17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022].

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
IET Generation, Transmission & Distribution

Special Section: Next Generation of Synchrophasor-based Power

System Monitoring, Operation and Control

Topology detection in power distribution ISSN 1751-8687

Received on 9th January 2020

system using kernel-node-map deep networks

Revised 24th March 2020
Accepted on 28th April 2020
E-First on 22nd July 2020
doi: 10.1049/iet-gtd.2020.0048
www.ietdl.org

Mengmeng Xiao1, Shaorong Wang1 , Zia Ullah1, Yan Li1, Reza Arghandeh2
1State Key Laboratory of Advanced Electromagnetic Engineering and Technology, Huazhong University of Science and Technology, Wuhan

430074, People's Republic of China

2Department of Computer Science, Electrical Engineering, and Mathematical Sciences, Western Norway University of Applied Science, Bergen

Area, 7030, Norway

E-mail: [email protected]

Abstract: This study is focused on real-time topology detection (TD) problems in the power distribution system. The advent of
distribution phasor measurement units offers additional opportunities to use deep learning methods for accurate TD of the
distribution system. In this study, a new concept named the kernel-node-map is presented, and then a novel topology detection
method, Kernel-Node-Map Deep Network (KNDN), for distribution system is developed. KNDN is based on a deep convolutional
neural network and the kernel-node-map concept. This neural network is adaptive to the physical topology of the distribution
system in the structure. The principle, training process, and application of KNDN are elaborated in detail. Testing results on the
modified IEEE 33-bus and 123-bus distribution systems verify the effectiveness and robustness of the proposed KNDN
methodology, and show its advantages compared with state-of-the-art methods.

1 Introduction group of studies, the energy conversion-based topology

identification is developed using only smart meter data [22].
Topology detection (TD) plays an important role in power system However, the hourly collected data may not be fast enough for TD
operation and control. In transmission networks, the TD methods in systems due to load-generation dynamics.
are more mature due to the data availability of the protection Recently the deep learning (DL) approach achieved high
system and SCADA system. However, the power distribution popularity in data science due to enhanced performance and
system is lagging behind transmission network in terms of scalability. In the subject of the power system, DL methods have
monitoring system coverage and data availability. Generally, been extensively used for the detection of false data injection
distribution system faces more topology changes comparing to the attacks [23–25], load forecasting [26–28], load modelling [29, 30].
transmission networks due to higher load dynamics, more fault The applied methods offer fast convergence on the issues of high
occurrence, and recently high penetration of distributed energy computational complexity.
resources (DERs). Moreover, the volt–var optimisation efforts can In this paper, we propose a novel TD algorithm, which is based
also increase the topology changes in distribution system [1]. on DL and D-PMU specially designed for use in the distribution
Furthermore, distribution phasor measurement unit (D-PMU) system. The new idea in the proposed method is to formulate the
with standardised technologies are proven sources to detect TD problem as a mapping process from the measurement data
accurately the network topology. As reported in [2] that the collection to the feasible topology set by convolutional neural
proposal of D-PMU combined with the advanced applications for networks (CNNs). The reason selected CNN is the requirement of
active distribution systems are in ascendant. Deployment of smart sensing local features to solve TD issues. Based on the above new
meters, D-PMU and advanced power electronic inverters facilitate idea we further develop Kernel-node-map Deep Network (KNDN)
grid operation functionalities such as TD applications. algorithm.
The advantages of high resolution, time-synchronised and The main contributions of this work are:
accurate D-PMU measurement data for distribution network
operators have been proven in different applications such as state • Utilise a feasibility test on grouping and enumerating topologies
estimation [3–5], fault detection [6–8], volt–var control [9, 10], and from N switches to reduce the number of possible topologies from
TD [1, 11–13]. For a comprehensive overview of the D-PMU data
application in the distribution network, readers can refer to the 2N to a smaller number. It consequently reduces computational
report by the North American SynchroPhasor Initiative [14–16] time.
review papers. • Propose the ‘kernel-node-map’ concept as a new feature of input
The classic TD methods are mainly based on the weighted data to feed the DL structure for the TD algorithm in the
least-squares state estimation [17]. In these methods, switch status distribution system.
is considered as one of the variables of the state estimator.
However, this approach is over-dependent on the accuracy of the The rest of this paper is organised as follows: Grid modelling
state estimator. preliminaries and problem formulation are described in Section 2.
A number of data-driven approaches are reported in the Section 3 offers the concrete structure and modified CNN
literature, such as graph theory and graphical modelling [18], algorithms based on the ‘Kernel-node-map’ concept, including the
probabilistic graphical models [19, 20]. In [1], a collection of all feature extraction layers and the multi-label classification layer.
possible topologies compared with the real-time time-series Section 4 presents the proposed model validation on modified
signatures of voltage measurements is presented, but it may be IEEE 33-bus and 123-bus distribution systems. The results
argued that line parameters are not available all the time. A obtained by the proposed KNDN TD algorithm are presented and
recursive grouping algorithm is proposed in [21] for topology discussed. Finally, Section 5 summarises the main contributions
verification based on the estimation of line parameters. In another and conclusions.

IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041 4033
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
magnitude Imag and phase angle I ph of current phasor I, active and
reactive power p + jq). Thus, the TD process can be described as a
mapping function as follows:

T = f (V, I, p, q) (3)

3 Methodology
The overview of KNDN methodology is shown in Fig. 2. The
process of the KNDN algorithm starts with calculating and
Fig. 1 Single-line diagram of a simple distribution system
obtaining feasible topologies of the objective distribution system.
The next step is to down-select the synthetic measurement data set
to fulfil requirement of identifying the feasible topologies. As the
following step, the selected inputs are rearranged based on ‘kernel-
node-map’ concept. Finally, rearranged input data is being sent to
KNDN network for model training and predicting. The output label
of the KNDN algorithm is in the presence of 0-1 arrays for feasible
topologies.
In this paper, the distribution system model is constructed in
Matlab® 2014b. The KNDN is developed and trained using
TensorFlow and Keras libraries. The proposed KNDN
methodology is elaborated in the following subsections.

