SEKE 2012 Proceedings - Knowledge Systems Institute

classified. Recall is the probability that a (randomly selected)
relevant data point is retrieved in a search. Precision and recall
are calculated using the formulas in Eq. 2 and Eq. 3
respectively.
measure = 2
( )
(1)
fault events are expected. If a fault event is expected, then the
remaining trained algorithms will determine in what zone,
substation, and feeder the fault is expected to occur, and also if
the fault will occur in an overhead or underground line. The
modules presented in Fig. 5 form the basis for the DSS for
forecasting fault events in the power distribution grid.
=
=
#
#
#
#
(2)
(3)
80
70
60
50
40
30
20
10
0
RPART f-measure
NB f-measure
KSVM f-measure
NN f-measure
Figure 4. Performance Comparison of Machine Learning Models
Based on the highest f-measure of the trained models for
each analysis, the following selections were made to become
the basis for the IOU Decision Support System:
a) NN model was selected to predict faults based on
weather;
b) NN model was selected for prediction of the zone in
the grid where a fault may occur;
c) RPART model was selected to predict the substation
where a fault may occur;
d) NN model was selected to predict if the fault occurs in
the overhead or underground lines;
e) RPART model was selected to predict the feeder
where a fault may occur in the grid.
VI.
The analytic models developed in thi s project and
summarized in Section V formed the basis for the development
of a decision support system (DSS) at IOU to forecast fault
events in their power distribution grid. The analytic models
have been trained using historical data which is pertinent to the
IOU infrastructure and weather patterns in the r egions where
their distribution grid is located. Fig. 5 depicts the modular
view of the architecture of the fault event forecasting DSS, and
the output types that are expected from the trained machine
learning models. Given a certain weather forecast or forecasts,
the Forecast Prediction NN model identifies if o ne or more
FORECASTING FAULT EVENTS AT IOU
Figure 5. Machine Learning Models as Basis for DSS
VII. CONCLUSIONS AND FUTURE WORK
This paper presented a machine learning approach to
forecast fault events in a p ower distribution grid. The
investigation developed and summarized in this paper was
conducted as an applied research project utilizing real-life
historical data from a power distribution utility (referred as
IOU) in the US. The data from IOU dates as far back as 2008,
and includes historical data on fault event occurrences, the
electrical values from the grid at the time that the fault event
occurred, as well as data associated with the topology of the
grid (overhead and underground lines). Historical weather
records collected from the US National Weather Service
(NWS) and from the WeatherBug (WBUG) weather service
during the time period when the fault events occurred were also
included in the datasets employed to train the machine learning
models. An additional historical dataset on lightning, also
provided by WBUG, was included as part of the weather data.
The fields used to fuse fault event data and weather data were
the location and time stamp associated with the fault event. In
situations when readings of weather are taken over certain
intervals of time, a decision has to be made about what weather
reading will be associated with a specific time stamp. In our
study, the weather reading utilized was the one closest after the
fault time stamp reading, since that weather reading covered
the observation interval just ended which encompassed the
fault time. However, an alternate approach would be to use a
weather reading preceding the time stamp reading.
462