> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
Abstract—Distribution Automation (DA) is deployed to reduce
outages and to rapidly reconnect customers following network
faults. Recent developments in DA equipment have enabled the
logging of load and fault event data, referred to as ‘pick-up
activity’. This pick-up activity provides a picture of the
underlying circuit activity occurring between successive DA
operations over a period of time and has the potential to be
accessed remotely for off-line or on-line analysis. The application
of data analytics and automated analysis of this data supports
reactive fault management and post fault investigation into
anomalous network behavior. It also supports predictive
capabilities that identify when potential network faults are
evolving and offers the opportunity to take action in advance in
order to mitigate any outages.
This paper details the design of a novel decision support system
to achieve fault diagnosis and prognosis for DA schemes. It
combines detailed data from a specific DA device with rule-based,
data mining and clustering techniques to deliver the diagnostic
and prognostic functions. These are applied to 11kV distribution
network data captured from Pole Mounted Auto-Reclosers
(PMARs) as provided by a leading UK network operator. This
novel automated analysis system diagnoses the nature of a
circuit’s previous fault activity, identifies underlying anomalous
circuit activity, and highlights indications of problematic events
gradually evolving into a full scale circuit fault. The novel
contributions include the tackling of ‘semi-permanent faults’ and
the re-usable methodology and approach for applying data
analytics to any DA device data sets in order to provide diagnostic
decisions and mitigate potential fault scenarios.
Index Terms—Distribution automation, distribution network
data, decision support system, fault activity, fault diagnosis and
prognosis, pick-up activity.
I. I
NTRODUCTION
S load grows and regulatory regimes focus increasingly on
the security of supply and reliability [1], distribution
systems have become much more complex to plan, control, and
This work was supported by SP Energy Networks.
X. Wang, S. D. J. McArthur, and S. M. Strachan are with the Institute for
Energy and Environment, Department of Electronic and Electrical Engineering,
University of Strathclyde, UK. (e-mail: xiaoyu.wang@strath.ac.uk,
s.mcarthur@strath.ac.uk, scott.strachan@strath.ac.uk).
J. D. Kirkwood and B. Paisley are with Engineering Services of SP Energy
Networks, UK. (e-mail: bruce.paisley@spenergynetworks.co.uk).
maintain. Therefore, in order to effectively manage these
aspects of network operation, major investments and
developments have been undertaken in the area of Distribution
Automation (DA). In particular, DA is applied for the purpose
of automatic sectionalising and localisation of network faults
and improving restoration times, with an aim to reduce
Customer Minutes Lost (CML) and Customer Interruptions
(CIs) [2], and generally improve customer service.
The ongoing implementation of DA has led to increasing
volumes of operational data becoming available from
Intelligent Electronic Devices (IEDs) [3] and other DA devices.
These data sets can be analysed to provide fault diagnosis and
network event information and predictions, but are not always
capitalised upon. There have been attempts to use the data
through diagnostic algorithms within automated decision
support systems [4].
Previous research has delivered diagnostic capability in
related areas. The data captured from Supervisory Control and
Data Acquisition (SCADA) or Advanced Metering
Infrastructure (AMI) combined with DA devices has been
automatically analysed using intelligent system methods [5].
This has provided fault analysis and diagnostic assistance for
protection engineers [6][7]. There has also been a focus on fully
automating diagnostics and integrating these into operational
systems [8][9], and the use of novel approaches for network
monitoring [10].
Meanwhile, other research has concentrated on the practical
use of IED data as the primary source to support automated data
analysis to assist engineers. This has focused on monitoring
system behavior and improving system operation [11][12].
Popovic et al. [13] proposed a fully automated data analytics
solution to improve decision-making processes by interpreting
the network data retrieved from digital protective relays.