3.1 Data preparation and generation

3.1.1 Feasible topologies: The total number of topologies of the
distribution system is growing exponentially with the number of
switches. For instance, if the number of switches is m, and the
number of all the topologies denoted by T total equals 2m. It is
worthy to notice that 2m for m switches is all possible combination
of switches status mathematically. However, practically in an
Fig. 2 Block diagram of KNDN methodology for TD in power distribution
electricity network, it is not preferred to open or close all switches
system
due to the connectivity requirements in the grid. Accordingly, the
TD problem is reduced to feasible topologies T f , which is defined
as radial and physically connected configuration in [31]. The
2 Problem formulation
algorithm of counting the feasible topologies of the distribution
The graph of a distribution system is represented as G = {Vℰ}, system is introduced in [32].
where V denotes the set of all vertices and ℰ represents the set of Generally, feasible topologies are based on graph properties,
all operational edges. The set of switch lines ℒs is a subset of ℰ. and then we introduce the definitions of degree, T-joint node and
Switch status s is labeled as 0 (open) and 1 (closed). The topology branch multiplicity factor, which are described as follows:
of the distribution system can be determined via switches status.  
And then, different 0-1 arrays of switch status yeild multiple Definition 1: The degree of a vertex is the number of edges
topologies T = { T1, T2, T3, …} of the distribution system. connected to it in an undirected graph.
Since historical data is limited to record all feasible topologies  
of the distribution system, we simulate all feasible topologies and Definition 2: A T-joint node has at least three degrees in an
generate synthetic measurement data by simulation software. For undirected graph.
each topology, there are multiple corresponding measurement sets  
M = {ℳ1, ℳ2, ℳ3, … }. The goal is to identify the real-time Definition 3: Multiplicity factor of a path equals to the number
topology of the distribution system with high load or DER of series branches in the path between two nodes.
dynamics. In the following example, we use a simple distribution The algorithm of counting the feasible topologies of the
system to better explain the TD task, as illustrated in Fig. 1. distribution system is inspired by the problem reduction procedure
where R is the resource of the circuit system, S1, S2, S3, S4, S5 and in [32]. By merging two-degree nodes together, the reduced
S6 are six switches in the circuit system. IL1, IL2, IL3 and IL4 are load network model is formed with only T-joint nodes remained. Using
currents. It is worth mentioning that power loads in Fig. 1 represent a set of vectors v=[v1 v2 v3 …], each of them contains a tree.
the aggregate loads with high dynamics. The distribution lines with Initially, v is empty. Then, the supply node (bottom position) is
switch are ℒs = {lci, l jb, ldk, lme, lgl, lnh}. Any topology T of this loaded as the first component v = [v1]. Constrained Conditions for
distribution system can be described as a vector of switch status finding the new component v j + 1 to the node v j are listed as follows:
{s1, s2, s3, s4, s5, s6}. For example, if S1, S2 and S4 are open while
others are closed, T = [1, 1, 0, 1, 0, 0]. • v j + 1 must be searched from the set of ordered sequential nodes
For the TD task, the status of each switch must need to be (bottom to top);
identified. It is supposed that switch is open if current phasor • there is no loop formed if v j + 1 is added;
flowing through the switch and voltage phasors at two ends of the • the same directed path (v j + 1 → v j ≠ v j + 1 ← v j) can not be
switch meet the following requirements: repeated in the same path set ℙ.
I12 = 0 (1) The backtracking algorithm in [32] is introduced to enumerate
feasible topologies of distribution systems, which is summarised in
V1 ≠ V2 (2) Fig. 3.
Finally, the total number of feasible topologies T f of the
Thus, the network topology T is identifiable if switch current original network can be generated by relating branch multiplicity
phasors ℐs = {Ici, I jb, Idk, Ime, Igl, Inh} and the two terminal voltage factor (see in Definition 3) between two T-joint nodes in the
phasors can be estimated according to the measurements (i.e. original network to the equivalent path of the reduced network
magnitude V mag and phase angle V ph of voltage phasor V, model.

4034 IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
is ±0.010° %rdg (1 standard deviation) and ±0.003° %FS while
the amplitude accuracy is ±0.050 %rdg and ±0.01 %FS.
Therefore, Gaussian noises with standard deviation of α = 0.025%
are added to D-PMU measurements.

3.2 Kernel-node-map deep networks for topology detection

problem
According to the physical law, relationships among voltage and
current phasors, real and reactive power flows are of high non-
linearity. It is a challenging task to solve all those complex and
non-linear equations, especially in a distribution system with
missing parameters. To address the above challenge, DL methods
are employed for TD. In the distribution system, CNN has been
explored extensively because of its excellent local sense ability and
employed in fault location [37], switch action identification [38]
and disturbance classification [39] so far. Thus, considering the
locally connected structure of the distribution system, CNN is
selected as a base neural network to implement the mapping
process from measurement data to switches status.
However, due to the sparse and various structure of distribution
system, weights sharing may hamper CNN's application for
topology identification. Unlike the application to image pixels,
where a shared weight matrix can extract feature by sliding the
filter, a specific CNN kernel can not be generalised to all inputs
relating to some topology. Thus, one CNN kernel only works for
one kernel-sized local input rather than all the inputs. It can be
Fig. 3 Algorithm 1: Backtracking algorithm for enumerating feasible argued that the modified CNN structure is required to overcome the
topologies of distribution systems above issues; thus, we proposed KNDN methodology based on
‘kernel-node-map’ concept.

