You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330754690

Adaptive Protection Combined with Machine Learning for Microgrids

Article  in  IET Generation, Transmission and Distribution · March 2019


DOI: 10.1049/iet-gtd.2018.6230

CITATIONS READS

98 5,566

6 authors, including:

Hengwei Lin Kai Sun


Aalborg University University of Tennessee
7 PUBLICATIONS   227 CITATIONS    278 PUBLICATIONS   5,800 CITATIONS   

SEE PROFILE SEE PROFILE

Zheng-Hua Tan Chengxi Liu


Aalborg University Shanghai Jiao Tong University
294 PUBLICATIONS   5,966 CITATIONS    5 PUBLICATIONS   215 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Control of Modular UPS View project

Fault Management of Inverter-Based Microgrids View project

All content following this page was uploaded by Josep M. Guerrero on 11 February 2019.

The user has requested enhancement of the downloaded file.


Adaptive Protection Combined with Machine Learning for Microgrids

Hengwei Lin 1,3, Kai Sun 1, Zheng-Hua Tan 2, Chengxi Liu 3, Josep M. Guerrero 3*, Juan C. Vasquez 3
1
Department of Electrical Engineering, Tsinghua University, Beijing, China
2
Department of Electronic Systems, Aalborg University, Aalborg 9220, Denmark
3
Department of Energy Technology, Aalborg University, Aalborg 9220, Denmark
*
corresponding author: joz@et.aau.dk

Abstract: This paper presents a rule-based adaptive protection scheme using machine learning methodology for microgrids
in extensive distribution automation. The uncertain elements in a microgrid are first analyzed 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 utilizes 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.

mho-based characteristics is proposed for a microgrid in [7].


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

1
proposed for microgrids in distribution system [19]. The systems. The relay settings and operating modes can be
calculation of the overcurrent protective settings is treated as reconfigured automatically for different operating states.
an optimization problem, while the system configuration can Moreover, the proposed scheme which carried out on two
be updated and modified based on the state recognition. The different models illustrates the effectiveness for electrical
state estimation algorithms are proposed with synchronized application.
phasor measurements for protection system in [20]-[21]. For This paper is organized as follows. Section 2
modern power system and future smart grid, the collected describes the test microgrid model. The data generating and
huge amount of valuable data may become possible to be processing are introduced in detail in this section. The data
utilized directly to obtain/train the mathematical model. This mining results reveal that the operating states influence the
tendency has been of concern as the electrical system keeps protection selectivity. In Section 3, the time-domain
evolving toward to an intelligent entity. simulation and the calculation of protective settings are
Machine learning is such a methodology that given. In Section 4, as a hybrid approach, ANN-SVM model
concerned with the pattern recognition and computational is integrated into the rule-based adaptive protection scheme
learning theory in artificial intelligence (AI) [26]-[29]. It for microgrid so that the system can adapt to the variable
may be inspiring for protection and control to combine with operating states with a “self-learning” and “self-training”
machine learning to fulfil the object of intelligent capability. Finally, the conclusion is given in section 5.
operation/control in smart grid, which may undergo a
process from huge data to meaningful information to 2. Data Generating and Processing
massive intelligence. The integration for microgrids may
contain distribution management system (DMS) where 2.1. Test Model
information layer allows for DA functions.
Therefore, this paper proposes a brand new adaptive The proposed methodology requires database
protection strategy combined with machine learning for available beforehand, which is obtained through simulation
medium voltage (MV) microgrids. The focuses of this paper in this paper due to the confidentiality from the utilities.
are threefold: 1. The analysis of uncertain elements in a However, it does not lose the representativeness since
microgrid is studied quantitatively by Pearson correlation simulation data are significant for power planning and
coefficients on database, which is the first time proposed for operation and other database in utilities can be considered in
microgrid protection and control system from a statistic practical application. The first test microgrid model is a part
view. 2. A rule-based adaptive protection scheme is of the distribution system owned by Himmerlands
proposed for microgrids in extensive distribution system. Elforsyning (HEF) in Aalborg, Denmark, which is a typical
This novel scheme can be integrated into distribution microgrid configuration on medium voltage level. This
automation system performing smartly and flexibly distribution system contains a combined heat and power
dependent on modern computer science. 3. The calculation (CHP) plant with three 3.3MW gas turbine generators
of protective settings is given both for overcurrent and (GTGs) and three small local wind farms. Each wind farm
distance protection that are widely used in distribution consists of two 2MW doubly fed induction generators

Bus1 Bus2 Bus3 Bus4 Bus5 Bus6

R12 R21 R23 R32 R34 R43 R45 R54 R56 R65
Line12 Line23 Line34 Line45 Line56 R60

R30 R40
60kV 20kV
Wind
R1 R2 Wind Farm1 Wind Farm2 Farm3
R61

G GTG1
CHP Plant

Line17
G GTG2
R17 R71
G GTG3
Collector Closed
Bus 7 Line16 Open
R16

Fig. 1. The test microgrid model in Aalborg.