This research takes the next step. It automates the analysis of
Pole Mounted Auto-Recloser (PMAR) data to diagnose PMAR
faults. In addition, it automatically identifies emerging circuit
and device faults to allow preventative measures to be taken to
avoid or minimise outages. Specifically, the contributions of
this paper are: research, design and proof of a predictive
capability to identify emerging faults within a DA application;
the combination of data science and knowledge based
techniques to deliver diagnostic and prognostic functionality; a
unique focus on distributed DA autoreclosing devices which
A Data Analytic Approach to Automatic Fault
Diagnosis and Prognosis for Distribution
Automation
Xiaoyu Wang, Student Member, IEEE, Stephen D.J. McArthur, Fellow, IEEE, Scott M. Strachan,
Member, IEEE, John D. Kirkwood and Bruce Paisley
A
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2
allows the identification of evolving or incipient network
faults; and, the full implementation of a prototype and
evaluation against actual industry data. This research addresses
a gap in the field of power system data analysis, which is the
use of Low-Voltage Distribution Network IED data for fault
diagnosis and prognosis. The methods can be adapted for other
DA devices.
II. C
IRCUIT PERFORMANCE AND UNSOLICITED OPENINGS
A PMAR placed on the overhead lines of a distribution
network [14] is an example of a protection IED [15] which
combines monitoring and communication capabilities. It
automatically isolates faulted sections of an overhead line
circuit, while simultaneously capturing the fault details and
storing them in log files (e.g. current and voltage amplitude,
affected phase, etc.) [16]. Fig. 1 diagrammatically shows the
operation of the device.
Fig. 1. PMAR operation
When a fault occurs, the PMAR attempts a set number of
reclosure attempts (typically 3 times, as set by the operator in
this study). If a reclosure attempt is unsuccessful (indicating the
continued presence of a fault) the PMAR device remains open
for a period of 10 seconds before attempting a reclose. If the
fault remains on the circuit during all three reclosure attempts,
the PMAR will ‘lockout’ on its third attempt, suggesting that a
‘permanent fault’ (or non-transient fault) may exist on the
circuit (see Fig. 1(a) above). Following a ‘lock-out’, the PMAR
can only be reclosed manually via telecontrol by the control
engineers. Therefore, it is clear that a PMAR can only remain
closed if the fault has been removed from the circuit or isolated
from the PMAR in question. Fig. 1(b) shows the operation of
the PMAR after a short-lived ‘transient fault’ has appeared and
then almost immediately been removed from the circuit. Here
the PMAR trips when the fault is initially detected and, as the
fault clears during the 10 second period of isolation, the PMAR
remains closed after its first attempt at reclosure.
‘Semi-permanent faults’ are intermittent faults that go
undiagnosed [17]. An example cause of such a fault is rain
affecting a cracked insulator on a wood pole (which can dry out
and then no longer provide a fault path). These can adversely
affect the quality of daily electricity service through ‘nuisance
tripping’, which incur regulatory penalties for network
operators. Semi-permanent faults result in unsolicited PMAR
trip activity, referred to as ‘Unsolicited Openings’ (UO) by the
network operator.
From a technical/asset management perspective, the
frequency of UOs on a particular circuit could provide an
indication of deterioration in a circuit’s (or indeed an individual
PMAR’s) performance and underlying condition. In addition,
the corresponding log files generated by the PMAR device may
contain ‘pick-up’ activity (line current and voltages) providing
a deeper insight into the device/circuit condition and evidence
of evolving fault conditions. For example, the cracked insulator
mentioned previously could ultimately cause an initial PMAR
trip when moisture ingresses the crack. As the PMAR is then
reclosed (in some instances multiple times) this can cause the
moisture to evaporate and the PMAR to remain closed after a
reclosure attempt. This restores the supply to the circuit for a
time until the scenario repeats itself in the future. As the circuit
supply is restored, there is usually the assumption that some
prolonged transient fault has cleared with no lasting damage to
the circuit, but this may not in fact be the case. In addition, from
a customer service perspective, assessment of a circuit’s UO
activity provides an indication of which customers may be most
prone to ‘nuisance tripping’, compromising the level of
customer service they experience. This exposes the
Distribution Network Operator (DNO) to regulatory penalties.
The PMAR device generates log files containing overhead
line current and voltage data sampled at 12.8 kHz. This is
referred to as ‘pick-up’ data. It has a Main Processor Module
(MPM), and along with its core protection functionality, it also
records the sampled data and generates alarms based on its
internal operation.