3.2.1 ‘Kernel-node-map’ concept: To improve feature extraction

with a physical base, a novel concept ‘kernel-node-map’ is
proposed. One set of kernels is matched with several subgraphs
containing considered nodes. The scheme of ‘kernel-node-map’
concept is shown in Fig. 4.
As illustrated in Fig. 4, the proposed ‘kernel-node-map’ concept
relates CNN kernel structure {Φ1, Φ2, Φ3, Φ4, Φ5} and subgraphs
{G1, G2, G3, G4, G5} of the distribution system. As following,
‘kernel-node-map’ concept is explained further using an example.
Fig. 4 Scheme of ‘kernel-node-map’ concept  
Example 1: In Fig. 5a, F1 and F2 are two passive closed surfaces
without loads, and relate to the topology configuration yielding by
the status of switches S1 and S2, respectively.
There are four current phasors {I1, I2, IL1, IL2} injected into the
passive closed surface F2.
{V 1, V 2, V 3, V 4, V 5} are voltage phasors while P( ± ) is the
direction of power flow. The input vector x= [I1, I2, IL1, IL2, V 3, V 4,
V 5, P( ± )], and the output ys represents two switches status of s1 and
s2. As indicated in Fig. 5b, the output is expressed as the following
equation:

ys = W 3 × (W 2 × (W 1 × x + b1) + b2) + b3 (4)

where W 1 is one of the CNN kernels matched with the physical

characteristics of nodes contained in F1 and F2.

3.2.2 Input selection analysis: In principle, we can determine the

switch status by estimating the switch current. However, it happens
Fig. 5 Example subgraph with corresponding NN that the injected load current (i.e. light load) is too small to
(a) Example subgraph with considered nodes, (b) Input and output accurately indicate the switch status. To address this issue, power
flow and voltage are considered to be combined with current
3.1.2 Loading conditions: To simulate scenarios with high measurement for TD. However, if power loss of the passive closed
dynamic behaviour, the load profile consists of fast-changing surface is comparable with the load power (i.e. light load) injected
industrial customers, commercial demands including intermittent to the end node of the switch, it will be difficult to identify the
loads, and slowly time-varying residential building data [33] with switch status using energy conversion. Even though, the direction
Gaussian random variables (standard deviation = 0.15) derived of power flow still plays an important role in TD since it can reveal
from its nominal values [34, 35]. the energy path of the distribution system. Voltage-correlation
based analysis provides another solution for TD [18–21, 33]. The
3.1.3 Measurement errors: To mimic measurement errors, literature presented that there exists a correlation relationship
Gaussian noises are added to the measurements. According to the among neighbouring nodes.
standards of micro-PMU as described in [36], the angle of accuracy

IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041 4035
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
2D input of conv1st be X1, then X2 is reformed by rearranging the
elements of X1.

C = X2 ⊗ W + b (5)

where the output (feature map) of conv 1st is C, which is generated

by the convolution operator ⊗ with the entries of X2 and kernels
W ∈ Rk1, k2, b is the bias. conv1st and conv2nd represent the first
and second convolutional layers, respectively.
In general CNN structure, fully connected (FC) layer is usually
Fig. 6 Input element rearrangement based on ‘kernel-node-map’ concept adopted at the end of the neural network if the outputs are
functions of all inputs. However, there is no FC layer in the
proposed KNDN network structure since switch currents can be
estimated only using local measurements. The output layer
indicates the predicted multi-label switches status, which can
reveal the detected topology T of the distribution network.
It is worth mentioning that the proposed KNDN method based
on ‘kernel-node-map’ concept is not specific for D-PMU data but
also works with other types of data, such as SCADA and smart
meter data.

4 Numerical simulations
Fig. 7 Overall structure of KNDN algorithm The proposed KNDN algorithm is demonstrated on the modified
IEEE 33-bus and IEEE 123-bus distribution networks. All cases
are conducted on Matlab® 2014b on a Core i5-8250U CPU @1.6 
GHz.

4.1 IEEE 33-bus model analysis and results discussions

4.1.1 Structure of IEEE 33-bus distribution system: IEEE 33-
bus distribution system has N = 33 nodes, N L = 33 loads and
B = 37 branches in baseline topology [41]. The modified single-
line diagram for IEEE 33-bus distribution system is illustrated in
Fig. 8.
Fig. 8 Single-line diagram of modified IEEE 33-bus test feeder Note that load at node 7 (L7) is modified to be a light load. S1,
S3, S6 and S9 are four additional switches to the original system. G1,
In summary, the input of KNDN network structure should be G2, G3, G4, G5, G6, G7, G8 and G9 are nine subgraphs relating to
selected appropriately that the topology distribution system can be switches S1, S2, S3, S4, S5, S6, S7, S8 and S9, respectively. According to
successfully detected. For instance, as indicated in Fig. 5a,
the feasible enumerating algorithm in [32], there are 48 feasible
representing L2 as a light load, S1 cannot be accurately estimated
topologies in the modified IEEE 33-bus distribution system.
only using KCL or energy conservation at the closed surface F1
based on the aforementioned analysis. Thus, V 3, V 4, V 5 and P( ± ) are
4.1.2 D-PMUs' location and input analysis: The metered buses
selected as combined input with current for TD.
are marked with red rectangles in Fig. 8. Loading data profile (e.g.
active power flow, power factor, current and voltage magnitude) is
3.2.3 KNDN network structure: Considering the structural assumed to be available and time-aligned with D-PMU data. After
sparsity of the distribution system, input elements' locations are analysing the structure parameter of the IEEE 33-bus distribution
rearranged to better feed CNN for feature extraction rather than system, it has been found that current phase angles have a slight
using the ‘blackbox’ function directly. Notes that the input influence on current law's application. Thus, in principle, switches
elements related to one passive closed surface should be arranged status can be identified if the currents relating to its corresponding
together and contained in one kernel-sized input area. Elements closed surface are available. Nevertheless, based on the analysis in
location rearrangement based on ‘kernel-node-map’ concept is Section 3.2.2, the status of switch S3 and S9 cannot be accurately
illustrated in Fig. 6. estimated using current's law or energy conservation since L7 is a
‘Kernel-node-map’ concept-based deep network (KNDN) is light load. Thus, voltage magnitudes at nodes 6, 7 and 8 are
proposed as a modified CNN methodology for TD. The overall exploited to improve the detection accuracy of S3 and S9.
architecture of the KNDN algorithm is depicted in Fig. 7, which
includes input layers, convolutional (Conv) layers and output layer.
There is no pooling layer after each convolutional layer which 4.1.3 ‘Kernel-node-map’ for IEEE 33-bus system: Elements'
provides the opportunity to remain the physical and topological locations of input array are changed based on the proposed ‘kernel-
characteristics of input data. node-map’ concept. Assuming that the kernel size of KNDN's first
It is standard practice to normalise and pre-process the selected convolutional layer is 4 × 4, the relationship between kernel-sized
inputs before applying them to the neural network. One- inputs and subgraphs involving corresponding switched in the
dimensional (1D) array is normalised and reshaped to 2D array. modified IEEE 33-bus distribution system is illustrated as Fig. 9.
Then the input element locations are rearranged based on the It is supposed that purple kernel is specific for relating the
‘kernel-node-map’ concept. Rectified Linear Unit function [40] is physical characteristics of the first kernel-sized input elements in
used as the activation function for all the convolutional layers. The G2, G3, G6 and G9 to status of switches S2, S3, S6 and S9. Other
kernel size of the first convolutional layer (k1 by k2) depends on the switches' status is detected in the same manner based on the
maximum number maxn of current phasors needed for the detection proposed ‘kernel-node-map’ concept. The relationship of kernel-
of related switch status in the distribution system. For example, if sized input mapped to switches through CNN kernels for the whole
maxn = 4, the kernel size is 4 × 2 at least, including four pairs of distribution system is illustrated in Fig. 10.
As indicated in Fig. 10, selected input is designed as a 5 × 6
magnitudes and phase angles of current phasors. Let the original
matrix including 25 current magnitudes and 3 voltage magnitudes.
The block with forward slash (/) will be filled with ‘0’ when

