IET Generation, Transmission & Distribution

Special Issue: Intelligent Protection and Control of Microgrids with

Energy Storage Integration

Adaptive protection combined with machine ISSN 1751-8687

Received on 9th February 2018

learning for microgrids

Revised 30th August 2018
Accepted on 18th October 2018
E-First on 8th March 2019
doi: 10.1049/iet-gtd.2018.6230
www.ietdl.org

Hengwei Lin1,2, Kai Sun1, Zheng-Hua Tan3, Chengxi Liu2, Josep M. Guerrero2 , Juan C. Vasquez2
1Department of Electrical Engineering, Tsinghua University, Beijing, People's Republic of China
2Department of Energy Technology, Aalborg University, Aalborg 9220, Denmark
3Department of Electronic Systems, Aalborg University, Aalborg 9220, Denmark

E-mail: [email protected]

Abstract: This paper presents a rule-based adaptive protection scheme using machine-learning methodology for microgrids in
extensive distribution automation (DA). The uncertain elements in a microgrid are first analysed quantitatively by Pearson
correlation coefficients from data mining. Then, a so-called hybrid artificial neural network and support vector machine (ANN-
SVM) model is proposed for state recognition in microgrids, which utilises the growing massive data streams in smart grids.
Based on the state recognition in the algorithm, adaptive reconfigurations can be implemented with enhanced decision-making
to modify the protective settings and the network topology to ensure the reliability of the intelligent operation. The effectiveness
of the proposed methods is demonstrated on a microgrid model in Aalborg, Denmark and an IEEE 9 bus model, respectively.

1 Introduction automation (DA) due to lack of considering the new requirements

in smart grids and energy Internets.
The expansion of the renewable energy sources has been a concern Adaptive protection is another effective solution against the
in the recent years. The future electricity network will need to aforementioned challenges with more auxiliary functions.
accommodate large scale of distributed generation (DGs) units and According to [11], adaptive protection is ‘an on-line activity that
facilitate the connection of grand scale of centralised generation at modifies the preferred protective response to a change in system
suitable locations. Microgrid has been considered as an effective conditions or requirements, which is usually automatic but can
way to manage the DGs and other distributed energy resources include timely human intervention’. Modern digital relays for low-
(DERs) on the distribution system level and the user level [1, 2]. voltage applications usually have several settings groups, which
The typical microgrid has the same size as a few medium-/low- can be changed by communication or the relays themselves. In
voltage distribution feeders and may exceed a capacity of several [12], the authors proposed an adaptive and coordinate scheme
MVA with modest geographical span. between fuses and reclosers on a distribution system. There is a
The transmission networks have always been a management main relay located at the substation controlling the downstream
and balancing role in the electricity supply chain. However, breakers. In [13, 14], the adaptive differential protection schemes
distribution grids played passively. The current problem is that lots are proposed to coordinate the threshold for different faulty
of services previously provided in transmission grids are required situations. An adaptive current protection is presented for
to be performed in the distribution networks, such as power flow microgrids in [15], which uses event tables for each relay recorded
optimisation, contingency analysis, balancing, and constraint in the control centre. An adaptive overcurrent protection is
management. The large amount of intelligent building services in proposed without online communications in [16]. Here, all the
user level and the utilisation of local generation to support the local other relays' settings are calculated and recorded offline in the relay
network sometimes influence on the main grid. As one of the basic itself. This action might be unreliable for the protection to store
aspects of the infrastructure against the severe disturbances and and identify all the settings for all the other relays with several
faults, protection needs to be enhanced for the modernisation of operating modes and uncertainties in the system. Meanwhile, the
distribution system. There are four main problems required to be previous literatures of microgrid protection did not consider the
taken into account for protection with the increasing penetration of state recognition which decides the state and the way to modify the
DGs: (1) the randomness and intermittence feature of renewable protective settings. State estimation is a mathematical tool that
energy sources that bring variable infeed currents, (2) bidirectional deals with estimating the values of state parameters based on
power flow on one feeder, (3) different fault currents subjected to empirical or measured data which has a random component [17].
various operating modes (grid-connected mode and islanded mode Kalman filter is adopted online to estimate the states of DERs for
etc.), and (4) looped or meshed networks. Conventional microgrid protection [18]. An adaptive overcurrent protection with
distribution power system is radial in nature. However, part of the ANN state recognition model is proposed for microgrids in
system may no longer be radial when DGs connects to the network. distribution system [19]. The calculation of the overcurrent
It has been a fact that the protection devices should be directionally protective settings is treated as an optimisation problem, while the
sensitive in a multi-sources system [3–5]. In [6], a storage unit is system configuration can be updated and modified based on the
used to support and increase the short-circuit current of the state recognition. The state estimation algorithms are proposed
microgrid to make the fault detection possible from conventional with synchronised phasor measurements for protection system in
relays. A distance protection scheme with mho-based [20, 21]. For modern power system and future smart grid, the
characteristics is proposed for a microgrid in [7]. A distance collected huge amount of valuable data may become possible to be
protection with inverse time tripping characteristics is presented in utilised directly to obtain/train the mathematical model. This
[8, 9]. A combined methodology is proposed in [10] for microgrid tendency has been of concern as the electrical system keeps
protection based on differential protection and analysis of evolving towards an intelligent entity.
symmetrical components. Whereas the protection schemes may Machine learning is such a methodology that concerned with
play passively for the intelligent operation in distribution the pattern recognition and computational-learning theory in

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 770
© The Institution of Engineering and Technology 2019
Fig. 1 Test microgrid model in Aalborg

different models illustrates the effectiveness for electrical

application.
This paper is organised as follows. Section 2 describes the test
microgrid model. The data generating and processing are
introduced in detail in this section. The data mining results reveal
that the operating states influence the protection selectivity. In
Section 3, the time-domain simulation and the calculation of
protective settings are given. In Section 4, as a hybrid approach,
ANN-SVM model is integrated into the rule-based adaptive
protection scheme for microgrid so that the system can adapt to the
variable operating states with a ‘self-learning’ and ‘self-training’
capability. Finally, the conclusion is given in Section 5.

