27.03.2014 Views

SEKE 2012 Proceedings - Knowledge Systems Institute

SEKE 2012 Proceedings - Knowledge Systems Institute

SEKE 2012 Proceedings - Knowledge Systems Institute

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

which the airport data did not contain, and it did not have<br />

weather condition comments), but the pre-processing was<br />

otherwise similar. Lightning stroke data was also obtained for<br />

2009 and 2010 from WeatherBug Total Lightning Network<br />

(WTLN) within a five-mile radius of each substation. Since this<br />

lightning stroke data were time-stamped to the micro-second,<br />

preprocessing it for data mining purposes required various<br />

aggregate counts and sums of strokes and amplitudes for both<br />

intra-cloud and cloud-to-ground lightning. These aggregates<br />

were calculated for both five-minute periods and one-hour<br />

periods so the lightning counts could be joined to the weather<br />

data.<br />

To enable selection of the “closest” weather data for each<br />

fault, preparing all o f this data for mining required aligning<br />

both the timestamps and the geographic locations. Calculating<br />

geographic location parameters was also dif ferent for each<br />

dataset. For the airport weather data, the precise (lat, lon) of the<br />

airport weather station was known. For the hourly weather<br />

data, the (lat, lon) values were not available for the weather<br />

stations, so the geographic center of each weather station’s zip<br />

code was used. Since each microsecond-time-stamped<br />

lightning stroke had its own unique (lat, lon), a reference (lat,<br />

lon) was determined algorithmically to tag the five-minute and<br />

one-hour aggregate records. The precise (lat, lon) had not been<br />

recorded for many faults, so the (lat, lon) of the associated<br />

substation was used for all fault records to ensure consistency.<br />

Each fault and weather dataset had a different reference<br />

time zone: some used local time and some used UTC, and<br />

some that used local time reflected Daylight Savings while<br />

others did not. Therefore, our preprocessing included<br />

calculation of a new timestamp parameter for each dataset,<br />

adjusted to the same reference time zone (local standard time<br />

was chosen). The lightning aggregates were then joined to the<br />

five-minute weather data and the hourly weather data.<br />

Approximate ‘Great Circle’ distances were calculated between<br />

each weather station and each substation, and for each lightning<br />

aggregate, the distances to the four substations were calculated.<br />

These distances were used to choose for each fault the “closest”<br />

five-minute and “closest” one-hour weather record, using both<br />

timestamp and distance.<br />

IV. FORECASTING FAULT EVENTS<br />

Several approaches to predicting fault events in a<br />

distribution grid have been proposed. Butler [1] discusses a<br />

failure detection system, which makes use of electrical property<br />

parameters (such as feeder voltages, phase currents,<br />

transformer windings' temperatures, and noise from the<br />

transformer during its operation) to identify failures.<br />

Gulachenski and Bsuner [6] use a Bayesian technique to<br />

predict failures in transformers. Some of the features<br />

considered in these studies include ambient temperature,<br />

varying loads on the transformer, and age-to-failure data of<br />

transformers. Quiroga et al. [12] search for fault patterns<br />

assuming the existence of r elationships between events. Their<br />

approach considers factors such as values of over-currents of<br />

past fault data and the sequence, magnitude, and duration of the<br />

voltage drops. Although the above mentioned approaches are<br />

predictive in nature, they do not consider weather properties to<br />

predict faults. The hypothesis associated with th e work<br />

presented in this paper is that a fault event occurrence is likely<br />

to follow a pattern with respect to weather conditions,<br />

infrastructure type, and electrical values in the distribution grid<br />

at the time of the fault (as seen in Fig. 1).<br />

The objective of the investigation discussed in this paper is<br />

to develop the basis of a Decision Support System (DSS)<br />

using machine learning that forecasts fault events in the power<br />

distribution grid based on expected weather conditions. The<br />

strategy has two primary phases as shown in Figure 4. In the<br />

first phase, the historical data collected (as explained in<br />

Section 3) is utilized to d evelop and train machine learning<br />

models that can foretell fault events using weather forecasts.<br />

During the first phase, four algorithms are utilized to perform<br />

five generic analyses (fault prediction, zone prediction,<br />

substation prediction, infrastructure prediction, and feeder<br />

prediction) and train the models. Four measurements are used<br />

to compare machine learning algorithms: precision, recall,<br />

accuracy, and f-measure. Precision determines the fraction of<br />

records that actually turns out to be positive in the group the<br />

classifier has declared as a positive class. Recall is computed as<br />

the fraction of correct instances among all instances that<br />

actually belong to the relevant subset. Accuracy is the degree<br />

of closeness of the predicted outcomes to the actual ones. The<br />

f-measure is the harmonic mean of precision and recall.<br />

During the second phase, the best performing algorithm is<br />

selected for each of the five analyses, to be uti lized as part of<br />

the DSS for future predictions at the IOU.<br />

V. CREATION OF MACHINE LEARNING MODELS<br />

Supervised classification techniques are utilized to forecast<br />

the occurrence of faults in the distribution power grid of the<br />

IOU. Four supervised classification machine learning<br />

algorithms are utilized to conduct the analyses: Neural<br />

Networks (NN), kernel support vector machines (KSVM),<br />

decision-tree based classification (recursive partitioning;<br />

RPART), and Naïve Bayes (NB). A NN is an interconnected<br />

group of artificial neurons that use a computational model that<br />

allows them to adapt and change their structure based on<br />

external or internal information that flows through the network.<br />

A SVM is a linear binary classification algorithm [1]. Since the<br />

datasets consist of more than two classes, we choose to use<br />

kernel support vector machine (KSVM), which has been found<br />

to work in the case of non-linear classifications. RPART is a<br />

type of decision tree algorithm that helps identify interesting<br />

patterns in data and represents them as a tree. RPART is chosen<br />

because it provides a suitable tree-based representation of any<br />

interesting patterns that ar e observed in the data sets, and also<br />

because it works well with both nominal and continuous data.<br />

NB is a probabilistic classification technique that works well<br />

with relatively small datasets. These four algorithms are<br />

selected because of their distinct properties and their ability to<br />

work with different types and sizes of data, and the objective is<br />

to select the algorithm that performs the best for the type of<br />

prediction being conducted. Five analyses are conducted<br />

utilizing these four algorithms: (a) fault event prediction; (b)<br />

grid zone prediction; (c) substation prediction; (d) type of grid<br />

infrastructure; (e) feeder number prediction. As mentioned<br />

earlier, the primary data attributes shown in the shaded areas of<br />

460

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!