4036 IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
S9, S5, S6}, {S5, S6, S7}, {S1, S2, S8, S9}, {S1, S4}, and {S1, S4, S7},
respectively.

4.1.4 Results & discussions in IEEE 33-bus model: In this

section, the performance of the proposed KNDN methodology is
evaluated and compared with conventional CNN network structure,
MLPs (multilayer perceptrons) and SVMs (support vector
machines) approaches. The proposed KNDN approach and other
referred methods applied for TD are tested on the Gaussian noisy
measurements. Results are evaluated using mean absolute error
(MAE) and root mean square error (RMSE), which are expressed
Fig. 9 Relationship between KNDN network and distribution system based as follows:
on ‘kernel-node-map’ concept
N
1
(6)
N n∑
MAE = (yt − yp)
=1

N
1
(7)
N n∑
RMSE = (yt − yp)2
=1

where yt is the labelled data of switch status, yp is the prediction of

Fig. 10 Kernel-sized inputs with corresponding subgraphs based on
switch status, and N represents the total number of sampling
‘kernel-node-map’ concept
datasets. The neural networks are trained using Adam optimiser
with the default hyper parameters (β1 = 0.9, β2 = 0.999, and
Table 1 Neural network parameters for IEEE 33-bus
distribution system ϵ = 10−8) [42] and early stopping monitor.
Items MLP CNN KNDN There are 2000 sampling datasets spitted to 1500 training
samples and 500 testing samples. The parameters of different
Inputn 30 5×6 5×6
methods are depicted in Table 1 . The MAEs and RMSEs of
layer1_n 100 64 64 different methods under different conditions of load L7 (1.light load
kernel size1 – 4×4 4×4 2.average load 3.heavy load) are compared in Table 2.
layer2_n 35 32 32 Clearly, it can be seen from Table 2 that all the methods except
kernel size2 – 2×2 2×2 SVM can obtain good performance if L7 is an average load or
layer3_n – 16 16 heavy load. However, when there is a light load connected to the
kernel size3 – 1×1 1×1 switches, the accuracy of all the methods would be influenced. In
this case, the proposed KNDN obtained higher accuracy than other
lr 0.002 0.001 0.0012
methods.
lr-dec 2 × 10−5 4 × 10−5 5 × 10−5 For the following case studies, results analysis is mainly
batch 12 12 12 focused on the scenarios that L7 is light load while other loads are
mindelta 0.001 0.001 0.001 under different loading conditions. For instant, 99.44% switch
patience 25 20 20 status accuracy conceived by KNDN, which is 2.82, 0.53, and
layer1_n, layer2_n and layer3_n denote the number of neurons for 0.42% higher than of SVM, MLP and traditional CNN,
the first, second and third convolutional layers, respectively. lr and lr- respectively. For the overall TD, 96.6% accuracy obtained by
dec denote learning rate and its decay while mindelta and patience KNDN, which is 20, 2.4, and 1.4% higher than of SVM, MLP and
are parameters of early stopping monitor. traditional CNN, respectively. In addition, comparing the results
employing different inputs, it is indicated that using voltage
magnitudes along with currents can dramatically improve the
detection performance under the situation that switch is connected
Table 2 Results of different methods for switch status with a light load. Further, the detection accuracy relating to each
detection under different loading conditions of L7 (1.light load switch obtained by KNDN along with other methods on the testing
2.average load 3.heavy load) dataset with 500 samples are depicted Fig. 11.
Index Input SVM MLP CNN KNDN These facts confirmed that KNDN offers the best optimal
MAE(1) Im 0.1531 0.1044 0.1027 0.0971 results in all states which are conceiving and relevant. The high
Im + V (m) 0.0338 0.0109 0.0098 0.0056 detection accuracy of switches S1, S2, S4, S5, S6, S7 and S8
RMSE(1) Im 0.3152 0.3232 0.3204 0.3116
demonstrated that current magnitudes can approximately reveal the
underlying current's law in IEEE 33-bus distribution system where
Im + V (m) 0.1394 0.1044 0.0989 0.0746 the differences of current phases can be neglected for current law's
MAE(2) Im 0.1162 0.0002 0.0022 0.002 application to determine switch status. S3 and S9 are detected with
Im + V (m) 0.0309 0.0013 0.0011 0.0007 lower accuracy than other switches due to the light load L7.
RMSE(2) Im 0.2863 0.0149 0.0471 0.0422 Comparing the detection accuracy of S3 and S9 of different
methods, KNDN has shown to be competitive for feature
Im + V (m) 0.1359 0.0365 0.0333 0.0258
extraction and status identification of switch connected with a light
MAE(3) Im 0.1001 0.0024 0.0011 0.0016 load. The results verified the local sense ability of KNDN
Im + V (m) 0.0491 0.0029 0.0018 0.0004 algorithm based on ‘kernel-node-map’ concept.
RMSE(3) Im 0.2667 0.0494 0.0333 0.0394 The loss and accuracy curves obtained by the proposed KNDN
algorithm on the training and validating dataset of IEEE 33-test
Im + V (m) 0.1898 0.0537 0.0422 0.0211 feeders are presented in Fig. 12.
Where the train_loss and train_acc, valid_loss and valid_acc
denote the training and validating loss and accuracy for TD,
implementing the algorithm. The purple, blue, red, orange, green respectively. It is observed that KNDN converges at around 200
and grey frames are matched with subgraphs {G2, G3, G9, G6}, {G2, epochs, while CNN and MLP converge at 176 and 227 epochs,
G3, G9, G5, G6}, {G5, G6, G7}, {G1, G2, G8, G9}, {G1, G4} and {G1, respectively, which indicates that both KNDN and CNN converge
G4, G7}, which correspond to switches sets {S2, S3, S9, S6}, {S2, S3, faster than MLP.
IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041 4037
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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. 14 IEEE 33-bus test feeder with network changes
Fig. 11 Comparison of different methods for each switch
Table 3 Results of different methods under network
changes in IEEE 33-bus distribution system
Index Methods (1.a) (1.b) (2.a) (2.b)
MAE MLP 0.0102 0.0176 0.0142 0.0218
CNN 0.0089 0.0129 0.0111 0.0147
KNDN 0.0067 0.0113 0.0093 0.0069
RMSE MLP 0.1011 0.1325 0.1193 0.1476
CNN 0.0943 0.1135 0.1054 0.1211
KNDN 0.0816 0.1065 0.0966 0.0830

