An Intelligent Overcurrent Protection Algorithm of

Distribution Systems with Inverter based

Distributed Energy Resources
Lina He Shuaiang Rong Chengwei Liu
Department of Electrical and Computer Department of Electrical and Computer Department of School of Electrical and
Engineering Engineering Electronic Engineering
University of Illinois at Chicago University of Illinois at Chicago Huazhong University of Science &
Chicago, USA Chicago, USA Technology
[email protected] [email protected] Wuhan, China
[email protected]

Abstract—To facilitate renewable energy, distributed energy conventional synchronous generators. However, in current
resources (DERs) have been significantly integrated into distribution systems, the fault current fed from IBDERs has
distribution systems through power electronics-based inverters. totally different characteristics due to their inverter control.
The control of these inverters can limit the fault currents fed This can cause the malfunction of the conventional
from DERs during a short-circuit (SC) fault. The resulting SC overcurrent protection system, leading to significant damage
current can be too low to trigger conventional overcurrent to costly electrical equipment in severe situations.
relays on distribution feeders, leading to a protection failure
(should operate but does not). To address this issue, this paper It is known that the IBDER has a weaker SC fault current
proposes an intelligent overcurrent protection scheme, which feeding capability compared with a conventional synchronous
applies a machine learning algorithm innovatively, i.e., the generator [4-5]. In our previous study [6], a quantitative
radial basis function neural network (RBFNN), to learn and analysis showed that the SC fault current fed from a
detect the SC fault currents fed by inverter based DERs conventional synchronous generator is about 5 times higher
(IBDERs) intelligently based on the features of their time series than its current rating. It can be easily recognized and detected
data. This algorithm can be implemented in the microprocessor by an overcurrent relay. However, the SC fault current fed
of a digital relay on a distribution feeder to online detect the SC from an IBDER is much lower and only 1.2 to 1.5 times of its
faults of distribution systems with IBDERs fast and accurately. current rating. The main reason is that the SC fault current fed
Numerical simulation has been performed in a distribution from an IBDER is limited by the current limitation controller
system benchmark with IBDER integration to verify the of its inverter, which is used to protect power electronic
effectiveness of the proposed algorithm.
switches of the inverter that are sensitive to a high current. The
resulting low SC fault current can be lower than the current
Keywords— Inverter based DER, intelligent overcurrent
setting of an overcurrent relay. As a result, the overcurrent
protection, RBFNN, SC fault
relay fails to activate the corresponding CB to interrupt the SC
I. INTRODUCTION fault. Even the low SC fault current is slightly beyond a
current setting, it still leads to a sluggish operation of the CB
Worldwide power systems are transitioning to a resource tripping due to the inverse time-current characteristics of
mix that relies less on coal and nuclear while integrating more overcurrent relays [7].
locally available DERs such as natural gas, wind, and solar
photovoltaic (PV) due to vast energy demand and climate In the existing literature, some researchers have proposed
impacts [1]. In the last decade, DERs have been greatly several methods to address the protection issue caused by the
promoted due to their various benefits, such as improving the penetration of DERs including distributed synchronous
resiliency of power systems, reducing the overall electricity generators and IBDERs. A protection scheme is proposed
cost, and increasing the efficiency of renewable energy based on predefined protection settings of conventional
conversion. It is reported that the total capacity of DERs overcurrent relays for different operation modes, i.e., the grid-
around the world is 158.3 GW in 2019, and will be up to 345 connected and island modes of the DERs [8]. However, this
GW by 2028, including 245 GW from solar PV installations method is not efficient for protection systems with high-level
[2]. penetration of DERs. With the fast development of
microprocessors, it is possible to implement computationally
By definition, the DER include both small scale diesel intensive algorithms for high speed and accurate detection of
driven synchronous generators and inverter based generation fault signals [9]. Reference [10-11] discussed strategies
technologies utilizing renewable energy. Despite relying on adaptive overcurrent relays to detect faults (normal
environmental and economic benefits, the significant or high impedance) in distribution systems that cannot be
penetration of these inverter based DERs (IBDERs) has led to detected due to the integration of DERs into the network. Most
undesired issues to the protection system of distribution of these strategies rely on modifying relay settings according
systems due to the decline of the SC fault current of to system operating conditions (i.e., newly connected DERs
distribution feeders [3]. Traditional distribution systems are and actual loading conditions). The methods proposed in [12-
radial and protected by circuit breakers (CB) on distribution 13] track and estimate both of the current and voltage, and
feeders, which are controlled by overcurrent relays. When a adaptively choose an optimal current setting for the
SC fault current exceeds the threshold setting of an overcurrent relay in different operation modes of the DERs.
overcurrent relay, the relay sends a trip signal to the However, all the research work mentioned above mainly
corresponding CB to isolate the fault. The settings of focused on the protection coordination of the overcurrent
overcurrent relays in distribution systems are determined relays due to the integration of generally DERs including
based on the characteristics of the SC fault current fed from

