Distribution Transformer Health Monitoring using

Smart Meter Data

Kavya Ashok, Dan Li, Deepak Divan and Nagi Gebraeel
Georgia Institute of Technology, Atlanta, GA, USA
Email: kashok6@gatech.edu

Abstract—The distribution electric grid has become a highly of hot-spot temperature for in-service transformers can be
complex and intelligent network with changing load and expensive and difficult [6], work has been done to estimate
customer types. This has generated unprecedented challenges transformer hot spot temperatures using local weather data
and opportunities for utility companies–opportunities especially
in the area of asset health/performance management. Moreover, and power consumption from smart meters [5]. However,
several utilities are increasingly moving from the traditional these equations were developed for large power transformers
reactive and time-based asset monitoring approach to a more (LPTs) and the empirical constants recommended for these
proactive condition based method. However, this needs to be equations can vary widely in real life, resulting in significant
done in a low-cost and efficient manner. This paper explores error. Certain parameters depend on the type and make of
how existing sensor infrastructure such as smart meters can
be utilized to provide utility operators with more visibility into the transformer which are not always recorded accurately in
the health and operation of their assets. The paper focuses the utility’s database or might be difficult to obtain from
primarily on service transformers. transformer manufacturers.
Analyzing of the condition of insulating oil and paper is
Index Terms- Distribution Transformer, Health Monitoring, In- one of the most common tests used to evaluate transformers
sulation Degradation, Smart Meter data.
in service. Chemical methods such as dissolved gas analysis
(DGA), degree of polymerization (DP), Furan testing etc. test
I. I NTRODUCTION
for the presence of certain compounds that are released during
The electric grid is one of the largest man-made infras- the breakdown of insulating material in the presence of heat,
tructures and is extremely critical for the development and partial discharge or arcing [2]. These methods have shown
progress of any economy. Traditionally, generation and trans- high accuracy and are routinely employed by utility operators
mission systems have been heavily monitored by utilities. for large power transformers. However, they require laboratory
Distribution system assets on the other hand are difficult to testing of multiple oil samples and are therefore not feasible or
monitor due to their lower relative cost, high volume and scalable for low cost service transformers that are not equipped
lower importance compared to their high voltage counterparts. with the necessary outlets to conduct these tests.
Utilities have traditionally opted a ‘run-to-failure’ approach Researchers have also developed various other sensors to
with distribution assets, i.e. replace it when the asset fails. detect insulation degradation caused by partial discharge (PD).
In recent years however, the distribution network has become Two commonly used PD detection methods include measure-
more complex, decentralized and dynamic with the addition ment of electrical signals or acoustic signals produced by the
of unpredictable loads and renewable energy. Utilities have PD. However, acoustic methods are severely limited by signal
begun moving towards a condition-based monitoring approach attenuation and ambient noise, and electrical measurements
rather than time-based and prefer to have more visibility into are limited by electromagnetic interference problems [2].
the health and performance of their distribution equipment. Moreover, unlike LPTs, retrofitting service transformers with
This paper focuses on providing a low-cost scalable health sensors to monitor its health and maintaining a communication
monitoring framework for service transformers in distribution back haul, or scheduling routine maintenance checkups can
networks. Our framework provides utilities with actionable be expensive at the scale required with very low return on
insights about their service transformer fleets. investment.
Various methods have been developed to monitor trans- This paper explores the possibility of using smart meter
former health. As transformers age, their internal insulation data to detect high voltage insulation degradation in service
degrades. This can increase their risk of failure and vul- transformers. Literature has shown that one of the common
nerability to severe conditions such as lightning strikes and failure modes involves inter-layer shorts on the high voltage
short-circuits. The primary cause for insulation degradation is winding [3], [4]. When inter-turn insulation deteriorates it
overloading and excessive temperatures. IEEE loading guide results in arcing and shorts between turns causing the effective
C57.91–1995, Clause 7 [1] has developed empirical equations number of turns to reduce which results in an increased
to determine top-oil and hot-spot temperature as well as secondary voltage. By utilizing existing sensor infrastructure
remaining lifetime. These parameters are a function of load to track voltage abnormalities on the secondary, utilities can
and ambient temperature. While direct on-line monitoring monitor transformer’s health at negligible cost and effort.