network changes with good or comparable performance. However,

if a new distribution line is added to the network, correspondingly,
Fig. 12 Loss and accuracy curves of KNDN methodology the network baseline topology is modified, which means the
original trained model is not adaptive for this change.
To study the impact of new loads and new generations on
detection accuracy, which are displayed in Fig. 14, another 500
testing samples are generated for each new scenario.
The corresponding new scenarios are listed as follows:

(1) Addition of new loads:

To study the impact of the addition of new loads to detection
accuracy, the loading conditions are modified as follows:

a. A new load, equivalent to L31, is added to node 31.

b. New equivalent loads are added to nodes 31,21 and 17.
(2) Addition of new generations:
To test the effectiveness of KNDN with the addition of the new
Fig. 13 KNDN's robustness to measurement errors in IEEE 33-bus test
generation to the distribution network, a solar power generation
feeders
with nominal power 100 kW is added to the IEEE 33-bus test
feeder.
Furthermore, regarding the time requirement of TD, we tested
the execution time of the proposed KNDN algorithm, which is
a. Added to node 25. The irradiance (W/m2) of photovoltaic
defined as TD time using the trained KNDN model, and found that
KNDN can achieve high accuracy at about 0.02 s, which can be (PV) Array is supposed to be a uniform random variable
used practically for operation and control in the distribution ranging from 200 to 1000 while the temperature is fixed at
system. 25°C.
b. Added to node 10. The PV array of new solar generation has
• Performance regarding measurement errors constant irradiance 1000 W/m2 and fixed temperature at
25°C.
To demonstrate the robustness of KNDN to measurement
errors, the standard deviation of Gaussian random noises added to The performances of KNDN on distribution network changes
input dataset is set to vary from 0.025 to 10%. The accuracy of with comparison to CNN and MLP are shown in Table 3.
KNDN algorithm for different measurement errors is shown in Fig. It can be observed that, without retraining the neural network
13. model, the accuracy obtained by KNDN algorithm is comparable
As standard deviation grows from 0.025 to 1%, which is a to that conceived in the original distribution network and still
reasonable range of measurement errors, KNDN algorithm remains outperforms other methods. It is indicated that KNDN algorithm
high switches status detection accuracy. Thus, it can be concluded has the adaptability to some proper network changes. However, if
that KNDN algorithm is robust to measurement errors. Besides, the the network changes a lot compared to the original network, the
situations with measurement errors being 2 to 10% are simulated to neural network still needs retraining for the new distribution
verify the correctness of adding Gaussian noises. network.

• Performance regarding distribution network changes 4.2 IEEE 123-bus model analysis and results discussions

Distribution networks change more frequently than 4.2.1 Structure of IEEE 123-bus distribution system: To
transmission networks [38]. Therefore, it is important to consider validate the proposed KNDN algorithm further, modified IEEE
either the future expansions of distribution network such as the 123-bus test feeder is deployed in this subsection. IEEE 123-bus
addition of new generators, new distribution lines, new loads etc. It feeder is a three-phase unbalanced distribution system. To focus on
is expected that the trained KNDN model can adapt to future more interesting scenarios with various feasible configurations, on-

4038 IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
S5, S6, S8 and S9 are four additional switches to the original system.
G1, G2, G3, G4, G5, G6, G7, G8 and G9 are nine subgraphs relating to
switches S1, S2, S3, S4, S5, S6, S7, S8 and S9, respectively. According to
the feasible enumerating algorithm in [32], there are totally 20
feasible topologies in the modified IEEE 123-bus distribution
system.

4.2.2 D-PMUs' location and input analysis: The metered buses

are marked with red rectangles in Fig. 15. Based on the analysis in
Section 3.2.2, the status of switches S3 and S6 are difficult to
identify only using current's law or energy conservation since G0 is
a subgraph without phase-A loads. Specifically, if the line between
node 108 and node 97 is at open state, it is difficult to distinguish
which of S5 or S6 is open or closed. Thus, voltage magnitudes at
nodes 67, 97, 197, 105 and 108 are exploited to improve the
detection accuracy of S3 and S6. Additionally, notes that S8 and S9
are two series switches between nodes 54 and 72. In the same
manner, the load conditions of L88 and L94 have an impact on the
identification process.
Fig. 15 Single-line diagram of IEEE 123-bus distribution system. Node To deal with this issue, the power flow direction P( ± ) between
150 is the slack node nodes 89 and 91, represented by 1 (89 → 91) and −1 (91 → 89), is
used to help determine where the power provided to loads L88 and
L94 comes from. The advantage of employing P( ± ) is that it can
indicate the energy path without being affected by various loading
conditions.