978-1-7281-5826-6/20/$31.00 ©2020 IEEE 2746

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.
 synchronous generators, and they did not specifically study and delivered to the grid via its current and voltage control.
the impact of the lower fault current fed by the IBDERs. (3) To ensure the voltage and current phasor are matched with
To the best of the authors’ knowledge, the discussed the main grid, a grid synchronization is achieved via a phase-
protective issue caused by the integration of IBDERs has not locked loop (PLL) in the inverter controller. The output of the
been solved yet. A voltage enhanced current detection method inverter controller is a gate pulse signal generated via a pulse
is discussed in [14] to detect the SC fault currents fed from an width modulator (PWM).
IBDER. This method added a voltage dip detection function
to the overcurrent protective relay to determine a SC fault. Grid
However, this protection scheme is not reliable since most of
IBDERs in power systems are required to have voltage PV Array Inverter
support capabilities, which can lead to a subtle voltage dip Filter
during faults that cannot be detected easily [15]. C
To address the above issue, this paper proposes an Vdc
ipv, vpv
intelligent protection algorithm using the machine learning PMW u, i

Current/voltage control
algorithm innovatively, i.e., RBFNN, to detect the low SC Local load
MPPT Grid synchronization
faults fed from IBDERs. By learning the features of time
series data of currents during normal and fault conditions in Fig.1: Single-line grid-connected PV system
an offline training, the gained model is applied to intelligently
detect the SC fault current fed from an IBDER. The proposed B. Inverter control strategy
algorithm can be implemented in the microprocessor of digital The inverter applies a double-loop control, where the outer
relays to detect faults in real time. In order to generate the loop is to control the DC link voltage to enable the maximum
training data, a distribution system benchmark with an PV output, and the inner loop is a current control loop of the
integrated IBDER is developed to obtain the current data fed inverter, as shown in Fig.2. The grid synchronization is
from the IBDER under both normal and fault conditions. The realized via a PLL control, which generates the grid voltage
RBFNN has been used for the fault diagnosis in many areas angle to ensure the output current of the inverter matches with
due to its excellent performance on the universal optimal the power grid in phase and frequency.
approximation and fast learning convergence [16]. However, In order to achieve an easy control of the output power, the
it is the first time to be applied in power systems for the SC three-phase (phase a, b, c) varying voltage and current signals
fault current detection. are transformed into constant values including the d-axis
This paper is organized as follows. Section II discusses the components and the q-axis components through Park’s