2 Data generating and processing

2.1 Test model
The proposed methodology requires database available beforehand,
which is obtained through simulation here due to the
confidentiality from the utilities. However, it does not lose the
Fig. 2 Configuration of the wind farm representativeness since simulation data are significant for power
planning and operation and other database in utilities can be
artificial intelligence (AI) [22–25]. It may be inspiring for considered in practical application. The first test microgrid model
protection and control to combine with machine learning to fulfil is a part of the distribution system owned by Himmerlands
the object of intelligent operation/control in smart grid, which may Elforsyning (HEF) in Aalborg, Denmark, which is a typical
undergo a process from huge data to meaningful information to microgrid configuration on MV level. This distribution system
massive intelligence. The integration for microgrids may contain contains a combined heat and power (CHP) plant with three 3.3 
distribution management system (DMS) where information layer MW gas turbine generators (GTGs) and three small local wind
allows for DA functions. farms. Each wind farm consists of two 2 MW doubly fed induction
Therefore, this paper proposes a brand new adaptive protection generators (DFIGs) and one 0.8 MW capacitor bank. The
strategy combined with machine learning for medium-voltage configuration of microgrid and wind farm 1 is given in Figs. 1 and
(MV) microgrids. The focuses of this paper are threefold: (1) The 2, respectively.
analysis of uncertain elements in a microgrid is studied More details of the model are given in [7, 16]. In the microgrid,
quantitatively by Pearson correlation coefficients on database, there is a data concentrator located at the substation working as a
which is the first time proposed for microgrid protection and central human–machine interface (HMI) between the downstream
control system from a statistic view. (2) A rule-based adaptive network and the utility operators. With the help of simple network
protection scheme is proposed for microgrids in extensive time protocol (SNTP) server or GPS or BeiDou Navigation
distribution system. This novel scheme can be integrated into DA Satellite System (BDS) controlled inter-range instrumentation
system performing smartly and flexibly dependent on modern group (IRIG) time code generator, real-time clocking can be
computer science. (3) The calculation of protective settings is guaranteed. Due to limited refresh rate of supervisory control and
given both for overcurrent and distance protection that are widely data acquisition (SCADA) system for microgrids in distribution
used in distribution systems. The relay settings and operating system, the paper proposes an independent online state diagnosis
modes can be reconfigured automatically for different operating model integrated with adaptive protection. The mathematical
states. Moreover, the proposed scheme which carried out on two model is trained dependent on database and can be integrated in the

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 771
© The Institution of Engineering and Technology 2019
Table 1 Correlation between the uncertain elements and • Capacitor bank switching events; (change from 0 to 100% of the
the voltage rated value with five steps in all. x7)
Voltage Amplitude Amplitude Angle Angle
(forward (backward (forward (backward The features dataset X = [x1, x2…xN ] includes the above seven
direction) direction) direction) direction) uncertain elements and matrix Y = [y1, y2…yN ] contains the
fault 6.5 × 10−4 −7.0 × 10−4 −0.06 −0.001 corresponding short-circuit measurements of voltage and current
location (both magnitude and angle). Pearson correlation coefficient is
fault 0.933 0.9335 0.225 0.514 adopted for quantitatively analysing the correlations between X and
resistance Y [26]. X and Y are transformed into normalised z-scores by the
operating 0.079 0.075 0.173 0.210 following equations:
mode
topology 0.078 0.085 0.173 0.260 xi − x̄
zxi = (1)
DFIG 0.001 sx
6.9 × 10−4 9.4 × 10−4 2.8 × 10−4
load 5.9 × 10−4 8.0 × 10−4 5.0 × 10−4 0.001 where x̄ is the mean and sx is the standard deviation of X given:
capacitor 0.002 0.003 −0.004 −0.009
bank N
1
(2)
N i∑
x̄ = xi
=1
Table 2 Correlation between the uncertain elements and
the current 1
N
(3)
N − 1 i∑
sx = (xi − x̄)2
Current Amplitude Amplitude Angle Angle =1
(forward (backward (forward (backward
direction) direction) direction) direction) The Pearson correlation coefficient for the two datasets X and Y is
fault −0.001 0.001 3.0 × 10−4 −3.0 × 10−4 given by:
location
fault −0.728 −0.945 0.944 0.908 1
N
(4)
resistance
N − 1 i∑
rxy = zxizyi
=1
operating 0.393 −0.155 0.068 0.126
mode
The above equation can also be reformulated as the covariance
topology 0.288 −0.003 0.064 0.192
between X and Y, normalised by the product of the standard
DFIG −5.3 × 10−4 0.002 0.002 0.002 deviations of X and Y:
load −4.7 × 10−4 0.002 0.001 0.002
N
capacitor 6.6 × 10−5 0.005 0.005 −0.013 ∑i = 1 (xi − x̄)(yi − ȳ)
bank
rxy =
N N
(5)
(1/(N − 1))∑i = 1 (xi − x̄)2 (1/(N − 1))∑i = 1 (yi − ȳ)2