4.2.3 ‘Kernel-node-map’ for IEEE 123-bus system: In the

same manner with Section 4.1.3, switches status are detected based
Fig. 16 Kernel-sized inputs with corresponding subgraphs based on on the proposed ‘kernel-node-map’ concept. The relationship of
‘kernel-node-map’ concept in IEEE 123-bus distribution system kernel-sized input mapped to switches through CNN kernels for the
whole distribution system is illustrated in Fig. 16.
Table 4 Neural network parameters for IEEE 123-bus As indicated in Fig. 16, selected input elements are rearranged
distribution system as a 6 × 6 matrix including 29 current magnitudes Im, 6 voltage
Items MLP CNN KNDN magnitudes V m and power flow direction P( ± ). The purple, blue,
Inputn 36 6×6 6×6 red, orange, green and grey frames are matched with subgraphs
{G4}, {G1}, {G2, G5}, {G7, G8, G9}, {G8, G9} and {G3, G6}, which
layer1_n 100 64 64
correspond to switches sets {S4}, {S1}, {S2, S5}, {S7, S8, S9}, {S8, S9},
kernel size1 – 3×3 4×4
and {S3, S6}, respectively.
layer2_n 35 32 32
kernel size2 – 2×2 2×2
4.2.4 Results & discussions: There are 1000 sampling datasets
layer3_n – 16 16 acquired under different loading conditions and partioned into 750
kernel size3 – 1×1 1×1 training samples and 250 testing samples. The parameters of
lr 0.001 0.001 0.001 different methods are tuned in the same way as Section 4.1.4 and
lr-dec 2 × 10−5 2 × 10−5 4 × 10−5 reported in Table 4. The MAEs and RMSEs of different methods
are compared in Table 5.
batch 10 10 12
It indicates that the proposed KNDN methodology performs
mindelta 0.001 0.001 0.001 much better than SVM, MLP and traditional CNN methods.
patience 30 25 25 Comparing the results employing different inputs, it is indicated
the detection performance can be greatly improved using integrated
inputs of current magnitudes, voltage magnitudes and power flow
Table 5 Results of different methods for switch status direction. Specifically, the added inputs of voltage magnitudes
detection in IEEE 123-bus distribution system contribute to the identification of switches S3 and S6 while the
Index Input SVM MLP CNN KNDN introduction of power flow direction improves the detection
MAE Im 0.0928 0.0449 0.0444 0.0489 accuracy of S8 and S9. KNDN can achieve 99.96% switches status
Im + V m 0.0427 0.0076 0.0080 0.0049
detection accuracy and 99.6% overall TD accuracy when using
integrated inputs.
Im + V m + P( ± ) 0.0080 0.0036 0.0031 0.0004 The loss and accuracy curves obtained by the proposed KNDN
RMSE Im 0.2283 0.2119 0.2108 0.2211 algorithm on the training and validating dataset of IEEE 123-bus
Im + V m 0.1332 0.0869 0.0894 0.0699 test feeders are presented in Fig. 17.
It is observed that KNDN converges at 124 epochs, while CNN
Im + V m + P( ± ) 0.0462 0.0632 0.0558 0.0211
and MLP converge at 123 and 216 epochs, respectively, which
reveals that both KNDN and CNN converge much faster than MLP.
Furthermore, we tested the TD time using the trained KNDN
load tap changer between node 60 and node 160 is excluded. The model in the IEEE 123-bus distribution system, and found that the
modified IEEE 123-bus test feeder has nine operational switches. detection time is 0.045 s on average, which is potential and
The single-line diagram of the modified IEEE 123-bus test feeder promising for practical use.
is depicted in Fig. 15.
To simplify the analysis, only phase-A switches status are • Performance regarding measurement errors
considered. The modified IEEE 123-bus distribution system is
divided into nine regions, which are described as subgraphs based To validate the robustness of KNDN to measurement errors, the
on graph theory. Note that G0 is a subgraph without phase-A loads. standard deviation varies from 0.025 to 10%. Fig. 18 reports the

IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041 4039
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
standard deviation to verify the correctness of adding gaussian
noises to the measurements.

• Performance regarding distribution network changes

To validate the trained KNDN model's adaptability to future

network changes to IEEE 123-bus distribution systems, as depicted
in Fig. 19. For each kind of network change as follows, another
250 testing data samples are regenerated:

(1) Shifting the nodes 102, 103 and 104 from phase C to phase A.
(2) Adding new phase-A loads to nodes 99, 100 and 450.
(3) Adding new solar power generation with nominal power 100 
Fig. 17 Loss and accuracy curves of KNDN methodology kW to node 67. Irradiance is 1000 W/m2 and temperature is 25°C.

The switches status detection performances of KNDN on

distribution network changes with comparison to other methods are
shown in Table 6.
It can be observed that, without retraining the neural network
model, the detection accuracy obtained by KNDN algorithm has
shown overwhelming performance over other methods. To some
extent, it validated the robust local sensing and physical feature
extraction ability of KNDN algorithm. On the other hand, the
accuracy is compromising with these great changes added to the
original distribution network. In these cases, neural network
retraining is needed to obtain higher accuracy.

5 Conclusion
Fig. 18 KNDN's robustness to measurement errors in IEEE 123-bus test
feeders In this paper, we propose a novel data-driven approach, KNDN, to
detect the topology of the distribution system using D-PMU
measurements. The proposed approach relates CNN kernels to the
physical topology of the distribution system by implementing the
presented Kernel-Node-Map concept. We validate the proposed
algorithm on the modified IEEE 33-bus and 123-bus distribution
networks. The dataset used in this work are acquired in various
scenarios, and are partitioned into training, validation and test
subsets according to 50, 25 and 25%. The obtained results reveal
the proposed method can detect the topology with high accuracy in
a robust computational time under various operating scenarios with
measurement errors. Besides, it is indicated that KNDN is robust to
errors and comparison with state-of-the-art methods validates the
outperformance of KNDN, especially for the situation where light
load or no load is connected to the distribution line between two
series switches.
For future work, the authors will develop an optimal algorithm
to automatically rearrange input elements for more practical use of
KNDN in a large distribution system and test it on real field data if
Fig. 19 IEEE 123-bus test feeder with network changes possible.