system configuration including a benchmark of a distribution transformation [18]. In this way, the actual active power P
system with an IBDER and the inverter control scheme. and reactive power Q could be expressed in (1) and (2).
Section III discusses the RBFNN and proposes the intelligent
overcurrent protection algorithm. Section IV shows a case 𝑃𝑃 = 𝑉𝑉𝑑𝑑 𝐼𝐼𝑑𝑑 + 𝑉𝑉𝑞𝑞 𝐼𝐼𝑞𝑞 (1)
study to verify the effectiveness of the proposed algorithm. 𝑄𝑄 = 𝑉𝑉𝑞𝑞 𝐼𝐼𝑑𝑑 − 𝑉𝑉𝑑𝑑 𝐼𝐼𝑞𝑞 (2)
Section V concludes this paper and discusses future work.
where the symbols 𝐼𝐼𝑑𝑑 (𝑉𝑉𝑑𝑑 ) and 𝐼𝐼𝑞𝑞 (𝑉𝑉𝑞𝑞 ) are the grid currents
II. SYSTEM CONFIGURATION (voltages) in the d-q coordinate system. In the inner current
This paper uses a typical IBDER model, i.e., a grid- loop, a decoupling current controller is applied to decouple P
connected distributed PV system to generate, to develop a and Q, so that they can be controlled separately via the d and
distribution system benchmark. q components of the current.
A. Grid-connected PV system ipv iga Us
Fig. 1 shows the sigle-line diagram of a grid conneted igb
distributed PV system in a distribution system. The PV system vpv Vdc
igc
is mainly composed of a PV array, a large DC-link capacitor
and an inverter. The functions of the main components are vpv , ipv vga,vgb,vgc
decribed as follows: (1) the PV array converts solar energy MPPT
PLL Gate
Volatge pulse
into DC power, which is affected by weather conditions, controller I θ signal
Vd
mainly including temperature and illumination [17]; (2) a
max
I d* +
Vmpp PI
+
large DC-link capacitor between the PV array and the inverter - Vdc -Imax - I
d d-q
d,q axis
stabilizes the DC link voltage; (3) the inverter can maintain Id Iq Vd Vq decoupling to PWM
the maximum output power of the PV array and transform the current
controller
a-b-c
output of DC power into the AC power injected into the grid a-b-c to d-q Iq =0 Imax
*
- Iq
+
via its control system; and (4) the LCL filter is used to reduce iga, igb, igc vga,vgb,vgc
θ
Vq
-Imax
the total harmonics distortion (THD) of the injected current to
the grid. Fig.2 Inverter control strategy
The inverter is the core component of the PV system, Fig. 2 shows the structure of the inverter control strategy.
which has a complex control system to achieve its functions. The symbol 𝑉𝑉𝑚𝑚𝑚𝑚𝑚𝑚 is the output of the MPPT and the
(1) In order to extract the maximum power from the PV array
symbol 𝑉𝑉𝑑𝑑𝑐𝑐 is the voltage of the DC-link capacitor. The 𝑉𝑉𝑑𝑑𝑐𝑐 is
under varying weather conditions, a control strategy
subtracted from the reference value 𝑉𝑉𝑚𝑚𝑚𝑚𝑚𝑚 and the result is sent
maximum power point tracking (MPPT) is applied to control
the PV array to work at the maximum power point. (2) The to the PI voltage controller. The output is the reference value
inverter transforms the output of DC power into the AC power 𝐼𝐼𝑑𝑑 ∗ of the d-q axis current. The symbol 𝐼𝐼𝑞𝑞 ∗ is determined by
the reactive power delivered to the grid. The value of 𝐼𝐼𝑞𝑞 ∗ is

