Use of smart-meter data to determine distribution system topology

Jeremy Donald Watson1, John Welch2, Neville R. Watson1

1
Electrical and Computer Engineering Department, University of Canterbury, Private Bag 4800, Christchurch, New Zealand
2
Vector Network Ltd., Auckland, New Zealand
E-mail: [email protected]

Published in The Journal of Engineering; Received on 25th February 2016; Accepted on 21st March 2016

Abstract: Smart-meter data presents an opportunity for utilities to improve their database records, and develop a low-voltage (LV) model
which may be useful for outage management and fault detection, isolation and response, phase balancing, and network planning. In addition,
impact assessment studies on new technologies can be performed. This study presents several contributions in the area of determining the
topology of the LV distribution system. This is in terms of identifying the transformer a particular installation control point is connected
to, and the phase if that customer is single-phase. First, harmonic voltage correlation is proposed as it is more robust to noise and missing
records than the prior algorithm of voltage correlation. Second, it is demonstrated that smart-meter data can be used to determine the trans-
former/phase to which a customer is connected and update database records in this regard. To achieve this, a new algorithm based on correl-
ation analysis with the Fisher Z transform is developed. Third, a method to estimate LV feeder and service main impedances is presented.
Further work is necessary; however, the results from trials in Auckland, New Zealand are highly promising.

1 Introduction Having reliable data is of immense usefulness for outage manage-

ment and fault detection, isolation, and response. It may be used
In New Zealand, as in many other countries, smart-meters are being
for phase balancing, which will reduce losses and improve the
deployed rapidly. As of December 2014, 62.3% of the 1.75 million
power quality supplied to any three-phase loads which may be
customer connections (called installation control points or ICPs) in
present in the distribution network. It is also valuable for network
New Zealand use smart-meters, with an installation rate of 17,000
planning, as well as for impact assessment studies on new technolo-
per month at present [1]. Owing to the vast amount of data recorded
gies such as distributed generation and electric vehicles.
by these meters [2], the concept of big data analytics is in the vogue.
Recent research has explored using the data recorded by the smart-
meter for: detection of electricity theft [3–5], determining typical
2 Background
customer consumption patterns [6], load forecasting [6], and deter-
mining the topology and parameters of the low-voltage (LV) Though this is a relatively new area, there is some previous litera-
network [7–19], among other aspects. This paper presents a contri- ture which considers the use of smart-meter data to determine LV
bution in the latter area of determining the topology of the LV dis- network connectivity and parameters. Several different approaches
tribution system. This is in terms of identifying the transformer a have been proposed for determining LV topology; however, the
particular ICP is connected to, and the phase if the customer is simplest and perhaps most successful to date is using voltage cor-
single-phase. relation to determine the connectivity of the LV network, as in
There are several reasons which may cause the LV network infor- Short [7], Arya and Mitra [8], Pezeshki and Wolfs [9, 10], Luan
mation to require updating. Not all utilities record the transformer to et al. [11, 12], Seal and McGranahan [13], Wen et al. [14], and
which each ICP is connected to, and very few record the phase. LV Bolognani et al. [15]. One of the most important advantages of
(415 V) networks are typically run as radial feeder systems with a this technique is that no additional hardware, monitoring equip-
normally open switch between one LV feeder and an adjacent LV ment, or signal injection equipment is required. Its disadvantage
feeder network. Sometimes, for example, due to loading issues on is that there tends to be a high correlation through the low-
a transformer, a line crew may move the breakpoint, thereby trans- impedance medium-voltage (MV) network, or between different
ferring some ICPs from one transformer to an adjacent transformer. phases close to the transformer.
At other times, ICPs are transferred to another phase in order to Short [7] developed a model which removes load distortion (i.e.
reduce the voltage unbalance. The LV feeder typically runs along the difference in voltage from the smart meter to the feeder at the
the street either overhead or underground, and a short service line connection point) in order to correlate each ICP voltage time-series
connects each customer to the feeder. The meters are located at with other ICPs and constructs the LV system bottom-up. One of
the customer end of the service line. the major hurdles in voltage correlation is that MV voltage
This trial was carried out on part of Vector Networks’ distribu- changes can cause sites near the distribution transformer to correlate
tion network. Vector is the largest utility in New Zealand and more strongly to similar sites on a different transformer. The per-
owns the distribution network which supplies Auckland, the formance of load correction removes this problem, but requires a
largest city in New Zealand by population. Vector is taking an in- very high ‘coverage ratio’ (i.e. a high proportion of sites with avail-
novative and proactive approach to identify the transformer and able metering data). The impedances are also determined to a very
phase to which an ICP is connected to and smart metering data pre- high accuracy by linear regression; once again requiring a relatively
sents an opportunity for improvement in this area without requiring high coverage ratio. Arya and Mitra [8] used clustering based on the
thousands of expensive and time-consuming field checks. The Pearson correlation coefficient and geographical location for phase
smart-meter data was obtained from metering companies and is and transformer identification. The geographical location data is,
actual ‘live’ data from the Vector system. most likely, used to prevent correlations through the MV network
The end goal is to use smart-meter data to build a model that can dominating LV correlations. Pezeshki and Wolfs [9, 10] showed
be used to simulate the LV network and thus manage it better. the application of the voltage correlation technique within a