This work has been funded through Department of Energy through grant DE-OE-0000877
 Section II explains the degradation detection method and B. Algorithm Implementation
algorithm. Section III describes the simulation setup and data We propose a data-driven method, to detect the degradation
extraction to test the developed approach. Section IV discusses of a transformer using data from its neighboring transformers.
the results of the algorithm on the synthetic smart meter dataset The following attribute the steps of our methodology.
followed by conclusion and next steps in Section V. 1) Estimating Transformer Secondary Voltage: The first
II. A LGORITHM D ESCRIPTION step is to estimate the secondary voltage V sec for each trans-
The algorithm builds on the assumptions that: former. This can be done by using voltage correlation to pair
meters with their respective service transformers [9] and then
1) The primary network topology of the feeder is known,
estimating the impedance of the cables from the transformer
i.e. transformers immediate neighbors can be identified.
to the meter using linear regression [10]. Alternatively, various
2) There has been negligible degradation for an initial time
IoT sensors as described in [11] can directly record the
period when smart meter data (voltage and power) was
secondary voltage.
recorded.
2) Formulating a linear regression problem for each Trans-
A. Algorithm Overview former Pair: For every transformer i in a feeder, we identify
Equations [1]-[5] and Fig.1 show that the voltages on the the set of its neighbors j, denoted as Ni . We use linear
secondary of neighboring transformers can be related in the regression to estimate the coefficients βj for all j ∈ Ni i.e.,
following method: X Pk (t)
Visec (t) = βjT [Vjsec (t), ] + j (6)
V1pri = N1 × V1sec (1) Vjsec (t)
k

V2pri = N2 × V2sec (2) where k ∈ D(j) is the set of all transformers downstream
V1pri = V2pri = V pri (3) of transformer j. Thus, we have |Ni | linear models for each
transformer. This regression model is formulated using an
V1sec /V2sec = N2 × N1 = constant (4)
initial subset of the data during which we assume negligible
pri
where, N indicates the turns ratio of the transformers, V degradation for t ≤ n.
and V sec are transformer primary and secondary voltages. 3) Computing Residuals: Each linear model is then used to
sec
calculate the voltage of transformer i, V̂i,j (t), based on the
sec
new observation [Vj (t), Pk (t)] where t > n. The residuals
rj , defined as the difference between the actual voltage of
transformer i and its estimation given by the linear model are
computed as:
rj (t) = Visec (t) − V̂i,j
sec
(t), (7)
i.i.d.
where rj (t) ∼ N (0, σj2 ).
Fig. 1. Primary network connection
We then take the average of the residuals rj (t) for all j ∈ Ni
to get the average residual ri :
Since N2 × N1 is a constant, the ratio of the transformer’s 1 X
secondary voltages should also remain a constant. However ri (t) = rj (t) (8)
|Ni |
during peak loading periods, assumption (2) above is not j∈Ni