For the purpose of diagnosing this root cause and predicting
evolving faults, this research focuses on the design and
implementation of a Decision Support System (DSS) that
assists engineers in recognising PMAR behaviour. This is
based on analysis of ‘pick-up’ data stored in PMAR log files.
The DSS has been developed and deployed as a
Knowledge-Based System (KBS) [18][19] with the objective of
diagnosing PMAR device faults. It also provides detection and
early warning of semi-permanent circuit faults responsible for
UOs, which threaten to evolve into more serious permanent
faults. These adversely impact on customer service, and the
data analysis functions to combat this are detailed in Section III.
The DSS includes visualisation tools. These support the
analysis of circuit behavior by engineers and the development
of predictive and diagnostic rules, as shown in the Case Studies
in Section V.
III. DATA ANALYSIS FOR DIAGNOSTIC AND CLASSIFICATION
FUNCTIONS
This section outlines how the KBS generates and implements
diagnostic rules from expert knowledge and engineers’
experience. It also describes how it uses the data analysis
methods and approaches to diagnose PMAR device faults and
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
3
detect semi-permanent faults on the circuits. This functionality,
in conjunction with the predictive capability for fault prognosis
described in Section IV, was built into the prototype software
DSS described in Section V.
A. Diagnosis of PMAR Device Faults
The condition monitoring of the PMAR is supported by its
integrated Main Processor Module, which also controls the
operation of the PMAR. The MPM records details of on-line
activity (i.e. fault current and voltage amplitudes, etc.) and the
condition of the PMAR’s components. These recorded
conditions or details contain the information on
existing/potential device faults or circuit faults.
Therefore, after the raw log file data is imported, the
knowledge-based system (KBS) automatically invokes the
diagnostic rules (based on the experience of operational
engineers), which characterise known PMAR faults. The aim is
to prevent potentially delayed or failed operations in response
to overhead line faults. These rules operate on the PMAR log
file data [20] and are shown in Fig. 2.
Fig. 2. Rules for diagnosing PMAR device fault
As illustrated in Fig. 2 there are four main device faults
which can be diagnosed and reported to engineers. These faults
are identified through: the interpretation of alarms generated by
the MPM; or, calculations on interval times between the PMAR
operations and status changes within the log file. As an
explanation:
• An MPM fault could be confirmed by detecting alarms
in the log file, which were generated automatically by
the MPM.
• A typical driver module fault always leads to the failure
of the PMAR operation for a fault within the protected
zone. The KBS identifies this when the ‘Excessive To’
(i.e. contact opening time exceeds setting time) or
‘Excessive Tc’ (i.e. contact closing time exceeds setting
time) events are indicated in the log file.
• A tank fault is characterised by multiple open and close
status changes within impracticable time periods (i.e.
duration from Open to Close is less than 10 seconds).
The MPM recognises incorrect driver operation and
generates the alarm “Undefined”.
• A microswitch fault generally reflects a problem in the
PMAR’s switch. The KBS identifies this fault by
comparing the actual time periods of status changes with
primary settings through automatic time stamp
calculations.
This is the first stage of automated decision support, which
focuses on PMAR device fault diagnostics. The next stage of
the system concentrates on overhead line fault diagnoses.
B. Semi-Permanent Fault Detection on Circuits
As discussed previously, a Semi-Permanent Fault (SPF)
manifests itself through sporadic periods of intense PMAR
operation on a circuit. It is therefore necessary to first classify
PMAR log file data into different classifications of PMAR
operation. The next step is to filter the activity consistent with
SPFs. Trends of classified SPF candidates can then be used to
confirm the existence of SPFs. To assist in the detection of
potential SPFs, a further rule-base was designed to
automatically classify the PMAR operations being experienced.
The rules were developed through engineers’ expert knowledge
and knowledge of the PMAR’s operating mechanism.
Following classification, a visualisation tool is used to allow
engineers to observe the results and allows them to identify data
trends and patterns associated with PMAR operations that
could be due to potential SPF activities. In addition, data can
also be visualised to substantiate the diagnosis of invoked rules.