where rxy ∈ [ − 1, 1]. The minus sign in the correlation coefficient

intermediate information layers or the control centres in power
grid. The forward direction of the fault current is defined from means that the measuring parameter decreases with the increasing
power grid side to the downstream buses in the network (from bus1 uncertain element. The more the absolute value of rxy is, the higher
to bus6), while the backward direction is from bus6 to bus1. the influence of the uncertain element on this measurement is.
Each variation of the uncertain elements in the analysis is
2.2 Data generating and data mining simulated separately for a three-phase short-circuit on Line23 in
DIgSILENT PowerFactory. The simulation contains 18,000 times
An electrical power system suffers unexpected faults on power calculation for all cases (6 × 6 × 2 × 2 × 2 × 5 × 5 × 5 = 18,000),
lines for various random causes with non-linear function. These while power flow calculation and complete short-circuit analysis
failures may hinder the stable operation of the system. Therefore, it are performed for both pre-fault and post-fault in each specific
is necessary to analyse the influences of the uncertain elements case.
before designing the protection schemes. This paper proposes to The simulation results are all recorded separately in matrix X
use data mining methodology to quantitatively extract the and Y as mentioned above and used further for training the state
meaningful information from the database, which is not the case in recognition models in Section 4. Therefore, this part can be treated
the previously reported reference. The uncertain elements (data) as a part of the data preprocessing for the adaptive protection
are performed as is shown below: scheme. Based on (5), the calculated correlations between the
uncertain elements (in X) and the post-fault measurements (in Y)
• The three-phase faults with different locations on one line are given in Tables 1 and 2.
segment; (change from 0 to 100% of the rated value with a Although the correlations are linear statistical values depending
variation of 20% every time, so six cases in all. x1) on the specific network, the results reveal quantitatively some
• The faults with different resistance; (change from 0 to 5 Ω with common phenomenon in microgrids and more cases or events can
a variation of 1 Ω every time, six cases in all. x2) be considered in practical applications. In Table 1, it can be
observed that the fault impedance, the operating mode (the
• Microgrid operation modes; (grid-connected = 1 and islanded = 
connected/islanded mode), the topology of the network, and the
0, realised by connecting/disconnecting R2, 2 cases in all. x3)
fault location have significant influence on the voltage and current
• Different topologies; (meshed = 1 and radial = 0, realised by (both forward and backward direction on line23) in descending
connecting/disconnecting relays on line16, two cases in all. x4) order. It shows that the operating mode and the network topology
• Different output of DFIGs (the randomness of DGs); (change should be operated under cautious planning. The backward voltage
from 0 to 100% of the rated value with a variation of 25% every angle has relatively large coefficients between the operating mode
time, 5 cases in all. x5) and the topology (0.210 and 0.260), which mean the angle stability
• Load variation events; (change from 0 to 100% of the rated and the critical clear time (CCT) may become serious during a
value with an interval of 25% every time, 5 cases in all. x6) fault. This is due to the capacity and the mechanical inertia of
backward wind farm are limited. Meanwhile, the capacitor bank
also has certain influence on the measurements. It shows that the

772 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779
© The Institution of Engineering and Technology 2019
 reactive compensators at proper locations are significant for the
stability of the network.
In Table 2, the correlation coefficient between the forward
current amplitude and the operating modes are relatively high
(0.393). It implies the short circuit current changes severely
between the grid-connected mode and islanded mode, because the
transmission grid has much higher power rating than the microgrid
itself. Meanwhile, the network topology influences the current (in
Table 2) more than the voltage (in Table 1). Therefore, the design
and calculation of overcurrent protection should consider the
influences in the system. On the other hand, the other variations
have almost the same correlation coefficients that are relatively
small. However, the intermittence and connection states of DGs
should be considered. Some storage units may be required around
the local power plants to avoid the complete loss of DFIGs' support
during a fault. Another countermeasure is to collect the real-time
states of DGs to the control centre so that it enables the
modification of the protective settings of relays based on the
current operating situation.

3 Protection calculation and test

3.1 Time domain simulation on different operating states
In order to reveal the influence of the operating mode and
topology, time domain simulation (electromagnetic) is studied for
different operating modes in this subsection. There are mainly four
operating modes considered for the microgrid test: grid-connected
radial mode, grid-connected looped mode, islanded radial mode,
Fig. 3 Fault current magnitude for different operating states and islanded looped mode, which are controlled by R2, R16, and
R61 as is shown in Fig. 1. In the proposed adaptive protection
scheme (in Section 4), the connection states of DGs are collected
periodically to the control centre. The microgrid is normally
operated in grid-connected radial mode to avoid power oscillations
in the looped/meshed grids during a fault. If there is a fault or
emergency in the system, the operating mode can be reconfigured
based on the online decision-making.
Figs. 3 and 4 present the time-domain simulation results
(electromagnetic transient) when a three-phase short-circuit with
0.2 Ω fault resistance occurs on line45 at 0 ms for the four
operating modes. It can be found that the fault current and voltage
change severely in different operating states, which may influence
the coordination of the relays. If the fault cannot be cleared fast,
oscillation may exist in the network (in islanded looped mode)
which may influence the system stability (e.g. lose synchronism). It
is mainly due to the power oscillation among the CHP and local
wind plants through line16. The simulation results illustrate that
even a same fault on the system may lead to different consequences
and measurements for different operating modes. Therefore, it is
imperative to calculate the protective settings based on the specific
operating states so that the fault can be cleared selectively and
quickly.