Table 6 Results of different methods under distribution 6 Acknowledgments

network changes
Index Methods (1) (2) (3) The authors would like to thank the sponsorship by the National
Key Research and Development Program of China Grant
MAE SVM 0.0222 0.0324 0.0258 2017YFB0902800.
MLP 0.0204 0.0413 0.0533
CNN 0.0164 0.0164 0.0329
7 References
KNDN 0.0098 0.0107 0.0107
[1] Cavraro, G., Arghandeh, R.: ‘Power distribution network topology detection
RMSE SVM 0.0765 0.1047 0.0959 with time-series signature verification method’, IEEE Trans. Power Syst.,
MLP 0.1430 0.2033 0.2309 2018, 33, (4), pp. 3500–3509
[2] Liu, Y.K., Wu, L., Li, J.: ‘D-PMU based applications for emerging active
CNN 0.1282 0.1282 0.1814
distribution systems: a review’, Electr. Power Syst. Res., 2020, 179, pp.
KNDN 0.0989 0.1033 0.1033 106063
[3] Carta, D., Benigni, A., Muscas, C.: ‘Model order reduction for PMU-based
state estimation in distribution grids’, IEEE Syst. J., 2018, 12, (3), pp. 2711–
2720
MAEs of KNDN method with noisy measurements of different [4] Pegoraro, P.A., Meloni, A., Atzori, L., et al.: ‘PMU-based distribution system
levels for the IEEE 123-bus test feeder. state estimation with adaptive accuracy exploiting local decision metrics and
With the standard deviation ranging from 0.025 to 2%, switches IoT paradigm’, IEEE Trans. Instrum. Meas., 2017, 66, (4), pp. 704–717
[5] Macii, D., Barchi, G., Schenato, L.: ‘On the role of phasor measurement units
status detection accuracy conceived by KNDN algorithm remains for distribution system state estimation’. Proc. 2014 IEEE Workshop on
over 99.8% while the overall TD accuracy keeps over 99%. Thus, Environmental, Energy, and Structural Monitoring Systems Proc., Naples,
it is confirmed that KNDN is robust to measurement errors. Italy, September 2014, pp. 1–6
Additionally, we tested the proposed method with 5 and 10% [6] Cordova, J., Soto, C., Gilanifar, M., et.al.: ‘Shape preserving incremental
learning for power systems fault detection’, IEEE Control Syst. Lett., 2018, 3,
(1), pp. 85–90