The following sections describe how the rules are generated
and the trends evaluated to detect SPFs. The visualisation tools
are described in Section V.
a) Classification of PMAR Operations
Using the log file ‘pick-up’ data recorded during the
operation of PMARs under different fault conditions, four
specific rules were defined. These were based on the experts’
knowledge and understanding of how these faults are likely to
manifest themselves in these data sets. The rules are then
implemented within the KBS to group ‘pick-up’ activity into
different classifications of PMAR operation depending on the
pick-up duration and the number of corresponding PMAR
operations. These are fault ‘pick-up’ activity resulting in: no
trip (FP); single trip (ST); multiple trip (MT); and, lockout (L).
Fig. 3 shows the rules for classifying PMAR operation (i.e. FP,
ST, MT and L) in order of increasing severity.
Using these rules the DSS can then identify, map out and
prioritise potential SPF activity on circuits which have
experienced pick-up activity exceeding acceptable time limits.
From this prioritised mapping of potential SPF activity across
the network, experts can then drill into the data using the
visualisation tool for a deeper analysis of the pick-up data. This
allows them to discern the details of PMAR operation against
phases affected, evidence of earth faults, etc. The Case Studies
in Section V show trends of interest identified using the
visualisations.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
4
Fig. 3. Rules for classification of PMAR operation
b) Evaluation of Behavioural Trends for SPF
Detection
Following the classification of PMAR operation, their trends
are then analysed. This provides an insight into potential SPFs.
Moreover, the statistical evaluation of ‘pick-up’ activities
associated with these PMAR operations could substantiate the
existence of potential SPFs. The trends could directly reflect
the evolving conditions associated with this form of fault. Fig. 4
shows an example of the trends and statistical features for
considering FP PMAR operation (as detailed in Fig. 3). These
support engineers in determining whether a SPF exists on the
circuit. These trends and statistical features are applied to other
classifications of PMAR operation (i.e. ST, MT and L) as well.
Fig. 4. Trends of statistical features
With respect to Fig. 4, the DSS will calculate and
demonstrate the frequency of a number of FP operations for a
specific PMAR. Where there is an increasing rate in the
cumulative frequency distribution (CFD) (as shown in Fig.
4(a)), this suggests the FP operations resulted from faults
becoming more frequent and the corresponding frequency
distribution (FD) (as shown in Fig. 4(b)) exceeds the average
number of PMAR operations. The high number of FP
operations during the same time period also indicates an
evolving fault on the circuit (i.e. a SPF). To confirm the
existence of the SPF, the system extracts the duration time (DT)
of each associated ‘pick-up’ activity and determines the
interval time (IT) between consecutive pick-ups through
calculation. Trends showing an increase in DT and a decrease
in IT suggest the circuit’s condition is worsening. This may also
be indicative of the stage of maturity of the semi-permanent
fault in terms of its closeness to becoming a permanent circuit
fault. Using the visualisation tool, the users obtain a perspective
on the evolution of faults and can make informed and faster
decisions regarding the existence of SPFs within the network.
IV. P
ROGNOSTIC FUNCTIONS
The KBS described in the previous sections can assist control
engineers by diagnosing PMAR device faults and identifying
potential SPFs present on overhead line circuits. In addition to
this functionality, further work has been conducted into
developing a prognostic capability which can predict potential
PMAR operations (or evolution of SPFs into permanent faults).
It grades the circuit’s pick-up activity in terms of the
imminence of such a threat. This would enable maintenance
staff to take evasive action and potentially avoid expensive and
prolonged outages required to repair damage from permanent
faults.
In developing this prognostic capability, this paper proposes
a structured method of data mining [21] to derive rules
operating on anomalous pick-up activity. This is defined as a
pick-up duration of greater than 30ms which has not yet led to a
PMAR operation (previously classified in Fig. 3 as “FP”).
These rules will then be capable of predicting future PMAR
operations (i.e. classified as ST, MTs, and L in Fig. 3). The
following sections detail the process of data mining applied to
the analysis of FP pick-up activity data to generate the
predictive rules. This can be reapplied to other DA and
operational devices.