3.2 Protective settings calculation for overcurrent relays

Since it has been known that the short-circuit measurements are
more related to the operating modes, topologies of the network and
the connection states of DGs, the relay settings should be modified
based on the current operating situations. As is illustrated in Fig. 5,
Fig. 4 Fault voltage magnitude for different operating states the transmission grid and CHP bring infeed currents for the
backward relays in looped mode (with R16 and R61 closed), whose
short-circuit capacity are much larger than the DGs. Meanwhile,
the different connection states of DGs also influence the infeed
current on the network. Here, the calculation algorithm is given for
directional overcurrent relay (DOR) and distance relay,
respectively. However, the calculation algorithms for protective
settings are not only limited on the presented ones here.
For DORs, the coordination of protective settings can be treated
as a linear programming problem for minimising the total function
time. Fig. 6 gives the characteristic of DOR provided by
international electrotechnical commission (IEC). The characteristic
Fig. 5 Infeed current on Line56 under looped operation modes (IEC255-3) is given below:

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 773
© The Institution of Engineering and Technology 2019
 A × TDS
t= B (6)
I f /I p − 1

where: A is a constant; B is a characteristic index; If is the fault

current magnitude; Ip is the pickup current; TDS is the time dial
setting.
The coordination algorithm is based on optimising an objective
function of operating times of the primary relays while subjected to
maintaining the operation of backup relays coordinated. The
objective function for a system that contains m relays can be
formulated as:
f max m Fig. 6 Characteristic curve of DOR
z = Min ∑ ∑ ti, i + ∑ ti, j (7)
f =1 i j

where: f indicates the fault location; ti, i is the operating time for the
primary relay located at i for a fault near at I; and ti, j is the
operating time for the backup relay located at j for the same fault
near at i.
The operating time ti, j for the backup relays should be >ti, i,
which can be expressed as following:

ti, j ≥ ti, i + Δt (8)

where Δt is the coordination time interval (CTI) and set as

constant. Fig. 7 Block diagram of time-stepped distance relay
When Ip and If are confirmed through power flow calculation
and short-circuit analysis, TDS of each relay can be obtained from
(7) as a linear programming problem. To keep the reliable
operation, Ip can be set as 1.5 times of the maximum anticipated
load current in the next several hours or days. More details of DOR
settings calculation can be found in [19, 27].

3.3 Protective settings calculation for distance relays

For simplicity, Ohm-based distance protection (with three time-
stepped zones) is chosen for its good selectivity and sensitivity. As
is shown in Fig. 7, Zone1 is an instantaneous protection, while
Zone2 and Zone3 have definite time-stepped delay. The measuring
impedance Zm is calculated from the measuring voltage U˙ m and
current I˙m at the relay point, which is shown in (9).

U˙ m
Zm = (9)
I˙m

The settings of each relay can be calculated as follows: Fig. 8 Protective settings for normal operation mode

I
Zset I
= Krel Zline scheme for distance protection and/or differential protection in
electrical systems.
II
Zset = II
Krel (Zline + I
KbZset 2) (10)
III
Zset = III
Krel (Zline + II
KbZset 2) 4 Rule-based adaptive protection
4.1 Adaptive protection with state recognition for microgrids
where the super script in the equations represents the different
protection zone. Kb is the branch coefficient that offsets the infeed The above calculation results of relay settings can be recorded in
currents, which can be obtained from short-circuit calculation. Krel I
, the control centre or relays beforehand, or calculated online, which
is dependent on the specific requirements in the electrical system.
Krel, Krel are the reliability coefficient of each protection zone,
II III
Due to the limited refresh rate of current SCADA system and
which is often set between 0.8 and 0.85. Zline is the impedance of the paucity of measurements in some microgrids, real-time least
the primary protected line. Zset
I
2 and Zset2 are the protective settings
II square estimation (LSE) usually are not available for microgrid
of Zone 1 and Zone 2 of the downstream relays, respectively. protection. When a fault occurs, the operators in workstation often
To ensure the protection selectivity, the delay settings of time- have to retrieve the on-field measurements manually or even go to
stepped distance relays are set at 0.35 s for Zone2 and 0.7 s for the site. Meanwhile, the variable connection states of DGs in the
Zone3, respectively. The pickup time of the relay is set to 20 ms. network influence the selectivity of protection. Another
The calculated protective settings for grid-connected radial mode phenomenon may rise in power grids that some mobile DERs and
with full DGs connection are shown in Fig. 8. The test of the electric vehicles (EVs) are connected to the grid as DGs with a
distance protection can be found in [7, 28]. For extreme fault cases feature of changing the operating modes among generators, loads,
in power grids, overcurrent protection, distance protection, and/or and compensation equipment on some locations sometime. The
differential protection can be activated together in modern relays, above mentions require an alternative approach to recognise the
e.g. the overcurrent protection can be used as the backup protection state of downstream network and make decision automatically for a
computer-based control system. As the huge historical and/or

774 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779
© The Institution of Engineering and Technology 2019
 Fig. 10 Simple ANN illustration

and post-fault are required. In this experiment, 108,000 (18,000 × 

6lines) cases with a three-phase short-circuit are performed in
DIgSILENT PowerFactory, which is realised as described in
Section 2. The fundamental frequency component of the voltage
and current phasors (amplitude and angle) are calculated in the
simulations, which are finally set as the inputs of the estimation
model. Moreover, the proposed scheme enables the self-learning
capability for improvement if various new data streams are
available during the operation and (or) test, which is the advantage
comparing with the conventional methods.