2747

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.
 set to 0 since the grid-connected inverter usually operates in the linear combination of radial basis functions [20]. Due to
the unit power factor and only provides active power to the its simple structure, fast convergence rate and universal
main grid. optimal approximation ability, RBFNN is widely applied in
The current control of the inverter can limit the current fed the pattern classification, function approximation, data
from an IBDER during a SC fault via a current limiter, which mining, etc.
is utilized to avoid the breakdown of the semiconductor A radial basis function refers to any real-valued function,
devices in the inverter. After passing through the current whose value only depends on the Euclidean distance between
limiter, the reference current values are subtracted with the the input and some fixed point, i.e., centers. The generally
actual 𝐼𝐼𝑑𝑑 and 𝐼𝐼𝑞𝑞 , and the result is sent to the d-q axis used radial basis function is the Gaussian function represented
decoupling controller to generate the signals in the d-q in (3).
coordinate system. After a reverse Park’s transformation to ‖𝑥𝑥−𝑥𝑥𝑖𝑖 ‖2
transform the signals in the d-q coordinate system back into 𝜙𝜙(‖𝑥𝑥 − 𝑥𝑥𝑖𝑖 ‖) = exp ( ) (3)
2𝜎𝜎 2
the a-b-c coordinate system, the gate pulse control signal is
generated via PWM to control the semiconductor switches of where the symbol x is the input, the symbol 𝑥𝑥𝑖𝑖 is the center
the inverter. In order to reduce the impact of grid side voltage and the symbol 𝜎𝜎 represents the standard deviation, which is
disturbances on the current control, the feedforward control also defined as the RBF width parameter.
of 𝑉𝑉𝑞𝑞 and 𝑉𝑉𝑑𝑑 is added in the inner current loop, which can The function approximation is usually achieved by the
improve the dynamic response speed of the inverter. radial basis functions in the form of the RBFNN as (4),
C. Benchmark system 𝑦𝑦(𝑥𝑥) = ∑𝑁𝑁
𝑖𝑖=1 𝑊𝑊𝑖𝑖 𝜙𝜙(‖𝑥𝑥 − 𝑥𝑥𝑖𝑖 ‖) (4)
A benchmark of a distribution system with an integrated where the approximating function 𝑦𝑦(𝑥𝑥) is the output of the
PV system is developed in this paper for case study, as shown RBFNN, represented as a sum of N radial basis function unit,
in Fig. 3. and each unit is associated with a different center 𝑥𝑥𝑖𝑖 and
Load 1 Load 2 weighted by an appropriate coefficient 𝜔𝜔𝑖𝑖 . The weights 𝜔𝜔𝑖𝑖
Bus 1 Bus 2 can be estimated using the matrix method of linear least mean
If_pv squares (LMS), since the approximating function is linear in
CB1 CB2
Utility Grid Feeder1
PV
the weights 𝜔𝜔𝑖𝑖 .
Substation
The function approximation process is interpreted as a
R1
F
R2
RBFNN, whose structure is shown in Fig.4. The hidden layer
CB3
Load 3 applies nonlinear radial basis function units as neurons, which
Feeder2 can directly map the input vector without weights for
connection. The mapping from the hidden layer to the output
R3
layer is linear. That is, the output of the network is the linear
Fig. 3: Benchmark of a distribution system with a PV system weighted sum of the outputs of hidden units, where the
weights are adjustable parameters of the network. Therefore,
In this system, each feeder segment is protected by a CB the mapping from the input to the output of the network is
controlled by an overcurrent relay. The relays monitor the nonlinear, while the output of the network is linear. As a result,
current flowing through the CBs and activate the CB tripping the weights of the network can be solved directly by the linear
when the viewed current exceeds its setting [19]. When a fault equations, which can greatly speed up the learning rate and
occurs at point F, both of the fault currents are fed from the avoid the local optimal solution problems.
main grid and PV system simultaneously. The SC fault current
flowing from the main grid is large enough to be detected by RBF units
Relay R1, which will trip circuit breaker CB1 to isolate the ɸ1(x) Linear
fault from the main grid. However, the SC fault current If_pv x1
W1 output
⁝
flowing through the circuit breaker CB2 will be limited to a unit

much lower magnitude due to the current limiter of the PV ⁝ W2 y(x)

ɸi(x)
inverter, so that it cannot be detected by Relay R2. As a result, xn
⁝
the fault current If_pv will constantly feed into the fault location. WN
It can result in significant damage of costly electrical ɸN(x)
equipment such as the inverter and the transformer in the PV
Input Hidden Output
system. Most of IBDERs have a current limiter in the current layer layer layer
control loop of inverters. It can limit the fault current fed from
IBDERs and make this fault current difficult to be detected by Fig.4 RBFNN structure
overcurrent relays. Therefore, it is necessary to develop an The RBFNN training is to generate the optimal two sets of
intelligent protection algorithm to detect the low SC fault parameters including: (1) the parameters of hidden layers
current from the IBDER fast and effectively. including the center of each RBF units, the RBF width, and
the number of hidden layer neurons; and (2) the weights
III. PROPOSED PROTECTION SCHEME between the hidden layer and the output layer. Generally, the
To detect low SC current fed from an IBDER, an centers can be generated by an unsupervised training, such as
intelligent overcurrent protection algorithm applying the k-means clustering. With the determined centers, the RBF
RBFNN is proposed in this section. width can be determined via (5)
A. Radial Basis Function Neural Network (RBFNN) 𝜎𝜎𝑖𝑖 =
𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚
𝑖𝑖 = 1, … , ℎ (5)
√2ℎ
RBFNN is a three-layer neural network (including the
input layer, the hidden layer and the output layer), which uses