A. Data Preparation
FP activity is grouped into one-month time windows, which
do not contain any recorded trip or lockout activity. The short
one-month window of interest was selected to focus in on
activity resulting from the same underlying cause
(semi-permanent fault condition).
The training data set was produced from 12 PMAR log files.
This represents pick-up activity captured at a sampling rate of
12.8 kHz across 7 circuits, spanning a period of up to 5 years. A
total of 100 FP groups were derived from this training data set
and subjected to the data mining process discussed in Section
IV.B below.
The features selected for the predictive algorithm were:
1. Number of Total Pick-ups (NTP) in each FP group.
2. Time to Trip (TTT), which is the time duration from
the last recorded pick-up activity in the FP group to
the next PMAR operation.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
5
3. Interval Time Trend (ITT) is a Boolean value which
represents the increasing or decreasing of time interval
between two consecutive fault pick-ups in a group of
activity (i.e. True for decreasing, False for not
decreasing).
4. Average Current Amplitude (ACA), which is the
average current amplitude of FPs in each group.
5. Average Pick-up Duration (APD), which is the
average pick-up duration of FPs in each group.
B. Data Segmentation and Visualisation
Since the PMAR’s operation associated with the five defined
features is unlabelled, data-class associations are unknown. In
order to extract hidden associations between the data
characteristics and those of FP activities, clustering techniques
were applied.
The K-Means algorithm is a clustering technique used for
data mining which is simple to implement and straightforward
to interpret. It partitions similar unlabelled data (i.e. data
without defined categories or groups) into a pre-set number (K)
of clusters [22]. When compared with alternatives, such as the
hierarchical clustering algorithm [23], K-Means provides
increased accuracy. That is, it has more possibilities of
partitioning the data point to the correct allocated cluster when
the relative weights of the features are not well understood.
Therefore, the application of a K-Means algorithm enables
segmentation of the data into distinct clusters where specific
clusters could be considered indicative of distinct PMAR
operating conditions [24].
To visualise the clustering output from the K-Means
algorithm, a dimensionality reduction technique is required to
process the clustered data. The t-Distributed Stochastic
Neighbour Embedding (t-SNE) methodology is one of a set of
algorithms that can transform high-dimensional data into two
or three dimensions [25]. Other techniques, such as Sammon
Mapping [26], can be used for this purpose. The t-SNE
algorithm was adopted to transform the data into a more
visually appreciable two-dimensional (2D) representation due
to its fast computation, high accuracy and efficiency of
processing non-linear data.
Fig. 5 demonstrates the comparison of cluster distributions of
feature vectors as the number of clusters in the K-Means
algorithm is increased. From the visualisation through the
t-SNE technique, the aim is to identify one or more specific
clusters with the indicative features’ values, which could be
used for generate the predictive rules.
Fig. 5. t-SNE visualisation of K-Means clustering
As shown in Fig. 5, each cluster is separated with boundary
lines. Each point represents one set of feature vector
coordinates representing an instance of FP activity. The relative
distance between points provides an indication of the
similarity/dissimilarity between plotted feature vectors. The
closer together these points are clustered, the greater the
similarity that exists between associated FP activity. From the
t-SNE visualisation, it is apparent that the data is distributed in
two main areas, i.e. that inside and outside the circled area.
While the membership of the clusters outside the circle changes
significantly as the number of initialising clusters increases, the
cluster membership inside the circled area does not.
It is evident that the distributed data in the circled area is
more stable and consistent, and demonstrates a higher
correlation coefficient. This means that the feature vectors
within this cluster are most likely to lead to improved accuracy
if used to build the predictive model. The purpose of the
prognostic function is to predict potential future operation.
Based on this analysis it is apparent that, if a new log file is
presented to the KBS, rules operating on the NTP, ITT, ACA
and APD features can be used to determine the TTT. That is,
given the four feature values, the TTT (i.e. time to the next
PMAR operation) can be predicted within a particular time
window. This allows engineers to take actions before the
occurrence of a fault causing PMAR operation.
C. Predictive Rule Implementation
In order to generate rules to predict future PMAR operation,
Fig. 6 shows a parallel chart describing the distribution of
values of the features characterising the cluster representing
PMAR operation.