This is an open access article published by the IET under the Creative Commons J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101
Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/) doi: 10.1049/joe.2016.0033
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
single distribution network, avoiding voltage correlations due This is not the same as the average voltage over a half-hourly period
to MV. (as in previous studies) and it is important to determine whether
Luan et al. [11, 12] presented a similar correlation analysis. voltage correlation still performs well.
Voltage correlation factors between all customers are calculated, Third, the LV network in New Zealand may be significantly dif-
and the ICPs with very strong correlations are assumed to be on ferent to overseas LV networks. Furthermore, the LV network is
the same transformer. Furthermore, the magnitude of the voltage often widely varied, including both cables and overhead lines,
is used to indicate the meter’s upstream/downstream position. long and short feeders etc.
Seal and McGranahan [13] correlate voltage information from the To overcome these potential problems and differences to previ-
utility’s supervisory control and data acquisition (SCADA) ous studies, several new methods are proposed. The key innova-
system with voltage information from customer electricity meters tions of this paper are:
to identify the service phase. In our paper, voltage information
from the utility’s SCADA system was not available. Wen et al. † proposing the use of harmonic voltage data for voltage correl-
[14] developed an algorithm based on analysing cross-correlations ation (Section 3.2);
over voltage magnitudes along with phase angle differences on two † establishing a method using the Fisher Z transform to determine
candidate phases to be matched. If two measurement points are on LV connectivity and update database records in cases where much
the same phase, large positive voltage magnitude correlations and of the load is unmetered, and smart-meter records were not well-
small voltage angle differences should be observed. However, high- synchronised (Section 3.3); and
precision phasor measurements are not present on the distribution † using smart-meter voltage and load data to estimate the service
networks in New Zealand. main and feeder impedance without requiring a high coverage
Bolognani et al. [15] estimated the grid topology from a correl- ratio (Section 3.4).
ation analysis of the voltage signals, proving via simulations that
the developed algorithm is indeed effective in reconstructing the
correct topology. The requirements are that the measurement 3 Methodology
noise is limited and the number of samples is sufficiently large. 3.1 LV connectivity: voltage correlation
Khumchoo and Kongprawechnon [16] proposed the use of cluster-
ing based on the wavelet reduction to classify customers correctly. Voltage correlation uses one of the most basic principles of electric-
A similar procedure in this trial was unable to identify all customers al engineering. When a current flows through an impedance, a
correctly, probably due to the low ‘coverage ratio’ in our trial. voltage drop occurs. Hence, in a power system, a similar voltage
Finally, Peppanen et al. [17, 18], Wang and Luan [19], and profile over time suggests that the two metering sites are closely
Berrisford [20] present methods to use smart-meter data to deter- connected electrically. This is demonstrated in Figs. 1 and 2,
mine distribution secondary parameters. The approach in [17] is where Fig. 1 is the correlation between two closely related sites
similar to the simplified approach used to estimate network impe- on the same phase of the same transformer, whereas Fig. 2 is for
dances in this trial. two sites which are physically close, but on different phases of a
The data obtained from smart-meters in this trial differs from the different transformer.
data used in previous studies in several ways. The Pearson correlation coefficient is calculated as follows
First, in most cases there was a significant proportion of un-
metered loads, or load at which smart-meter data was not available
 cov(x, y)
x, y = (1)
for various reasons. The data reliability/completeness on the trial sx sy
sites ranged from <50% to about 75%, much lower than the
97.28% recorded in [17], for example. It should be noted that the strength of the correlation (or the mag-
Second, most of the data recorded instantaneous voltage and the nitude of the Pearson coefficient) is affected by factors such as:
timing between meters was not synchronised (up to 3 min apart).
Instantaneous load data was also included. One batch of data (i) The frequency of sampling.
recorded the maximum and minimum voltages in each half-hourly (ii) Whether instantaneous or average values are recorded, or
period instead, as well the voltage total harmonic distortion (THD). otherwise.