completely accurate. A small voltage drop between the two The standardized residuals are given as:
nodes can be correlated with the power flowing downstream.
Accounting for this, the equation can be updated and formu- r̃i (t) = ri (t)/σ i , (9)
lated into a linear regression problem as follows: such that the standardized residual r̃i (t) follows a standard
X normal distribution N (0, 1).
V1 sec = β1 .V2sec + β2 . (P )/V2pri +  (5)
P 4) Degradation Alarms: We monitor the degradation pro-
where, (P ) is the sum of all power flowing downstream of cess of transformer i at each time t > n. For a healthy
V2pri and accounts for the voltage drop between the transform- transformer, the linear model should give an unbiased esti-
ers. β1 is a function of the transformers’ turns ratio while β2 mation of the voltage Visec (t), i.e., r̃i (t) ∼ N (0, 1). When the
is proportional to the resistance of the medium voltage cable transformer is degrading, the actual secondary voltage of the
between the two transformers. transformer i is higher than the estimated voltage calculated
If there is no insulation degradation, i.e. change in effective using its neighbors. There will be a positive mean shift µ in
turns ratios, the coefficients of regression will remain the the standardized residual: r̃i (t) ∼ N (µ, 1) where µ > 0.
same for every time step. However, for transformers with Since the degradation process is cumulative and irreversible,
degrading high voltage insulation, the residuals defined as the a cumulative sum control chart (CUSUM) is used to detect
difference between subsequent predicted and actual voltages, degradation [12], [13]. The chart plots the cumulative sum
will increase over time. of deviations from the target for individual measurements or
subgroup means. It is described using two tuning parameters;
k detects the shift from the mean, and h, the threshold limit.
Here we use a one-sided CUSUM control chart to detect the
positive mean shift in the residuals. The CUSUM statistic S(t)
is calculated using [10]

Si (t) = max{0, Si (t − 1) + r̃i (t) − µ0 − k} (10)

where µ0 = 0 is the in-control mean. The detection alarm is

raised when S(t) > h, implying a consistent and significant
deviation between expected and measured secondary voltages.
5) Ranking Degraded Transformers: Since transformers
can degrade at different rates, utilities can identify and pri-
oritize failing transformers by regularly updating and ranking
S(t) in descending order.
Fig. 3. Total active feeder load during the year long simulation
This method ensures that expected deviations in voltage
due to loading, tap changes, solar power injection etc. do
not raise false alarms. By tracking every transformers’ voltage classes 0.2 (accuracies within +/- 0.2%) [8]. More details of
combined with its neighbors’, only transformers with changing the simulation can be found in [10]
turns ratio due to shorted windings can be flagged.
A. Simulating Transformer Degradation
III. S IMULATION S ETUP AND DATA E XTRACTION To simulate transformer insulation degradation 6 out of 590
service transformers were chosen at random. During the 1 year
To generate synthetic smart meter data, a realistic distri- long simulation, their primary base voltage rating (effectively
bution feeder was simulated in OpenDSS. The simulation is the number of turns on the primary) was reduced gradually to
based on EPRI’s Test circuit 5 [7]; a 12.47 kV residential a total of less than 0.6%. The first 3 months were considered
feeder with 1379 residential loads and 591 transformers. as a base period during which no degradation occurred. The
12 months of consumption data at 1-minute intervals were transformers experienced different speeds and duration of
extracted from Pecan Street and high-resolution voltage data simulated degradation as described in Table I and Fig.4. This
was recorded at each consumer node during the power flow variability was included in the simulation to test the robustness
simulation. This consumption data was collected from homes of the detection algorithm.
located in Austin, Texas between 2013-2017. Descriptions of
the simulated feeder and load profiles are shown below from TABLE I
Fig. 2 to 3. R ATE , SPEED AND DURATION OF TRANSFORMER DEGRADATION
The feeder contains rooftop solar power injection, tap
changes and capacitor banks. Noise was injected into the data Transformer Start Day Months of Rate of
Number of the Year Degradation Degradation
with a uniform distribution of 0.2% of the base voltage, 240V.
This corresponds to ANSI standard C12.20-2010 with meter 5 90 3 Fast
175 90 3 Fast
250 160 5.5 Slow
343 120 6 Slow
483 160 5.5 Slow
562 200 1 week Very Fast