4.2 ANN model for classification

As a hybrid approach, the ANN-SVM estimation model is critical
in the proposed scheme. The training and test of ANN are
implemented in MATLAB while SVM is performed in LIBSVM.
Out of the data available, 30% are randomly selected as the test set
while the remaining data are used for the model training for both
models.
ANNs are a class of powerful non-linear machine-learning
Fig. 9 Flow chart of the rule-based adaptive protection scheme predicted class from the ANN model and each row represents
models [23, 24]. Among ANN models, the feed-forward neural
simulation data have become possible in power system, the paper network, also known as the multilayerperceptron (MLP), is most
utilises the methodologies in machine learning to train the successful. As can be seen in Fig. 10, each layer of MLP contains
mathematical model such that the protection and control can adapt numbers of neurons that take inputs from the previous layer
to the variable operating states. Furthermore, an integrated rule- neurons. The excitation signals are fed into the input layer. Each
based scheme is adopted, which has distinctive advantages of neuron is a logistic regression model that acts as a non-linear
applying on modern industrial computers. The simulation data are function of its inputs (xi) and generates an output (y) as an input for
proposed as a partial source of the machine learning–based model the next layer or as a final output, which can be found in Fig. 11.
to implement smart grid type protection and control. The The neurons are connected by links, and each link has a weight (w)
implementation may consist of a DMS installation at distribution associated with it. The link weights and the biases of neurons are
substation which combines SCADA system and bus/feeder adjusted according to the training data set during the training
measurements. process. After training, the ANN is ready to be used for the testing
The flow chart of the proposed scheme is given in Fig. 9. The data or data from real-world applications. The ANN model is
processor/computer in the control centre (or the intermediate level) determined by its network architecture, transfer function, and
processes the proposed protection scheme periodically, which learning rules [25]. The output neuron model (Fig. 11) as a
works in an ‘interrupt mode’. Once the interrupt is accepted by the function of d inputs can be expressed as:
processor, tripping signal is retrieved first (tripping action has the
highest priority) to confirm whether the system operates in normal d
state or fault state. If a tripping action occurs, control centre begins y(x, w) = f (u) = f ∑ w jφ j(x) (11)
to retrieve the measurements automatically and sends them to the j=1
ANN model to estimate whether there is a fault and which line is
faulted. If the model confirms there is a permanent fault on the where y is the output, φ j(x) is the basis function and f (.) is the
line, fault location estimation is followed using SVM. Since the neuron activation function. u is the summation output signal:
fault has been confirmed on the specific power line, fault location
estimation is limited on the tripped line to enhance the accuracy. u = wT x (12)
For redundancy and further ‘self-learning’ and ‘self-training’, the
results from ANN-SVM can work together with LSE. Then, a where w = [w0, w1, …wd], x = [x0, x1, …xd]. The logistic-sigmoid
decision table is retrieved whether it is needed to reconfigure the transfer function is used for the signal delivering between the
network and the corresponding relay settings, depending on the neurons at two layers:
pre-defined settings or online enhanced decision-making. If it is
required to reconfigure the topology, activation signal is sent to the 1
corresponding on-filed relays. Some shedding actions and online f (u) = (13)
1 + e− k(u − u0)
calculation may be required in some situations. On the other hand,
if the system is operated in normal state, the connection states of
where u0 is the value of the midpoint of sigmoid, k is the steepness
DGs are retrieved from each local wind farm centre or DG.
According to the connection states of DGs, the protective settings of the curve. The ANN's capability of dealing with non-linear and
are modified as similar as in fault state. complex problems is realised by making the basis function φ j(x)
The proposed scheme requires the parameters and knowledge of depend on parameters and then to allow the parameters to be
the network. To train the hybrid estimation model, power flow adjusted, along with the coefficients {wj}, during training:
calculation and complete short-circuit analysis both for pre-fault

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 775
© The Institution of Engineering and Technology 2019
 M D around 80%, while they are higher for Line12, Line 56, and
∑w ∑w (14)
(2) (1)
yk(x, w) = f kj
h ji
xi Line16. In the first scenario, the diagnosis model cannot identify
j=0 i=0 all the faults in the microgrid. It indicates that measurements on
bus2, bus3, or bus4, which is inside the microgrid, may be
where h(.) is the differentiable, non-linear activation functions. D necessary to improve the estimation accuracy. Therefore, the
and M are the numbers of neurons on different layers. enhanced decision-making is necessary for the system to main the
Back propagation learning rule is adopted to adjust the weights reliable operation. For comparison, the ANN model is trained again
and biases [25]. The problem of finding a suitable architecture and with full measurements from all the six buses. The accuracy is
the corresponding weights of the network is of central importance shown in Table 4. It can be found that the accuracy keeps at 100%
in the area of training ANNs. (with 0 error) for different hidden layers since the measurements
Used as a classification model, the ANN model in the scheme (from all buses) used in this case implicate the direction for a fault
has seven outputs with relevant labels to represent the six power on each feeder, which is similar as the absolute selectivity of the
lines (line12, line23, line34, line45, line56, and line16) and no differential protection. Moreover, the comparison results of the test
fault. The input neurons for ANN Model are the measurements show that the performance of the learning-based models can keep
(voltage and current) on different buses. The number of hidden being improved with the increasing measurements and the relevant
neurons is set as 64, while the number of training epochs is set as data.
1000.
Concerning the input neurons, two cases are considered for
4.3 SVM model for fault location
comparison (the overdetermined set of measurements is available
or not): only measurements from bus1 and bus6 are available, and Used as a regression model in the scheme, the SVM model outputs
full measurements from the six buses (bus1, bus2, bus3, bus4, the estimated fault location on the power line. SVM is a supervised
bus5, and bus6) are available. If the measurements from bus1 to learning method for a linear model, while it performs well for non-
bus6 are available for training (the worst condition), the average linear problems by using kernel trick. With the basis function φ(x),
accuracy of ANN for all test cases is 90.3%, while the confusion the input vector x in SVM is mapped into a higher D-dimensional
matrix (error matrix) is given in Table 3. Each column of the feature space [29]. The output y(x) in the high-dimensional space is
confusion matrix represents the instances in an actual class. The defined as (15):
accuracy of ANN is shown in Fig. 12. From Table 3 and Fig. 12, it
can be found that the accuracies of Line23, Line34, and Line45 are