Fig. 1 Voltage correlation between two sites on the same phase of the same
transformer Fig. 2 Voltage correlation between two sites on different transformers

J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101 This is an open access article published by the IET under the Creative Commons
doi: 10.1049/joe.2016.0033 Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/)
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
(iii) The time synchronisation between meters. 3.2 LV connectivity: voltage THD correlation
(iv) Line losses and unmetered load.
The inclusion of voltage THD in several of the datasets presented
the opportunity of using this for correlation instead of the actual
Load correction is one method used to improve the effectiveness voltage. This is done very simply by calculating an N × N matrix
of voltage correlation. The procedure of using the instantaneous R of voltage THD correlation coefficients between each pair of
power measurements to improve voltage correlations (i.e. to ICPs, where N is the number of ICPs included in the analysis
reduce uncertainties) is well known in literature, e.g. [2, 11, 12].
The idea is that the voltage at the feeder, rather than at the smart- ⎛ ⎞
meter, will be more closely correlated to related ICPs. This is
r(1, 1) = 1 · · · r(1, N )
only possible with the datasets that included instantaneous load/
⎜ .. .. .. ⎟
R= ⎜
⎜ . . . ⎟
current measurements. In most other studies, the service main im-
⎟
⎝ r(N − 1, 1) r(N − 1, N ) ⎠
pedance was already known. Here, the service main impedance r(N , 1) · · · r(N , N ) = 1
must be estimated. Only the power Pins and the voltage Vins were
provided by the smart-meters. In this project, the method was
Fig. 4 shows an example of the extremely high correlation coeffi-
adapted as follows.
cients typically obtained between closely related ICPs using
voltage THD.
Step 1: Calculate the instantaneous current Iins
For sites that are not on the same transformer/phase, the Pearson
correlation coefficient ranged from 0.5 to 0.9. Fig. 5 shows an
Pins example of the voltage THD profiles, where ICPs 2 and 7 are on
Iins = (2) the same transformer/phase while ICPs 2 and 8 are not.
Vins

3.3 LV connectivity: database correction

The reactive power was ignored, since it was very small in the Once the voltage correlations or THD voltage correlations are cal-
data provided, and did not have a notable effect on the voltage (cor- culated, this needs to be converted into reliable ICP to transformer/
relation magnitude <0.1).

Step 2: Estimate average service main impedance Zest and calculate

new voltages. For every recording

Vfeeder(time, ICP)  Vins (time, ICP) − Iins × Zest (3)

where Vfeeder(time, ICP) is the estimated voltage at the coupling

point of the customer service wire to the feeder, VICP(time, ICP) is
the meter voltage, and Iins is the current calculated in step 1.
A more refined approach would be to estimate the service main
impedance at each ICP separately, but this poses new challenges.
As explained in step 3, the ideal impedance is when the average cor-
relations are at a maximum. The estimated service main impedance
at any ICP is dependent on that at other ICPs, leaving a multivariate
problem which is nearly impossible to solve. Fig. 3 Change in correlation with load correction
Step 3: Evaluate the average correlation of the dataset. A higher
average correlation indicates that ‘noise’ is being suppressed in
the voltage data.
Step 4: Solve for the estimated service main impedance which max-
imises the correlations.

The typical figures for the average service main impedance are
shown in Table 1.
These were verified by accessing the utility’s records to deter-
mine a typical conductor size and length for the service mains
and then calculating the impedance. By comparing the result for a
row of shops to residential areas (Table 1), the lower figure for
Street 1 is expected as it is a row of shops, compared with residen-
tial areas in the other two streets (2 and 3). The result of this process
for the data of Table 1 is shown in Fig. 3.