2
Bus3  The faults with different resistance; (change from 0 to 5Ω
R30 with a variation of 1Ω every time, 6 cases in all. x2)
Wind Farm 1 Line  Microgrid operation modes; (grid-connected=1 and
Collector Bus islanded=0, realized by connecting/disconnecting R2, 2
cases in all. x3)
WTGXmr2 WTGXmr3
 Different topologies; (meshed=1 and radial=0, realized
WTGXmr1
by connecting/disconnecting relays on line16, 2 cases in
all. x4)
 Different output of DFIGs (the randomness of DGs);
(change from 0 to 100% of the rated value with a
DFIG G DFIG G Capacitor
variation of 25% every time, 5 cases in all. x5)
Bank  Load variation events; (change from 0 to 100% of the
rated value with an interval of 25% every time, 5 cases in
all. x6)
 Capacitor bank switching events; (change from 0 to
100% of the rated value with 5 steps in all. x7)
The features dataset X  [ x1, x2 L xN ] includes the above
Fig. 2. The configuration of the wind farm.
seven uncertain elements and matrix Y  [ y1, y2 L yN ] contains
(DFIGs) and one 0.8MW capacitor bank. The configuration the corresponding short-circuit measurements of voltage and
of microgrid and wind farm1 is given in Fig.1 and Fig.2, current (both magnitude and angle). Pearson correlation
respectively. More details of the model are given in [7], [16]. coefficient is adopted for quantitatively analyzing the
In the microgrid, there is a data concentrator located at the correlations between X and Y [24]. X and Y are transformed
substation working as a central human machine interface into normalized z-scores by the following equations:
(HMI) between the downstream network and the utility x x
operators. With the help of simple network time protocol z xi  i (1)
(SNTP) server or GPS or BeiDou Navigation Satellite sx
System (BDS) controlled inter-range instrumentation group where x̅ is the mean and sx is the standard deviation of X
(IRIG) time code generator, real-time clocking can be given:
guaranteed. Due to limited refresh rate of supervisory 1 N
x   xi (2)
control and data acquisition (SCADA) system for N i 1
microgrids in distribution system, the paper proposes an
1 N
independent on-line state diagnosis model integrated with
adaptive protection. The mathematical model is trained
sx   ( xi  x )2 (3)
N  1 i 1
dependent on database and can be integrated in the The Pearson correlation coefficient for the two
intermediate information layers or the control centers in datasets X and Y is given by:
power grid. The forward direction of the fault current is
1 N
defined from power grid side to the downstream buses in the rxy   zx z y (4)
N  1 i 1 i i
network (from bus1 to bus6), while the backward direction
is from bus6 to bus1. The above equation can also be reformulated as the
covariance between X and Y, normalized by the product of
the standard deviations of X and Y:
2.2. Data Generating and Data Mining N

 ( x  x )( y
i i  y)
rxy  i 1
(5)
1 N 1 N
An electrical power system suffers unexpected faults
on power lines for various random causes with nonlinear 
N  1 i 1
( xi  x ) 2 
N  1 i 1
( yi  y ) 2
function. These failures may hinder the stable operation of
where rxy ∈[-1,1] . The minus sign in the correlation
the system. Therefore, it is necessary to analyze the
influences of the uncertain elements before designing the coefficient means that the measuring parameter decreases
protection schemes. This paper proposes to use data mining with the increasing uncertain element. The more the
methodology to quantitatively extract the meaningful absolute value of 𝑟𝑥𝑦 is, the higher the influence of the
information from the database, which is not the case in the uncertain element on this measurement is.
previously reported reference. The uncertain elements (data) Each variation of the uncertain elements in the
are performed as is shown below: analysis is simulated separately for a 3-phase short-circuit
 The three-phase faults with different locations on one line on Line23 in DIgSILENT PowerFactory. The simulation
segment; (change from 0 to 100% of the rated value with contains 18000 times calculation for all the cases
a variation of 20% every time, so 6 cases in all. x1) (6×6×2×2×2×5×5×5=18000), while power flow calculation
and complete short-circuit analysis are performed for both
pre-fault and post-fault in each specific case.