2748

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.
 Where the symbol 𝐶𝐶𝑚𝑚𝑚𝑚𝑚𝑚 is the maximum distance between the 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 =
𝑛𝑛
∑(
𝑥𝑥𝑖𝑖 −𝑥𝑥̅ 4
) (8)
selected centers and h is the number of hidden layer units. (𝑛𝑛−1)(𝑛𝑛−2) 𝑠𝑠

The connection weights of neurons between the hidden where the symbol 𝑥𝑥̅ is the data mean and the symbol s is the
layer and the output layer are determined by a supervised sample standard deviation.
method, i.e., gradient descent, which can be directly calculated
by the least mean square LMS through (6) (5) Kurtosis: Kurtosis is a measure of the shape of the data to
determine whether the data is peaked or flat with respect to a
𝑇𝑇
𝑾𝑾 = (𝜱𝜱 𝜱𝜱) 𝜱𝜱 𝒅𝒅 −1 𝑇𝑇
(6) normal distribution, as presented in (9).
where the symbols 𝑾𝑾 and 𝜱𝜱 are the wight and RBF unit 𝑛𝑛(𝑛𝑛+1) 𝑥𝑥𝑖𝑖 −𝑥𝑥̅ 4 (𝑛𝑛−1)2
matrix respectively, and the symbol d is the expected output 𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘𝑘 = ∑( ) −3 (9)
(𝑛𝑛−1)(𝑛𝑛−2)(𝑛𝑛−3) 𝑠𝑠 (𝑛𝑛−2)(𝑛𝑛−3)
vector in the training data.
The output values of the RBFNN in the training dataset are
B. Intelligent overcurrent protection algorithm defined as the healthy sample data labeled as (1, 0) and the SC
Fig. 5 shows the proposed intelligent overcurrent fault current sample data labeled as (0, 1). The obtained
protection algorithm based on the RBFNN. Through an offline trained RBFNN model is used to process the test current
training, the trained RBFNN can be implemented in the dataset. When this algorithm outputs “1” in the first value of
microprocessor of a relay on the connecting feeder of the PV the label, a fault is detected. Then the relay will send a tripping
system to intelligently detect the SC fault current fed from a signal to activate the CB tripping for isolating the fault.
PV system in real-time. Otherwise, the first value of the label stays with “0” that does
During system operation, the time series current signals not trigger any operation.
are sent to relays continuously from the current transformer IV. CASE STUDY
(CT). In order to generate a training dataset for the offline
training of the RBFNN, both of the healthy current sample In order to verify the effectiveness of the proposed
data during normal operation and SC fault current sample data intelligent overcurrent protection algorithm, a case study is
in a time interval of one cycle (16ms) are collected from the conducted in the developed distribution system benchmark in
recorded current data. This training dataset consists of five section II. A fault scenario with the proposed algorithm is
statistical features, which are explained as follows [21]. reported in this section. Their simulation results are compared
with that of the same scenario with the traditional overcurrent
Data acquisition relay.
A. Fault detection with traditional overcurrent relay
The fault detection with traditional overcurrent relays is
carried out in the developed distribution system benchmark
in Fig.3, where the voltage of the feeder 1 is 10.48kV, the
load on feeder 1 is 90kW and the output power of the PV
system is 46kW. The minimum current setting of a
traditional overcurrent relay is normally 1.5 to 2 times of
the current rating [7]. When a fault occurs at 0.5s at point F,
the SC fault current viewed by Relay R2 is shown in Fig. 6.
Fault Current data
3

2.5

Fig. 5 Intelligent overcurrent protection relay algorithm using

RBFNN
1.5
Current (p.u.)