Table 1 Service main impedance estimates

Area Best-fit estimate of service main impedance, Ω

Street 1 (shops) 0.050

Street 2 (mostly residential) 0.100
Street 3 (mostly residential) 0.115 Fig. 4 Voltage THD correlation between two sites on the same phase of the
same transformer

This is an open access article published by the IET under the Creative Commons J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101
Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/) doi: 10.1049/joe.2016.0033
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Step 10: Calculate the mathematical confidence of each ICP to
group pairing. The probability that a particular ICP to group
pairing is correct is

p(correct group) = 100% − probability of error

= 100% − p(other groups) (6)

Using the assumption that the Fisher Z transform yields a Gaussian

distribution [21, 22], the probability of each other group being the
‘correct’ one is
1
1 2
/2s2
p(group) = e−(t−m) dt (7)
2p2 x

where x is the difference between the group correlation found in

Fig. 5 Voltage THD profiles compared; same transformer/phase (upper),
step 7 from the best correlation, and the standard deviation σ is [21]
different transformer/phase (lower)

√ 2
phase pairings. This section details and discusses the method used s= variance = √ (8)
N samples − 3
to achieve this and detect possible inaccuracies in the Vector data-
base. As mentioned previously, ICPs close to a transformer present
where N samples is the number of records per ICP.
a particular challenge, as do low coverage ratios (due to not all
smart-meter data being available). For this reason, Fisher Z trans-
form correlation analysis (FZTCA) has been developed to look Step 11: For each ICP, check that the mathematical confidence, as-
for the database inaccuracies. The FZTCA algorithm was more suc- suming a Gaussian distribution, is higher than a pre-determined
cessful than simply taking the highest correlations as in [12], or threshold (95% appears reasonable) – otherwise, revert the ICP/
averaging the Pearson correlation coefficients themselves. The transformer relationship to the database one. A confidence of
FZTCA method is outlined as follows: <95% (or a chosen acceptable confidence threshold) indicates that
there is no strong correlation match for the ICP in the data.
Step 1: Load voltage (THD voltage if available) data into two-
dimensional (2D) matrix. Indices are (sample, ICP).
Step 2: Calculate Pearson correlation coefficient between each pair, 3.3.1 LV network impedance: The LV network appears to be pri-
R(ICP1, ICP2). This is a matrix of dimensions N × N, where N is the marily resistive. The correlation between reactive power and the
number of ICPs in the analysis. voltage is not particularly clear – which is quite different to
Step 3: Perform Fisher Z transform on the matrix R. This is required higher voltage levels, where reactive power is the most important
in order to normalise the distribution of the correlations, and it also factor influencing the voltage magnitude. However, it must be
allows calculation of the average correlation of an ICP to a group. noted that the reactive power magnitude is on average much
At each point the Fisher Z transform is smaller than the real power magnitude. This is shown in Fig. 6 at
a randomly selected ICP. At some ICPs in the data, the magnitude
1 + R(ICP1 , ICP2 ) of the Pearson correlation reaches 0.1, but no higher (which shows
z(ICP1 , ICP2 ) = (4) no significant correlation).
2(1 − R(ICP1 , ICP2 ) In contrast, the real power has a marked influence on the voltage,
as demonstrated at the same ICP (Fig. 7).
where z is a positive real number bound between zero and infinity. Hence, the method takes each change in real power consumption,
Step 4: Search the database to determine starting ICP to transformer and the associated change in voltage. With enough points, this
relationship.
Step 5: Assign random phases to each ICP: however, ensure all
three phases are represented. Each ICP now has a preliminary trans-
former and phase.
Step 6: Cluster ICPs related to each transformer into three best-fit
phases. K-means clustering was used here.
Step 7: For each ICP, calculated its correlation to each existing
transformer/phase group based on the Fisher Z transform

N
x=1 z(ICP, groupICPs(x))
groupCorr(ICP, group) = (5)
N

Assign each ICP to the most strongly correlated group [i.e. highest
value in (5)]. ICPs which are alone in a group are not moved:

Step 8: Loop through groups and assign them to the transformer that
the majority of the ICPs belong to in the database. Step 7 can cause
an ICP or a group of ICPs to end up on a different transformer to the
one they all are assigned to in the Vector database.
Step 9: For each ICP, calculate its correlation with its own group
and the other groups in the database. Fig. 6 Reactive power – voltage correlation

J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101 This is an open access article published by the IET under the Creative Commons
doi: 10.1049/joe.2016.0033 Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/)
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
The slope of the least-squares fit may then be used to estimate the
ICP to transformer impedance.
One way to solve for branch impedances is to solve for the esti-
mated branch impedances that maximises the correlation. Fig. 9
shows this.
ICP1 and ICP2 are known to be on the same phase. In theory, the
correlation between the voltage profiles of ICP1 and ICP2 will be at
a maximum if load correction (Section 3.1) is performed as follows