4040 IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041
© The Institution of Engineering and Technology 2020
 17518695, 2020, 19, Downloaded from https://ietresearch.onlinelibrary.wiley.com/doi/10.1049/iet-gtd.2020.0048 by University Of Macau, Wiley Online Library on [04/11/2022]. 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
[7] Zhou, Y.X., Arghandeh, R., Zou, H., et al.: ‘Nonparametric event detection in [25] Wang, H.Z., Ruan, J.Q., Wang, G., et al.: ‘Deep learning-based interval state
multiple time series for power distribution networks’, IEEE Trans. Ind. estimation of AC smart grids against sparse cyber attacks’, IEEE Trans. Ind.
Electron., 2019, 66, (2), pp. 1619–1628 Inf., 2018, 14, (11), pp. 4766–4778
[8] Zhou, Y.X., Arghandeh, R., Spanos, C.J.: ‘Partial knowledge data-driven [26] Fan, C., Xiao, F., Zhao, Y.: ‘A short-term building cooling load prediction
event detection for power distribution networks’, IEEE Trans. Smart Grid, method using deep learning algorithms’, Appl. Energy, 2017, 195, pp. 222–
2018, 9, (5), pp. 5152–5162 233
[9] Zhang, J.Y., Lu, C., Fang, C., et al.: ‘Load shedding scheme with deep [27] Shi, H., Xu, M.H., Li, R.: ‘Deep learning for household load forecasting-A
reinforcement learning to improve short-term voltage stability’. Proc. 2018 novel pooling deep RNN’, IEEE Trans. Smart Grid, 2017, 9, (5), pp. 5271–
IEEE Innovative Smart Grid Technologies-Asia (ISGT Asia), Singapore, 5280
Singapore, May 2018, pp. 13–18 [28] Kong, W.C., Dong, Z.Y, Jia, Y.W., et al.: ‘Short-term residential load
[10] Sankur, M.D., Dobbe, R., Von, M.A., et al.: ‘Model-free optimal voltage forecasting based on LSTM recurrent neural network’, IEEE Trans. Smart
phasor regulation in unbalanced distribution systems’, IEEE Trans. Smart Grid, 2017, 10, (1), pp. 841–851
Grid, 2020, 11, (1), pp. 884–894 [29] Zheng, C., Wang, S.H, Liu, Y.L., et al.: ‘A novel equivalent model of active
[11] Cavraro, G., Arghandeh, R., Poolla, K., et al.: ‘Data-driven approach for distribution networks based on LSTM’, IEEE Trans. Neural Netw. Learn.
distribution network topology detection’. Proc. 2015 IEEE Power & Energy Syst., 2019, 30, (9), pp. 2611–2624
Society General Meeting, Denver, CO, USA, July 2015, pp. 1–5 [30] Zheng, C., Wang, S.H, Liu, Y.L., et al.: ‘A novel RNN based load modelling
[12] Cavraro, G., Arghandeh, R., Barchi, G., et al.: ‘Distribution network topology method with measurement data in active distribution system’, Electr. Power
detection with time-series measurements’. Proc. 2015 IEEE Power & Energy Syst. Res., 2019, 166, pp. 112–124
Society Innovative Smart Grid Technologies Conf. (ISGT), Washington, DC, [31] Cavraro, G., Bernstein, A., Kekatos, V., et al.: ‘Real-time identifiability of
USA, February 2015, pp. 1–5 power distribution network topologies with limited monitoring’, IEEE
[13] Arghandeh, R., Gahr, M., Von, M.A., et al.: ‘Topology detection in microgrids Control Syst. Lett., 2020, 4, (2), pp. 325–330
with micro-synchrophasors’. Proc. 2015 IEEE Power & Energy Society [32] Andrei, H., Chicco, G.: ‘Identification of the radial configurations extracted
General Meeting, Denver, CO, USA, July 2015, pp. 1–5 from the weakly meshed structures of electrical distribution systems’, IEEE
[14] Von, M.A., Arghandeh, R., Brady, K., et al.: ‘Synchrophasor monitoring for Trans. Circuits Syst. I: Regul. Pap., 2008, 55, (4), pp. 1149–1158
distribution systems: technical foundations and applications a white paper by [33] Bolognani, S., Bof, N., Michelotti, D., et al.: ‘Identification of power
the NASPI distribution task team’. North American SynchroPhasor Initiative distribution network topology via voltage correlation analysis’. Proc. 2013
(NASPI), Berkeley, CA, USA, 2018 IEEE 52nd Annual Conf. on Decision and Control (CDC), Florence, Italy,
[15] Von, M.A., Stewart, E., McEachern, A., et al.: ‘Precision micro- December 2013, pp. 1659–1664
synchrophasors for distribution systems: a summary of applications’, IEEE [34] Ahmad, A., Azmira, I., Ahmad, S., et al.: ‘Statistical distributions of load
Trans. Smart Grid, 2017, 8, (6), pp. 2926–2936 profiling data’. Proc. 2012 IEEE Int. Power Engineering and Optimization
[16] Von, M.A., Culler, D., McEachern, A., et al.: ‘Micro-synchrophasors for Conf., Melaka, Malaysia, June 2012, pp. 199–203
distribution systems’. Proc. 2014 IEEE Power & Energy Society Innovative [35] Chaturvedi, A., Prasad, K., Ranjan, R.: ‘Use of interval arithmetic to
Smart Grid Technologies Conf. (ISGT), Washington, DC, USA, February incorporate the uncertainty of load demand for radial distribution system
2014, pp. 1–5 analysis’, IEEE Trans. Power Deliv., 2006, 21, (2), pp. 1019–1021
[17] Hayes, B., Escalera, A., Prodanovic, M.: ‘Event-triggered topology [36] Power Standards Lab: ‘Micro-pmu data sheet’, 2017. Available at https://
identification for state estimation in active distribution networks’. Proc. 2016 www.powerstandards.com/download-center/micropmu/, accessed on 20th
IEEE PES Innovative Smart Grid Technologies Conf. Europe (ISGT-Europe), October 2019
Ljubljana, Slovenia, October 2016, pp. 1–6 [37] Alhalaseh, R., Kämmer, R., Nath, N.C., et al.: ‘Convolutional neural network
[18] Deka, D., Backhaus, S., Chertkov, M.: ‘Structure learning in power based fault location detector for power grids’. Proc. 2019 4th Int. Conf. on
distribution networks’, IEEE Trans. Control Netw. Syst., 2018, 5, (3), pp. Smart and Sustainable Technologies (SpliTech), Split, Croatia, Croatia, June
1061–1074 2016, pp. 1–5
[19] Weng, Y., Liao, Y.Z., Rajagopal, R.: ‘Distributed energy resources topology [38] Duan, N., Stewart, E.M.: ‘Deep-learning-based power distribution network
identification via graphical modeling’, IEEE Trans. Power Syst., 2017, 32, switch action identification leveraging dynamic features of distributed energy
(4), pp. 2682–2694 resources’, IET Gener. Trans. Distrib., 2019, 13, (14), pp. 3139–3147
[20] Liao, Y.Z., Weng, Y., Liu, G., et al.: ‘Urban mv and lv distribution grid [39] Cai, K., Cao, W., Aarniovuori, L., et al.: ‘Classification of power quality
topology estimation via group lasso’, IEEE Trans. Power Syst., 2019, 34, (1), disturbances using wigner-ville distribution and deep convolutional neural
pp. 12–27 networks’, IEEE Access, 2019, 7, pp. 119099–119109
[21] Cavraro, G., Kekatos, V., Veeramachaneni, S.: ‘Voltage analytics for power [40] Hara, K., Saito, D., Shouno, H.: ‘Analysis of function of rectified linear unit
distribution network topology verification’, IEEE Trans. Smart Grid, 2019, used in deep learning’. Proc. 2015 Int. Joint Conf. on Neural Networks
10, (1), pp. 1058–1067 (IJCNN), Killarney, Ireland, July 2015, pp. 1–8
[22] Pappu, S.J., Bhatt, N., Pasumarthy, R., et al.: ‘Identifying topology of low [41] Zhao, Y., Chen, J., Poor, H.V.: ‘Learning to infer power grid topologies:
voltage distribution networks based on smart meter data’, IEEE Trans. Smart performance and scalability’. Proc. 2018 IEEE Data Science Workshop
Grid, 2018, 9, (5), pp. 5113–5122 (DSW), Lausanne, Switzerland, June 2018, pp. 215–219
[23] He, Y.B., Mendis, G.J., Wei, J.: ‘Real-time detection of false data injection [42] Kingma, D.P., Ba, J.: ‘Adam: A method for stochastic optimization’. arXiv
attacks in smart grid: a deep learning-based intelligent mechanism’, IEEE preprint arXiv:1412.6980, 2014
Trans. Smart Grid, 2017, 8, (5), pp. 2505–2516
[24] Yu, J.J.Q., Hou, Y.H., Li, V.O.K.: ‘Online false data injection attack detection
with wavelet transform and deep neural networks’, IEEE Trans. Ind. Inf.,
2018, 7, (14), pp. 3271–3280

IET Gener. Transm. Distrib., 2020, Vol. 14 Iss. 19, pp. 4033-4041 4041
© The Institution of Engineering and Technology 2020