(1) Range: Range refers to the difference between the

maximum and minimum value of a sample data. 1

(2) Mean value: The average value of a sample data is termed

as the mean value. 0.5

(3) Standard deviation: Standard deviation is a measure of the 0

degree of dispersion of the current sample data, as expressed 0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6

in (7).
Time (seconds)

Fig. 6 SC fault current fed from the PV system

𝑛𝑛 ∑ 𝑥𝑥 2 −(∑ 𝑥𝑥)2
𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠 𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑𝑑 = � (7)
𝑛𝑛(𝑛𝑛−1) It shows that during the fault, this SC fault current fed
from the PV system fluctuates and decreases to 1.6 per unit
where the symbol x is the sample data set and the symbol n is
(p.u.) due to the current limitation. It cannot be detected by
the sample number.
the relay with a setting higher than 1.6 p.u.. Even the relay
(4) Skewness: Skewness is a measure of symmetry, or more setting is lower than 1.6 p.u. and the fault is detected, the
precisely, the lack of symmetry of a time series signal, which operation of the time inverse overcurrent relay will be
is calculated in (8). sluggish with a large time delay due to the narrow margin

2749

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.
 between the SC fault current and the current setting. This selected in a one-cycle interval, say from 0.42s to 0.436s,
long-lasting SC fault can result in severe device damage. and a SC fault current data sample that is selected in a one-
cycle interval from 0.5 s to 0.516 s. Accounting four fault
B. Fault detection with proposed protection algorithm scenarios, eight data samples in total are included in the
Based on the proposed protection algorithm, the training training datasets, which have the extracted features,
dataset needs to be generated for the offline training of the including the range, mean value, standard derivation,
RBFNN. The SC fault current data flowing through CB2 in skewness and kurtosis. These features and labels listed in
one cycle time window is collected and recorded by Relay TABLE I are the inputs and outputs of the RBFNN,
R2 in real-time. Four fault scenarios are developed in this respectively. Using the training datasets, the RBFNN model
paper to generate four datasets of the SC fault current with is trained to approximate the feature mappings of the current
the Fault F on Feeder 1 located at 5 km, 10 km, 15 km and data in both normal and fault situations.
20 km away from Bus 2 in Fig. 3 (the line impedance is To test the overcurrent algorithm with the trained
given as 0.35 Ω/ km). Fig. 7 shows the current dataset with RBFNN, we developed four more fault scenarios with the
the fault location at 5 km from Bus 2 as an example. Fault F on Feeder 1 located at 4 km, 8 km, 12 km and 16 km
As shown in Fig.7, two data samples are selected from away from Bus 2. Similar to the process of generating the
each current dataset to generate the offline training datasets, training datasets, eight test data samples including four
including a healthy current data sample that is randomly healthy current samples and four fault current samples are

Fault Current data -- 5 km Fault Current data -- 4 km

3 3

2.5 2.5 Fault Detected

Healthy Fault + CB Trip
dample data sample data
2 2

1.5
Current (p.u.)

1.5
Current (p.u.)

1 1

0.5 0.5

0 0
0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6 0.4 0.42 0.44 0.46 0.48 0.5 0.52 0.54 0.56 0.58 0.6

Time (seconds) Time (seconds)

Fig. 7 Current sample data Fig. 8 Test of trained RBFNN

TABLE I. TRAINING DATASET

Training data
Features
SC fault current data Healthy current data
Range (p.u.) 1.589 1.590 1.586 1.588 1.039 1.038 1.039 1.038
Mean value (p.u.) 1.502 1.500 1.500 1.518 0.125 1.500 1.518 0.125

Standard deviation (p.u.) 0.133 0.131 0.131 0.133 0.024 0.022 0.025 0.025

Skewness -2.916 -3.130 -3.039 -3.105 -0.058 0.036 -0.082 -0.163

Kurtosis 18.926 20.006 19.602 20.136 2.613 2.634 2.796 2.700

Label (1, 0) (1, 0) (1, 0) (1, 0) (0, 1) (0, 1) (0, 1) (0, 1)

TABLE II. TEST RESULTS

Test current data

Features
SC fault current data Healthy current data
Range (p.u.) 1.591 1.587 1.589 1.586 1.038 1.037 1.039 1.038
Mean value (p.u.) 1.503 1.531 1.503 1.531 0.139 0.131 0.133 0.143