3
Table 1
Correlation between the Uncertain Elements and the Voltage
Voltage Amplitude Amplitude Angle Angle
(Forward direction) (Backward direction) (Forward direction) (Backward direction)
Fault Location 6.5×10-4 -7.0×10-4 -0.06 -0.001
Fault Resistance 0.933 0.9335 0.225 0.514
Operating mode 0.079 0.075 0.173 0.210
Topology 0.078 0.085 0.173 0.260
DFIG 6.9×10-4 9.4×10-4 2.8×10-4 0.001
Load 5.9×10-4 8.0×10-4 5.0×10-4 0.001
Capacitor Bank 0.002 0.003 -0.004 -0.009

Table 2
Correlation between the Uncertain Elements and the Current
Current Amplitude Amplitude Angle Angle
(Forward direction) (Backward direction) (Forward direction) (Backward direction)
Fault Location -0.001 0.001 3.0×10-4 -3.0×10-4
Fault Resistance -0.728 -0.945 0.944 0.908
Operating mode 0.393 -0.155 0.068 0.126
Topology 0.288 -0.003 0.064 0.192
DFIG -5.3×10-4 0.002 0.002 0.002
Load -4.7×10-4 0.002 0.001 0.002
Capacitor Bank 6.6×10-5 0.005 0.005 -0.013

The simulation results are all recorded separately in higher power rating than the microgrid itself. Meanwhile,
matrix X and Y as mentioned above and used further for the network topology influences the current (in Table II)
training the state recognition models in Section 4. Therefore, more than the voltage (in Table I). Therefore, the design and
this part can be treated as a part of the data preprocessing for calculation of overcurrent protection should consider the
the adaptive protection scheme. Based on (5), the calculated influences in the system. On the other hand, the other
correlations between the uncertain elements (in X) and the variations have almost the same correlation coefficients that
post-fault measurements (in Y) are given in Table 1 and are relatively small. However, the intermittence and
Table 2. connection states of DGs should be considered. Some
Although the correlations are linear statistical values storage units may be required around the local power plants
depending on the specific network, the results reveal to avoid the complete loss of DFIGs’ support during a fault.
quantitatively some common phenomenon in microgrids and Another countermeasure is to collect the real-time states of
more cases or events can be considered in practical DGs to the control center so that it enables the modification
applications. In Table 1, it can be observed that the fault of the protective settings of relays based on the current
impedance, the operating mode (the connected/islanded operating situation.
mode), the topology of the network, and the fault location
have significant influence on the voltage and current (both 3. Protection Calculation and Test
forward and backward direction on line23) in descending
order. It shows that the operating mode and the network 3.1. Time Domain Simulation on Different
topology should be operated under cautious planning. The Operating States
backward voltage angle has relatively large coefficients
between the operating mode and the topology (0.210 and In order to reveal the influence of the operating mode
0.260), which mean the angle stability and the critical clear and topology, time domain simulation (electromagnetic) is
time (CCT) may become serious during a fault. This is due studied for different operating modes in this subsection.
to the capacity and the mechanical inertia of backward wind There are mainly four operating modes considered for the
farm is limited. Meanwhile, the capacitor bank also has microgrid test: grid-connected radial mode, grid-connected
certain influence on the measurements. It shows that the looped mode, islanded radial mode and islanded looped
reactive compensators at proper locations are significant for mode, which are controlled by R2, R16 and R61 as is shown
the stability of the network. in Fig.1. In the proposed adaptive protection scheme (in
In Table 2, the correlation coefficient between the section 4), the connection states of DGs are collected
forward current amplitude and the operating modes are periodically to the control center. The microgrid is normally
relatively high (0.393). It implies the short-circuit current operated in grid-connected radial mode to avoid power
changes severely between the grid-connected mode and oscillations in the looped/meshed grids during a fault. If
islanded mode, because the transmission grid has much
4
4.5
circuit with 0.2Ω fault resistance occurs on line45 at 0ms
4 grid-connected looped
islanded looped for the four operating modes. It can be found that the fault
3.5 grid-connected radial current and voltage change severely in different operating
islanded radial
states, which may influence the coordination of the relays.
Forward Current (kA)