The simulation allows us to determine the minimum extent

of degradation required for accurate detection using AMI data.
It is not possible to distinguish a degrading transformer by
simply tracking its secondary voltage. Based on the simulated
data for example, Fig. 5 (a) shows the change in voltage for
transformer #175 with and without degradation for 1 year. The
increase in secondary voltage can be negligible and difficult to
identify. However, a significant jump in voltage ratio between
the transformer and a neighbor (#49) can be seen in 5 (b), that
sec
is V175 /V49
sec
. This forms the essence of the algorithm described
in section II.
IV. R ESULTS
The results of the algorithm applied to the dataset above
Fig. 2. Circuit plot of EPRI circuit 5 is described from Fig. 6 and 7. Fig. 6 shows the change in
 The red plot in Fig. 6 correspond to degrading transformers.
During the base period (first 3 months) no alarm is raised.
Transformer 5 begins to degrade first and it can be seen
that the CUSUM statistic rises consistently with time as its
residual r̃5 is greater than k. This occurs as long as its
present turns ratio is different from the original value. On
the other hand, transformer 483 experiences a much slower
and delayed degradation process. As a result, the CUSUM
statistic is significantly lower and will be ranked below the
other degrading transformers in terms of priority. Over time,
#483 will also generate a higher CUSUM as more data is
observed.
Fig. 7 shows the number of alarms generated by all trans-
formers over the entire year and the last three months. Two
false positives are initially generated by transformers 159 and
Fig. 4. Change in transformer turns ratio over one year. Default = 30. 217. This is due to peak summer loading periods (as seen in
Maximum of 0.6% change in turns ratio Fig.3) that cause a higher voltage drop across the primary
cables than expected by the linear model. However, their
CUSUM statistic starts decreasing when their standardized
residuals r̃ reduces, and eventually drops below the control
limit h.
We tune parameters (k, h) based on the precision-recall
measure, which is suitable for imbalanced data; only 6 degrad-
ing transformers out of 591. The precision (TP/TP+TF) and recall
(TP/TP+FN) are calculated based on true positives (TP), false
positives (FP) and false negatives (FN). The tuning parameters
are selected based on the F1 score which is the harmonic mean
of precision and recall as given below:

Precision + Recall
F1 = (11)
Precision · Recall
The tuning parameters were swept across a wide range
from: k ∈ {0.1, 0.2, 0.5, 1.0, 1.5, 2.0, 2.5, 3.0}, and h ∈
{0, 5, 10, 100, 500, 1000, 5000, 10000}. k = 2.5, h = 100
were selected as the optimal parameters that maximized the
F1 score. The confusion matrix for degrading transformers
under this setting is given in Table II. A 99.6% accuracy was
achieved, where
(T P + T N )
Accuracy = (12)
(T P + F P + F N + T N )

TABLE II
C ONFUSION MATRIX FOR DEGRADING TRANSFORMER DETECTION

Detected as Degraded Detected as Not Degraded

Transformers
6 (TP) 0 (FN)
Degraded
Transformers
2 (FP) 583 (TN)
Not Degraded

Fig. 5. a) Secondary voltage and b) voltage ratio with neighbor 49 for The detection delay computes the minimum insulation
transformer 175 with and without degradation degradation required for accurate detection as a percentage
change in turns ratio from a base period of 30. This has been
calculated for each failing transformer in Table III. An average
CUSUM statistic for all transformers over a period of 1 year. of only 0.33% reduction was required for accurate detection.
 TABLE III
5
D EGRADATION DETECTION DELAY FOR FAILING TRANSFORMERS