Standard deviation (p.u.) 0.135 0.135 0.129 0.133 0.025 0.023 0.0230 0.025

Skewness -2.861 -2.988 -3.195 -2.863 0.043 0.089 0.062 0.036

Kurtosis 18.256 19.050 21.245 18.767 2.804 2.757 2.967 2.849

Output (1, 0) (1, 0) (1, 0) (1, 0) (0, 1) (0, 1) (0, 1) (0, 1)

Signal for tripping 1 1 1 1 0 0 0 0

2750

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.
 selected from the currents of these four fault scenarios. The [6] S. Rong and L. He, “Impact of Integration of Photovoltaic Generation
features of these current samples are abstracted and then on Protection of Distribution Systems”, in Proc. MIT A+B Applied
Energy Symposium, 2020. Accepted.
function as inputs of the RBFNN. Using the trained RBFNN, [7] J. L. Blackburn and T.J. Domin. Protective relaying: principles and
the label outputs indicating whether the selected test data applications.3rd ed. CRC press, Boca Raton, FL, U.S., 2006.
samples are in healthy or fault conditions are determined. [8] P. Mahat, C. Zhe, B. Bak-Jensen, and C. L. Bak, “A simple adaptive
The corresponding tripping signals sent by Relay R2 during overcurrent protection of distribution systems with distributed
these fault scenarios are also recorded. All the features, the generation,” IEEE Trans. Smart Grid, vol. 2, no. 3, pp. 428–437, Sep.
2011.
label outputs of these test data samples, and the tripping
[9] M. S. Sanchdev and R. Das. “Understanding microprocessor based
signals are reported in TABLE II. The detection results technology applied to relays”, IEEE power system relaying committee,
show that all the four SC fault current data samples are Rep. WG I01, Jan. 2009.
detected accurately with the outputs (1,0), and these faults [10] A. Conde and E. Vazquez, “Operation logic proposed for time
are successfully isolated by sending the trip signal “1” to overcurrent relays,” IEEE Trans. Power Del., vol. 22, no. 4, pp. 2034–
2039,Oct. 2007.
CB2. The healthy current data samples are also detected
with the label output of (0,1) and the tripping signal of 0, [11] D. S. Kumar, D. Srinivasan and T. Reindl, "A Fast and Scalable
Protection Scheme for Distribution Networks With Distributed
which leads to no operation of CB2. One of fault test Generation," IEEE Trans. Power Del., vol. 31, no. 1, pp. 67-75, Feb.
scenarios are reported in Fig. 8 as an example, i.e., the fault 2016.
current flowing through CB2 with the proposed algorithm [12] A. D. Oliveira, L. R. M. Silva, C. H. Martins, etc., “An improved DFT
applied in Relay R2 and Fault F located at 4 km away from based method for phasor estimation in fault scenarios,” in Proc. IEEE
Bus 2. It is seen that the fault current became 0 at 0.52 s that Pow. Eng. Soc. Gen. Meet., San Diego, CA, USA, Jul. 2012, pp. 1–5.
is about one cycle after the fault occurrence at 0.5 s with [13] H. Wen, S. Guo, Z. Teng, and F. Li, “Frequency estimation of
distorted and noisy signals in power systems by FFT-based approach,”
together the calculating time delay. It is indicated that this IEEE Trans. Power Syst., vol. 29, no. 2, pp. 765–774, Mar. 2014.
SC fault is successfully detected and isolated. [14] A. Agrawal, M. Singh and M. V. Tejeswini, "Voltage current based
The offline training time for the RBFNN is 0.414 time inverse relay coordination for PV feed distribution systems,"
seconds and the detection time of the proposed algorithm NPSC, Bhubaneswar, 2016, pp. 1-6.
using the trained RBFNN is 0.0136 seconds (less than one [15] N. Mahmud and A. Zahedi. “Review of control strategies for voltage
cycle) on a lab platform including an intel core i7 2.6GHz regulation of the smart distribution network with high penetration of
renewable distributed generation”, Renew. Sust. Energy Rev., vol. 64,
CPU, an NVIDIA GTX 950M GPU, and 8GB RAM. This pp. 582-595, Oct. 2016.
algorithm can be easily implemented in the microprocessor [16] C. M. Bishop, Pattern Recognition and Machine Learning, Springer
relays. The simulation results demonstrate that the proposed Science Business Media, New York, NY, USA, 2006.