Fig. 11 Single neuron model

Table 3 Confusion Matrix of ANN with limited measurements only from bus1 and bus6
Line12 Line23 Line34 Line45 Line56 Line16 No-fault Accuracy, %
line12 5171 29 245 2 3 0 0 94.9
line23 81 4714 591 25 34 0 0 86.5
line34 237 808 4237 134 17 0 0 77.5
line45 0 86 619 4353 348 0 0 80.5
line56 0 0 192 139 4981 32 0 93.2
line16 0 0 0 0 0 5359 0 100
no-fault 0 0 0 0 0 0 5327 100

Fig. 12 Accuracy of ANN only with measurements on bus1 and bus6

Table 4 Error of ANN with full measurements

Layers Error for test cases Error for all cases
2 0 0
3 0 0
4 0 0
5 0 0

776 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779
© The Institution of Engineering and Technology 2019
Table 5 Error of SVM
Line 12 Line 23 Line 34 Line 45 Line 56 Line 16
test error 0.24% 0.31% 0.57% 0.44% 0.16% 0.1%

n
∑ (λi − λ′i) = 0
i=1
(21)
0 ≤ λi ≤ C
0 ≤ λ′i ≤ C

where K(xi, x j) = φT(xi)φ(x j) is an inner-product kernel.

Equation (21) can be finally solved as a quadratic programming
optimisation problem. The choice of the coefficients ε and C is the
critical problem during training SVM. Through adjusting and
Fig. 13 Accuracy of SVM model testing, the final parameters used are C = 1000 and ε = 0.05.
In the SVM model, there are six sub-models built separately for
D the six power lines using the same measurements in ANN. One of
y(x) = wTφ(x) = ∑ w jφ j(x) (15) the six sub-models in the SVM model will be activated once the
j=0
fault is confirmed on the corresponding power line in ANN. Full
measurements on each bus are used according to the test results in
where w = [w0, w1, …, wd]T is the weight vector, ∥ w ∥2 < γ. γ is a the section above. The following metric is used to assess the
constant. φ(x) = [φ0(x), φ1(x), …, φd(x)]T is the basis function performance of SVM for fault location:
vector.
Introducing the ε−insensitive loss function Lε(y(x), β): de − da
err = ⋅ 100% (22)
l
0, for β − y(x) < ε
Lε(y(x), β) = (16) where de is the estimated value, da is the actual value, and l is the
β − y(x) − ε, for β − y(x) ≥ ε
length of the power line.
where ε is a free parameter that serves as the threshold of the loss The test results for each sub-model in SVM model are given in
Table 5 and Fig. 13. It can be found the error is always within 1%
function. x is the input vector, β is the target value. The main task
for all the tests, which are acceptable for the requirements on the
of training a SVM model is minimising the error function:
estimation tasks.
n As is shown in Fig. 9, after the state diagnosis in the scheme,
1 the corresponding decisions will be followed to decide whether
(17)
n i∑
E= Lε(y(xi), βi)
=1 reconfigure or not. Meanwhile, alarm is sent to the utility
operators. Some additional decisions can be considered further by
where n is the number of learning data pairs (xi, βi). Considering the utility. Finally, the actuators perform the reconfiguration
the slack variables ζ and ζ′ for the outliers in the data, the training automatically, while the related events are recorded with excite
task can be treated as: time.
This proposed adaptive protection scheme is practical for smart
n grids when the penetration of DGs and microgrids are increasing in
1 T distribution system, which requires more intermediate layers
min w w+C ∑ (ζi + ζi′) (18)
2 i=1 controlling the downstream network via various computers. Due to
its flexible integration in the protection and control system, the
subject to: rule-based protection scheme reduces the complexity of fault
retrieve for distribution system operator (DSO) and enhances the
wTφ(xi) − βi ≤ ε + ζi reliability of operation. With the help of SNTP server and/or IRIG
time code generator, synchronous phasor measurement can be
βi − wTφ(xi) ≤ ε + ζi′ adopted for microgrids in DA.
(19)
ζi ≥ 0
4.4 Test on IEEE 9 bus model
ζi′ ≥ 0
To verify the feasibility of the recognition model, the test is
C is an adjustable parameter that determines the compromise performed on IEEE 9 bus standard model for its modest
between the model's complexity and accuracy. In order to solve the configuration (Fig. 14). The data are obtained with a smaller step
training task, Lagrangian function and Lagrange multipliers are as following:
introduced so that (18) is transformed into the dual problem:
• A three-phase faults with different locations on one line
n n segment; (change from 0 to 100% of the rated value with a
max ∑ βi(λi − λ′i) − ε ∑ (λi + λ′i) variation of 10% every time, 11 cases in all)
i=1 i=1
(20) • Several faults with different resistance; (change from 0 to 10 Ω
1
n n with a variation of 1 Ω every time, 11 cases in all)
2 i∑ ∑ (λi − λ′i)(λ j − λ′ j)K(xi, x j)
± • Variation on Load1; (change from 0 to 100% of the rated value
=1 j=1
with an interval of 25% every time, 5 cases in all)
subject to: • Variation on Load2; (change from 0 to 100% of the rated value
with an interval of 25% every time, 5 cases in all)
• Variation on Load3; (change from 0 to 100% of the rated value
with an interval of 25% every time, 5 cases in all)