Vload corrected ICP1 = Vins ICP1 − Iins ICP1 × Z(service, 1)



 (9)
Vload corrected ICP2 = Vins ICP2 − Iins ICP2
× (Z (service, 2) + Z (feeder, 1)) (10)

where Vload corrected is the voltage profile after load correction, Vins is
the meter voltage recording at the load, Iins is the instantaneous
power recorded by the meter divided by the instantaneous voltage
Fig. 7 Real power – voltage correlation
Vins, and Z is the branch impedance as shown in the graph. By
solving for Z(service, 1) and Z(service, 2) + Z(feeder, 1 to 2) such
that the correlation between Vload corrected(ICP1) and Vload corrected(2)
is at a maximum, an estimate of the network impedance is given.

4 Results and discussion

4.1 LV connectivity
The connectivity results can be presented in graphical form, as in
Fig. 10, illustrating the use of the voltage correlation technique.
The data samples were from 15 min instantaneous recordings, and
there were 352 samples per customer. The colour code in Fig. 10
indicates which phase the ICP is connected to and which transform-
er is supplying the ICP. The number inside the marker is the average
voltage at the ICP. The open point may be clearly identified by
visual means in Fig. 10, and two ICPs of the 56 at which data
was available had their transformer relationship corrected
(numbers 17 and 17A).
Other feeders in the network were also analysed throughout the
project. Verification was achieved, where possible, by a mixture
of actual site visits, Google Street View, and the Vector geograph-
ical information system (GIS) system; and the transformer/phase
allocations as shown in Fig. 10 were correct in every verified
Fig. 8 Change in real power against change in voltage with best-fit line
case. However, underground connections were not able to be veri-
fied in many cases. Note that the FZTCA algorithm to determine
provides an estimate of the impedance back to the transformer. The LV connectivity is independent of any of the methods or pro-
major assumptions are: grammes used for validation.
Table 2 shows a comparison of the voltage THD and fundamen-
tal voltage correlation methods for a test site, using the FZTCA al-
† Very low correlation with the load at other ICPs (true for the gorithm. The average mathematical confidence is taken over all
dataset used in the trial). ICPs in the analysis according to Section 3.3, and is relatively
† Transformer secondary voltage does not change notably due to low due to the low coverage ratio of the particular feeder at
load fluctuation at a single ICP. which voltage THD data was available.
† Reactive power changes are small enough that they can be As shown in Table 2, voltage THD correlation was found to be
ignored. more effective than using the actual voltage profile in the conditions
of this test. The voltage THD correlation between ICPs on the same
This provides a graph of data points in 2D, as illustrated in Fig. 8. phase of the same transformer is usually much higher than the

Fig. 9 Sample section of LV feeder

This is an open access article published by the IET under the Creative Commons J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101
Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/) doi: 10.1049/joe.2016.0033
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 10 LV connectivity result for one street

voltage correlation in all the data samples encountered, whereas power electronic equipment often causes high levels of 3rd or 5th
ICPs on different transformer did not appear to be strongly corre- harmonic. Thus, the correlation in voltage THD from one trans-
lated. The results are shown in Table 2. Note that the ICP to trans- former to another is not as strong as when the voltage, which is
former/phase mappings were verified by a combination of actual regulated, is used. Moreover, discriminating phases should be
site visits, Google Street View, and the Vector GIS system. easier with harmonic correlation. Using harmonics for correlation
Hence, the average mathematical confidence is a fair indication of is a promising idea and more research should be performed. If
the robustness of the method, and a higher confidence corresponds the theoretical explanation above is correct, one would expect an
to a more robust algorithm. analysis of the 3rd or 5th harmonic level to be even more effective
A theoretical reason for this is perhaps because the LV voltage than voltage THD.
THD level depends on what loads are present. In particular, In terms of computation time, Table 3 shows the results for a
number of sites in this regard on an Intel® Core™ i5-4300U
Table 2 Identification accuracy and confidence
central processing unit at 1.90 GHz. The programme is written in