2.5
If the fault can not be cleared fast, oscillation may exist in
the network (in islanded looped mode) which may
2
influence the system stability (e.g. lose synchronism). It is
1.5 mainly due to the power oscillation among the CHP and
1 local wind plants through line16. The simulation results
0.5
illustrate that even a same fault on the system may lead to
different consequences and measurements for different
0
0 0.05 0.1 0.15 0.2 0.25
Time (s)
0.3 0.35 0.4 0.45 0.5 operating modes. Therefore, it is imperative to calculate
the protective settings based on the specific operating
4 states so that the fault can be cleared selectively and
3.5
grid-connected looped quickly.
islanded looped
grid-connected radial
3 islanded radial 3.2. Protective Settings Calculation for
Backward Current (kA)

2.5 Overcurrent Relays


2
Since it has been known that the short-circuit
1.5 measurements are more related to the operating modes,
1 topologies of the network and the connection states of DGs,
the relay settings should be modified based on the current
0.5
operating situations. As is illustrated in Fig.5, the
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
transmission grid and CHP bring infeed currents for the
Time (s)
backward relays in looped mode (with R16 and R61
Fig. 3. The fault current magnitude for different operating closed), whose short-circuit capacity are much larger than
states. the DGs. Meanwhile, the different connection states of
DGs also influence the infeed current on the network. In
1
1 this paper, the calculation algorithm is given for directional
0.9 0.9 grid-connected looped overcurrent relay (DOR) and distance relay, respectively.
islanded looped
0.8 0.8 grid-connected radial However, the calculation algorithms for protective settings
0.70.7
islanded radial are not only limited on the presented ones in this paper.
Forward Voltage (p.u.)

For directional overcurrent relays, the coordination


0.60.6
of protective settings can be treated as a linear
0.50.5
programming problem for minimizing the total function
0.40.4
time. Fig. 6 gives the characteristic of DOR provided by
0.30.3
international electrotechnical commission (IEC). The
0.20.2 characteristic (IEC255-3) is given below:
0.10.1 A  TDS
000 t B (6)
 If 
0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
   1
Time (s)

11  Ip 
0.90.9 grid-connected looped
islanded looped where: A is a constant
0.80.8 grid-connected radial
islanded radial
B is a characteristic index
Backward Voltage (p.u.)

0.70.7 If is the fault current magnitude


0.60.6
Ip is the pickup current
0.50.5
TDS is the time dial setting
0.40.4 The coordination algorithm is based on optimizing
0.30.3 an objective function of operating times of the primary
0.20.2 relays while subjected to maintaining the operation of
0.10.1 backup relays coordinated. The objective function for a
000 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
system that contains m relays can be formulated as:
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Time (s) f max m

Fig. 4. The fault voltage magnitude for different operating z  Min  ( (ti ,i   ti , j )) (7)
f 1 i j
states.
where: f indicates the fault location.
there is a fault or emergency in the system, the operating ti,i is the operating time for the primary relay located
mode can be reconfigured based on the online decision at i for a fault near at i.
making. ti,j is the operating time for the backup relay located
Fig.3 and Fig.4 present the time-domain simulation at j for the same fault near at i.
results (electromagnetic transient) when a three-phase short- The operating time ti,j for the backup relays should be
5
Zone 1
Bus1 Bus6
Current
Line16 Line56 Measurement
Trip
R16 R61 R65 Signal
I1 I3 Zone 2 Delay 2 OR
Grid/CHP

Voltage
Measurement
Wind Farm3
Zone 3 Delay 3
I2
Fig. 5. Infeed current on Line56 under looped operation
Fig. 7. The block diagram of time-stepped distance relay.
modes.

Forward
10 1.00s
0.80s

1 0.60s
Time (s)