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 777
© The Institution of Engineering and Technology 2019
Fig. 14 IEEE 9 bus model

Table 6 Error of ANN with different measurements

Measurements Accuracy for test cases Accuracy for all cases
bus 5, 6, 8 7.18% 7.14%
bus 4, 7, 9 0.74% 0.62%
all buses 0 0

Table 7 Error of SVM

Line1, % Line2, % Line3, % Line4, % Line5, % Line6, %
bus 5,6, 8 0.25 0.11 0.04 0.23 0.14 0.25
bus 4,7, 9 0.12 0.2 0.23 0.03 0.12 0.02
all buses 0.25 0.21 0.25 0.23 0.13 0.22

The collected data (11 × 11 × 5 × 5 × 5 × 6lines = 90,750) are improved with various new coming data and knowledge, which is
used to train the ANN and SVM model. 144 hidden layers and 500 also the excellent advantage of machine learning. In practical
training epochs are set for ANN. There are seven outputs in ANN, application, massive variables and scenarios can be performed with
which are fault on line1, line2, line3, line4, line5, line6, and no certain accuracy in computers/machines for self-learning and/or
fault. The error of ANN models with different collected data as the self-training. With the development of technology, the proposed
inputs is given in Table 6. It can be found that the accuracy of methodology may play a role in protection and control systems in
ANN used data only from Bus4-7-9 is better than Bus5-6-8 due to future smart grid.
the connection of branch line and transformer on Bus4-7-9. It
implies the measurements from Bus4-7-9 contain more information 5 Conclusion
for the state recognition. The error of ANN for case ‘all busses’ is
0, because the used data are ideal without corruption for training Here, we proposed a rule-based adaptive protection scheme
and testing ANN. On the other hand, C = 1000, ε = 0.02 is set for combined with machine-learning methodology for microgrids. The
SVMs in this test, while the output is the estimated fault location correlation between the uncertain elements and the short-circuit
for the specific lines. As is shown in Table 7, all SVM models with measurements was analysed quantitatively by Pearson correlation
different measurements data work well with small errors. From the coefficients from a statistic view. The results reveal that the
test results, it can be found the ANN-SVM model can also operating states and topology influence the short-circuit
recognise the states accurately for transmission systems. Moreover, measurements and stability of the system significantly. An ANN-
it shows that the ANN-SVM model can still work well with SVM diagnosis model was integrated into the adaptive protection
enhanced decision-making when the overdetermined equations are scheme for state recognition in DA to extract and utilise the
not available for LSE. information and knowledge from the growing massive data streams
The paper utilises data mining technology first to extract the in power system. For reliable operation and system resilience,
meaningful information and knowledge in the electrical system, reconfiguration can be performed automatically and intelligently to
then the learning-based model is trained and integrated into the change the topology of the network and update the protective
automatic protection and control system. It undergoes a process settings. With the assistance of machine learning, this paper has
from huge data to meaningful information to final intelligence. The showed that the adaptive protection scheme can modify the settings
proposed methodology requires the original data (or prior properly and intelligently for different operating situations. At the
knowledge) to hold the accuracy to avoid the subsequent influence same time, the integrated adaptive protection system enables
on the final results. Some data processes may be required before accurate information for further intelligent operation/control in
the practical application. In case of data corruption, some smart grid and energy internet, which has a ‘self-learning’ and
additional countermeasures should be considered. For other diverse ‘self-training’ capability.
fault or event types, the data can be obtained in a similar way from
precise simulation or historical database in a system. It should be 6 References
emphasised that the machine-learning-based model can work
[1] Lasseter, R.H.: ‘Microgrids’. Power Engineering Society Winter Meeting,
together with the other mathematical models (LSE etc.) to improve New York, NY, USA, 2002, vol. 1
the accuracy and compatibility. In the specific situations, the [2] Hengwei, L., Chengxi, L., Guerrero, J.M., et al.: ‘Modular power
operators can also intervene the automatic process to maintain the architectures for microgrid clusters’. Green Energy, Sfax, Tunisia, 2014, pp.
reliable operation. The performance of ANN-SVM can keep being 199–206