Metric Voltage Fundamental

THD voltage

customer phase/transformer correctly 46/46 45/46

identified
average mathematical confidence 92.47% 84.65%

Table 3 Computational speed of the FZTCA algorithm

Number of ICPs present Number of records Time, s

46 88,366 0.18
291 419,331 1.37
582 419,331 12.49
582 838,662 14.19
Fig. 11 Computation time against number of ICPs

J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101 This is an open access article published by the IET under the Creative Commons
doi: 10.1049/joe.2016.0033 Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/)
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Fig. 12 Estimated impedances from each ICP to the transformer

MATLAB and run on version R2014b. The results show that the cost of a manual survey and desktop calculations for line and cable
FZTCA algorithm computation time is reasonable, especially as a parameters.
more powerful computer could easily be obtained. There is also Branch impedance values were calculated according to the
plenty of scope for optimisation, and the FZTCA algorithm method in Section 3.4: however, these are too numerous to
would be faster in a language such as C++ or C#. present here. In a few instances, it was possible to verify that
The computation time depends primarily on the number of ICPs these branch impedances matched well, but further work is required
in the data, though it also is affected by how many iterations are in this area.
needed to converge to the ideal ICP groupings. The relationship
is approximately time ∼O(N 3), where N is the number of ICPs in
5 Conclusions
the data. Although this would not scale well for very large
numbers of ICPs, by processing small geographical regions in This paper has presented work on mapping out the topology of the LV
turn it is computationally feasible to verify the entire utility data- network using smart-meter data. It is clear that having access for all the
base. Fig. 11 shows the relationship, where the data points are smart-meter data rather than a subset (those connected to one advanced
obtained by varying the number of ICPs included in the analysis. metering infrastructure) would greatly enhance the robustness and ac-
curacy of the estimation. However, it is possible to derive useful
results with limited data. The main findings from this paper are:
4.2 LV network impedance
(i) If voltage THD data, or other harmonic data, is available, this
The ICP to transformer impedance results are shown in Fig. 12, is likely to be much more reliable than the voltage magnitude
where the numbers inside the circles are the impedance from the correlation.
ICP to the transformer. (ii) The FZTCA algorithm has been developed to look for data-
There are a few suspect impedances, but the overall pattern of base inaccuracies, using a correlation analysis with the
low impedance near the transformer and high impedance at the Fisher Z transform. As found in previous research [7–15],
end of the feeder may be seen. Furthermore, the magnitude of the voltage correlation is a promising technique for ICP/trans-
result is very reasonable, and comparable with the actual conductor former mapping. However, it is not 100% infallible and a
impedance. In theory the method described could be used to acquire measure of conservatism is necessary. ICPs close to a trans-
an accurate estimate of the impedance. GIS data coupled with the former present a particular challenge, as do low coverage
results of this analysis (to fill in unknown parameters) would be ratios (due to not all smart-meter data being available).
the most obvious way to construct a complete LV network model (iii) A simple method to estimate network impedances, based on a
as it will match the actual system accurately and save the enormous least-square fit to the power against voltage relation for ICP to