intelligent overcurrent protection algorithm can detect the [17] X. H. Nguyen and M .P. Nguyen, “Mathematical modeling of
SC faults in distribution system with IBDERs fast and photovoltaic cell/module/arrays with tags in Matlab/Simulink”,
Environmental Systems Research., pp. 1-13, 2015.
accurately. [18] A. V. Sant, V. Khadkikar, W. Xiao, H. Zeineldin and A. Al-Hinai,
"Adaptive control of grid connected photovoltaic inverter for
V. CONCLUSIONS maximum VA utilization," in Proc. Annu. Conf. IEEE Industrial
This paper proposes an intelligent overcurrent Electron. Soc., Vienna, 2013, pp. 388-393.
[19] M. Kezunovic, J. Ren and S. Lotfifard, Design, Modeling and
protection algorithm applying the RBFNN to learn and Evaluation of Protective Relays for Power Systems, Springer
detect the SC fault current fed from IBDERs. This fault International Publishing, Switzerland, 2016.
current is limited by the inverter control and therefore [20] J. Park and I. W. Sandberg, “Universal Approximation Using Radial-
difficult to be detected by a conventional overcurrent relay. Basis-Function Networks”, Neural Compu., vol. 3 no. 2, pp. 246-257,
A case study shows that the proposed algorithm can detect 1991.
the low fault current fed by an IBDER in a distribution [21] P. K. Kankar, S. C. Sharma and S. P. Harsha, “Fault diagnosis of ball
bearings using machine learning methods”, Expert Syst. Appl., vol. 38,
system fast and accurately. The proposed algorithm can be no. 3, pp. 1876-1886, Mar. 2011.
easily implemented in a digital relay for online detection of
the low fault current. In the future, we will develop an
advanced intelligent protection scheme for distribution Lina He received the Ph.D. degree in electrical engineering
systems with IBDER integration. A deep learning method at University College Dublin, Ireland in 2014, and the B.S.
and M.S. degrees from Huazhong University of Science
will be applied to improve the robustness and universality and Technology, China, in 2007 and 2009, respectively.
of the protection algorithm. The effectiveness of this She is currently an Assistant Professor in department of
algorithm will be verified in the real-time simulator OPAL- electrical and computer engineering, University of Illinois
RT. at Chicago, Chicago, USA. She was a Project Manager and
Senior Consultant at Siemens headquarter in Germany and Siemens US from
REFERENCES 2014 to 2017. She was also an Electrical Engineer at State Grid Corporation
of China from 2009 to 2010. Her research interests include modeling, control
[1] "EU Energy Policy to 2050: Achieving 80-95% emissions and protection of power electronics based power systems, HVDC control and
reductions," European Wind Energy Association, pp. 68, Mar. 2011. operation, renewable energy integration, and wide-area protection and
[2] “Global DER Overview: Market Drivers and Barriers, Technology cybersecurity.
Trends, Competitive Landscape, and Global Market Forecasts”,
Navigant Research, Chicago, IL, U.S.. June 2019. Shuaiang Rong received the B.S degree in Automation in
[3] J. Driesen and F. Katiraei, "Design for distributed energy resources," 2015 and M.S. degree of electric engineering in 2018 both
IEEE Power Energy Mag., vol. 6, no. 3, pp. 30-40, June 2008. from Shanghai University of Electric Power. She is
[4] J. Keller and B. Kroposki, “Understanding Fault Characteristics of currently pursuing Ph.D. degree in electrical engineering
Inverter-Based Distributed Energy Resources,” National Renewable at the University of Illinois at Chicago, Chicago, USA. Her
Energy Laboratory (NREL), Golden, CO, U.S.. Rep. TP-550-46698, research interest includes machine learning, renewable
January 2010. energy integration and power system protection.
[5] A. Hooshyar and R. Iravani, "Microgrid Protection," Proceedings of
the IEEE, vol. 105, no. 7, pp. 1332-1353, July 2017.

2751

Authorized licensed use limited to: University of Illinois at Chicago Library. Downloaded on December 28,2020 at 23:11:50 UTC from IEEE Xplore. Restrictions apply.