778 IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779
© The Institution of Engineering and Technology 2019
[3] Anderson, P.M.: ‘Power system protection’ (Wiley, New York, 1999), pp. [17] Haughton, D.A., Heydt, G.: ‘A linear state estimation formulation for smart
201–240 distribution systems’, IEEE Trans. Power Syst., 2013, 28, (2), pp. 1187–1195
[4] Anthony, M.A.: ‘Electric power system protection and coordination’ [18] Mirsaeidi, S., Said, D.M.: ‘A protection strategy for micro-grids based on
(McGraw-Hill, New York, 1995), pp. 342–346 positive-sequence component’, IET Renew. Power Gener., 2015, 9, pp. 600–
[5] Blackburn, J.L.: ‘Protective relaying principles and applications’ (Marcel 609
Dekker, New York, 1998), pp. 383–408 [19] Hengwei, L., Guerrero, J.M., Chenxi, J.: ‘Adaptive overcurrent protection for
[6] Jayawarna, N., Jenkins, N., Barnes, M., et al.: ‘Safety analysis of a microgrids in extensive distribution systems’. 42nd Annual Conf. of the IEEE
microgrid’. Future Power Systems, Amsterdam, Netherlands, 2005 Industrial Electronics Society (IECON2016), Italy, Florence, October 2016,
[7] Hengwei, L., Chengxi, L., Guerrero, J.M., et al.: ‘Distance protection for pp. 4042–4047
microgrids in distribution system’. 41st Annual Conf. of the IEEE Industrial [20] Joe-Air, J., Jun-Zhe, Y., Ying-Hong, L., et al.: ‘An adaptive PMU based fault
Electronics Society (IECON2015), Japan, Yokohama, November 2015, pp. detection/location technique for transmission lines. I. Theory and algorithms’,
731–736 IEEE Trans. Power Deliv., 2000, 15, pp. 486–493
[8] Dewadasa, M.: ‘Protection for distributed generation interfaced networks’. [21] Joe-Air, J., Ying-Hong, L., Jun-Zhe, Y., et al.: ‘An adaptive PMU based fault
PhD thesis, Dept. Electrical Engineering, Faculty of Built Environment and detection/location technique for transmission lines. II. PMU implementation
Engineering, Queensland University of Technology, Australia, 2010 and performance evaluation’, IEEE Trans. Power Deliv., 2000, 15, pp. 1136–
[9] Dewadasa, M., Majumder, R., Ghosh, A., et al.: ‘Control and protection of a 1146
microgrid with converter interfaced micro sources’. Third Int. Conf. on Power [22] Cannas, B., Celli, G., Marchesi, M., et al.: ‘Neural networks for power system
Systems, Kharagpur, India, 2009, pp. 1–6 condition monitoring and protection’, Neuro Comput., 1998, 23, pp. 111–123
[10] Nikkhajoei, H., Lasseter, R.H.: ‘Microgrid fault protection based on [23] Tan, Z.-H.: ‘Hybrid evolutionary approach for designing neural networks for
symmetrical and differential current components’. Power System Engineering classification’, IEE Electron. Lett., 2004, 40, (15), pp. 955–957
Research Center, 2006 [24] Saha, M.M., Izykowski, J.J., Rosolowski, E.: ‘Fault location on power
[11] Rockefeller, G.D., Wagner, C.L., Linders, J.R., et al.: ‘Adaptive transmission networks’ (Springer Science & Business Media, Springer-Verlag, London,
relaying concepts for improved performance’, IEEE Trans. Power Deliv., UK, 2009), pp. 395–399
1988, 3, pp. 1446–1458 [25] Bishop, C.: ‘Pattern recognition and machine learning’ (Springer, UK, 2006),
[12] Brahma, S.M., Girgis, A.A.: ‘Development of adaptive protection scheme for pp. 232–240
distribution systems with high penetration of distributed generation’, IEEE [26] Benesty, J., Chen, J., Huang, Y.: ‘Pearson correlation coefficient’, ‘Noise
Trans. Power Deliv., 2004, 19, pp. 56–63 reduction in speech processing’ (Springer, Berlin, Heidelberg, 2009), pp. 1–4
[13] Dhar, S., Dash, P.K.: ‘Differential current-based fault protection with adaptive [27] Chattopadhyay, B., Sachdev, M.S., Sidhu, T.S.: ‘An on-line relay coordination
threshold for multiple PV-based DC microgrid’, IET Renew. Power Gener., algorithm for adaptive protection using linear programming technique’, IEEE
2017, 11, pp. 778–790 Trans. Power Deliv., 1996, 11, pp. 165–173
[14] Kar, S., Samantaray, S.R.: ‘Time-frequency transform-based differential [28] Hengwei, L., Guerrero, J.M., Vasquez, J.C., et al.: ‘Adaptive distance
scheme for microgrid protection’, IET Gener. Transm. Distrib., 2014, 8, pp. protection for microgrids’. 41st Annual Conf. of the IEEE Industrial
310–320 Electronics Society (IECON2015), Japan, Yokohama, November 2015, pp.
[15] Oudalov, A., Fidigatti, A.: ‘Adaptive network protection in microgrids’, Int. J. 725–730
Distrib. Energy Res., 2009, pp. 201–226 [29] Salat, R., Osowski, S.: ‘Accurate fault location in the power transmission line
[16] Mahat, P., Zhe, C., Bak-Jensen, B., et al.: ‘A simple adaptive overcurrent using support vector machine approach’, IEEE Trans. Power Syst., 2004, 19,
protection of distribution systems with distributed generation’, IEEE Trans. pp. 979–986
Smart Grid, 2011, 2, pp. 428–437

IET Gener. Transm. Distrib., 2019, Vol. 13 Iss. 6, pp. 770-779 779
© The Institution of Engineering and Technology 2019