather station in latitude and longitude, observation
time, wind direction, wind speed, air temperature,
sea level pressure, precipitation time and
accumulation, etc. The total amount of data is over
220 GB. In this study we used some of the weather
conditions as inputs to the predictive models. The
weather stations are selected within half a degree
from the boundaries of a Public Utility. Although the
number of stations is increased over the years, it is
still very small if we want to have one station every
few square miles. In this project we relied on data
interpolation to derive weather condition for
locations at fine granularity.
DataAnalyticsforPowerUtilityStormPlanning
309