This is an open access article published by the IET under the Creative Commons J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101
Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/) doi: 10.1049/joe.2016.0033
 20513305, 2016, 5, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/joe.2016.0033 by CochraneArgentina, Wiley Online Library on [31/10/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
transformer impedance and on maximising of correlation coef- Innovative Smart Grid Technologies Europe (ISGT Europe), Berlin,
ficients for branch impedances, has been presented. Further 2012
work in this area is necessary for the utility to make use of [10] Pezeshki H., Wolfs P.: ‘Correlation based method for phase identifi-
cation in a three phase LV distribution network’. 22nd Australasian
smart-meter data in this way.
Universities Power Engineering Conf. (AUPEC), Bali, 2012
[11] Luan W., Peng P., Maras M., ET AL.: ‘Smart meter data analytics for
6 Acknowledgments distribution network connectivity verification’, IEEE Trans. Smart
Grid, 2015, 6, (4), pp. 1964–1971
The authors acknowledge the input of Dr. Alan Wood and Dr. Allan [12] Luan W., Peng J., Maras M., ET AL.: ‘Distribution network topology
Miller, both of the University of Canterbury. In addition, the error correction using smart meter data analytics’. IEEE Power and
authors thank the members of the Asset Management team at Energy Society General Meeting 2013, 21–25 July 2013, pp. 1–5
Vector Limited for their helpfulness and input during the project. [13] Seal B.K., McGranahan M.: ‘Automatic identification of service
phase for electric utility customers’. IEEE Power and Energy
7 References Society General Meeting 2010, pp. 1–3
[14] Wen M., Arghandeh R., von Meier A., ET AL.: ‘Phase identification in
distribution networks with micro-synchrophasors’. IEEE Power &
[1] Campbell M., Watson N.R., Miller A.: ‘Smart meters to monitor Energy Society General Meeting 2015, pp. 1–5
power quality at consumer premises’. EEA Conf. & Exhibition, [15] Bolognani S., Schenato L.: ‘Identification of power distribution
Wellington, NZ, 24–26 June 2015 network topology via voltage correlation analysis’. 52nd IEEE
[2] Luan W., Sharp D., Lancashire S.: ‘Smart grid communication cap- Conf. on Decision and Control, pp. 1659–1664
acity planning for power utilities’. Proc. IEEE Power Energy
[16] Khumchoo K.Y., Kongprawechnon W.: ‘Cluster analysis for primary
Society Transmission Distribution Conf. Exposition, New Orleans,
feeder identification using metering data’. Sixth Int. Conf. of
LA, USA, April 2010, pp. 1–4
Information and Communication Technology for Embedded
[3] Jokar P., Arianpoo N., Leung V.C.M.: ‘Electricity theft detection in
Systems (IC-ICTES), 2015, pp. 1–6
AMI using customers’ consumption patterns’, IEEE Trans. Smart
[17] Peppanen J., Reno M.J., Thakkar M., ET AL.: ‘Leveraging AMI data
Grid, 2016, 7, (1), pp. 216–226
[4] Bandim C.J., Alves E. R., Pinto A.V., ET AL.: ‘Identification of energy for distribution system model calibration and situational awareness’,
theft and tampered meters using a central observer meter: a mathem- IEEE Trans. Smart Grid, 2015, 6, (4), pp. 2050–2059
atical approach’. Proc. IEEE PES Transmission Distribution Conf. [18] Peppanen J., Grinaldo J., Reno R.J., ET AL.: ‘Increasing distribution
Exposition, 2003 system model accuracy with extensive deployment of smart
[5] Depuru S.S., Wang L., Devabhaktuni V.: ‘Support vector machine meters’. Proc. IEEE Power Energy Society General Meeting,
based data classification for detection of electricity theft’. Proc. Washington, DC, USA, July 2014, pp. 1–5
IEEE/PES Power Systems Conf. Exposition (PSCE), 2011 [19] Wang B., Luan W.: ‘Generate distribution circuit model using AMI
[6] Da Silva P.G., Iliæ D., Karnouskos S.: ‘The impact of smart grid pro- data’. 2014 China Int. Conf. on Electricity Distribution (CICED),
sumer grouping on forecasting accuracy and its benefits for local elec- pp. 1251–1255
tricity market trading’, IEEE Trans. Smart Grid, 2014, 5, (1), pp. [20] Berrisford A.J.: ‘A tale of two transformers: an algorithm for estimat-
402–410 ing distribution secondary electric parameters using smart meter
[7] Short T.: ‘Advanced metering for phase identification, transformer data’. Proc. 26th IEEE Canadian Conf. Electrical and Computer
identification, and secondary modeling’, IEEE Trans. Smart Grid, Engineering (CCECE), Regina, SK, Canada, May 2013, pp. 1–6
2013, 4, (2), pp. 651–658 [21] Blanchet G., Charbit M.: ‘Digital signal and image processing using
[8] Arya V., Mitra R.: ‘Voltage-based clustering to identify connectivity MATLAB’ (John Wiley and Sons, Hoboken, NJ, USA, 2015, 2nd
relationships in distribution networks’. IEEE SmartGridComm Symp. edn.), p. 244
– Architectures and Models for the Smart Grid, 2013 [22] Carbonell F., Worsley K.J., Trujillo-Baretto N.: ‘On the Fisher’s Z
[9] Pezeshki H., Wolfs P.: ‘Consumer phase identification in a three transformation of correlation random fields’, Stat.: Probab. Lett.,
phase unbalanced LV distribution network’. Third IEEE PES 2009, 79, (6), pp. 780–788

J Eng, 2016, Vol. 2016, Iss. 5, pp. 94–101 This is an open access article published by the IET under the Creative Commons
doi: 10.1049/joe.2016.0033 Attribution-NoDerivs License (http://creativecommons.org/licenses/by-nd/3.0/)