Transformer Detection Turns Ratio Min Change

50000

562 Number Delay (TR) at Detection Required in TR

CUSUM satatistic

5 92 29.895 0.347 %
175 387 29.947 0.173 %
30000

483 10782 29.932 0.225 %

343
343 6174 29.923 0.254 %
562 26 29.794 0.684 %
175
10000

250

control limit h 483 tively identify and replace deteriorating transformers before
0

an outage occurs. By tracking the voltage ratio of neighboring

0 100 200 300 transformers, the algorithm avoids false positives caused by
time (days of the year) changes in voltage due to tap changes, solar power injection
etc. The algorithm averages the alarm over multiple neighbors
Fig. 6. The CUSUM statistics over 1 year for all the transformers. Control
chart parameters k = 2.5, h = 100
to increase confidence in detection accuracy.
Certain limitations to this algorithm will be explored in
future work. For example, the affect of insulation degradation
on the transformer’s secondary, although less common, can
5 negate the effect of that on the primary. Future work will also
25000
number of alamrs

include simulations and tests at an individual transformer level

175 562
instead of feeder level to validate the degradation assumptions
159
343 listed in this paper.
10000

217250 483
R EFERENCES
[1] “IEEE Guide for Loading Mineral-Oil-Immersed Transformers,” in IEEE
Std. C57.91-1995, 1996.
0

[2] M. Wang, A. J. Vandermaar and K. D. Srivastava, “Review of condi-

tion assessment of power transformers in service,” in IEEE Electrical
0 100 200 300 400 500 600 Insulation Magazine, vol. 18, no. 6, pp. 12-25, Nov.-Dec. 2002.
transformer index [3] “Secondary (low-side) surges in distribution transformers,” in IEEE
(a) Transactions on Power Delivery, vol. 7, no. 2, pp. 746-756, April 1992.
10000

[4] D. Hayes, “Investigation of Secondary High Voltage caused by Turn-to-

5 175 343 483 562 Turn Failures in Pole Type Distribution Transformers”, Presented to the
250 Missouri Valley Electric Association, April 21-23, 1987
number of alarms

[5] G. Atkinson and M. Thottan, “Leveraging advanced metering infras-

tructure for distribution grid asset management,” 2014 IEEE Confer-
6000

ence on Computer Communications Workshops (INFOCOM WKSHPS),

159 Toronto, ON, 2014, pp. 670-675.
[6] S. Kulkarni, K. Ashok, F. Lambert, D. Divan, “Asset monitoring using
smart sensing and advanced analytics for the distribution network,” to
0 2000

217 be presented at 51st IEEE North American Power Symposium, 2019.

[7] EPRI Circuits [Online]
Available: http://svn.code.sf.net/p/electricdss/code/trunk/Distrib
[8] ANSI 12.20-2010 for Electricity Meters - 0.2 and 0.5 Accuracy Classes.
American National Standard Institute (ANSI)
0 100 200 300 400 500 600
[9] T. A. Short, “Advanced metering for phase identification transformer
transformer index identification and secondary modeling”, IEEE Transactions on Smart
(b) Grid, vol. 4, no. 2, pp. 651-658, 2013.
[10] K. Ashok, M. Reno, L. Blakely, D. Divan, “Systematic Study of
Fig. 7. Number of alarms raised by the CUSUM control chart a) throughout Data Requirements and AMI Capabilities for Smart Meter Analytics,”
the year, b) in the last 3 months for all the transformers 2019 IEEE International Conference on Smart Energy Grid Engineering
(SEGE), Oshawa, ON, 2019, pp. 53-58
[11] S. Kulkarni et al., “Enabling a decentralized smart grid using au-
tonomous edge control devices”, in IEEE Internet Things J., pp. 1-1,
V. C ONCLUSION Feb. 2019.
[12] Cao, Y., Xie, Y., & Gebraeel, N. (2018). Multi-sensor slope change
The aim of this paper was to demonstrate a robust method detection. Annals of Operations Research, 263(1-2), 163-189.
for distribution transformer insulation and health monitoring [13] Mousavi, M. J., & Butler-Purry, K. L. (2010). Detecting incipient
without the need for installing additional sensors. By using faults via numerical modeling and statistical change detection. IEEE
Transactions on Power Delivery, 25(3), 1275-1283.
smart meter voltage and power data, which is widely deployed
on the US electric grid today, it is possible to identify those
transformers that have experienced degradation in their high
voltage insulation. This will allow utility operators to proac-