L
0.40s
0.20s
0.1 0
S Bus7 Bus1 Bus2 Bus3 Bus4 Bus5
0
If max 0.20s
0.01 2 3 4 5
10 10 10 10 0.40s
Current (A) 0.60s
Fig. 6. The characteristic curve of DOR. 0.80s
1.00s
greater than ti,i , which can be expressed as following: Backward
R34 R43
R71 R17
ti , j  ti ,i  t (8) R12 R45 R21 R54
R23 R56 R32 R65
where ∆t is the coordination time interval (CTI) and set as
constant. Fig. 8. The protective settings for normal operation mode.
When Ip and If are confirmed through power flow
calculation and short-circuit analysis, TDS of each relay can coefficient of each protection zone, which is often set
be obtained from (7) as a linear programming problem. To between 0.8 and 0.85. Zline is the impedance of the primary
I II
keep the reliable operation, Ip can be set as 1.5 times of the protected line. Zset2 and Zset2 are the protective settings of
maximum anticipated load current in the next several hours Zone 1 and Zone 2 of the downstream relay’s, respectively.
or days. More details of DOR settings calculation can be To ensure the protection selectivity, the delay settings
found in [19], [23]. of time-stepped distance relays are set at 0.35s for Zone2
and 0.7s for Zone3 respectively. The pickup time of the
3.3. Protective Settings Calculation for Distance relay is set to 20ms. The calculated protective settings for
Relays grid-connected radial mode with full DGs connection are
shown in Fig.8. The test of the distance protection can be
For simplicity, Ohm-based distance protection (with found in [7] and [22]. For extreme fault cases in power
three time-stepped zones) is chosen for its good selectivity grids, overcurrent protection, distance protection and/or
and sensitivity. As is shown in Fig.7, Zone1 is an differential protection can be activated together in modern
instantaneous protection, while Zone2 and Zone3 have relays e.g. the overcurrent protection can be used as the
definite time-stepped delay. The measuring impedance 𝑍𝑚 is backup protection scheme for distance protection and/or
calculated from the measuring voltage 𝑈̇𝑚 and current 𝐼𝑚̇ at differential protection in electrical systems.
the relay point, which is shown in equation (9).
𝑈̇𝑚 4. Rule-based Adaptive Protection
𝑍𝑚 = ̇
(9)
𝐼𝑚
The settings of each relay can be calculated as 4.1. Adaptive Protection with State Recognition for
follows: Microgrids
 Z setI  K relI Z line
 II The above calculation results of relay settings can be
 Z set  K rel ( Z line  K b Z set 2 ) (10)
II I
recorded in the control center or relays beforehand, or
 III
 Z set  K rel ( Z line  K b Z set 2 )
III II
calculated online, which is dependent on the specific
where the super script in the equations represents the requirements in the electrical system.
different protection zone. Kb is the branch coefficient that
offsets the infeed currents, which can be obtained from
I II III
short-circuit calculation. Krel , Krel , Krel are the reliability

6
scheme periodically, which works in an “interrupt mode”.
Begin
Once the interrupt is accepted by the processor, tripping
Tripping No signal is retrieved first (tripping action has the highest
action? priority) to confirm whether the system operates in normal
Yes
state or fault state. If a tripping action occurs, control center
Fault event Normal event
begins to retrieve the measurements automatically and sends
Redo
Retrieve the measurements them to the ANN model to estimate whether there is a fault
Retrieve the connection
states of DGs and which line is faulted. If the model confirms there is a
Estimate in
ANN permanent fault on the line, fault location estimation is
No
State No followed using SVM. Since the fault has been confirmed on
a fault occurs? changed?
the specific power line, fault location estimation is limited
Yes Yes Reclose Yes
Estimate in the breaker on the tripped line to enhance the accuracy. For redundancy
No
Times N? SVM Retrieve the
Decision Table
and further “self-learning” and “self-training”, the results
Counting
LSE No from ANN-SVM can work together with LSE. Then a
available?
No
decision table is retrieved whether it is needed to
Yes Modify protective
Record results
of ANN-SVM
settings? reconfigure the network and the corresponding relay
Compare with LSE
Yes settings, depending on the pre-defined settings or on-line
No
Bias A? On-line enhanced decision making. If it is required to reconfigure
Yes
Calculation the topology, activation signal is sent to the corresponding
Record the on-filed relays. Some shedding actions and/or on-line
mean value
Activate relevant calculation may be required in some situations. On the other
Alarm
on-filed actuators hand, if the system is operated in normal state, the
connection states of DGs are retrieved from each local wind
farm center or DG. According to the connection states of
Enhanced decision making DGs, the protective settings are modified as similar as in
fault state.
Record all actions with exact time The proposed scheme requires the parameters and
knowledge of the network. To train the hybrid estimation
Over model, power flow calculation and complete short-circuit
Fig. 9. The flow chart of the rule-based adaptive protection analysis both for pre-fault and post-fault are required. In this
scheme. experiment, 108000 (18000×6lines) cases with a 3-phase
short-circuit are performed in DIgSILENT PowerFactory,
Due to the limited refresh rate of current SCADA which is realized as described in Section 2. The fundamental
system and the paucity of measurements in some frequency component of the voltage and current phasors
microgrids, real-time least square estimation (LSE) usually (amplitude and angle) are calculated in the simulations,
are not available for microgrid protection. When a fault which are finally set as the inputs of the estimation model.
occurs, the operators in workstation often have to retrieve Moreover, the proposed scheme enables the self-learning
the on-field measurements manually or even go to the site. capability for improvement if various new data streams are
Meanwhile, the variable connection states of DGs in the available during the operation and (or) test, which is the
network influence the selectivity of protection. Another advantage comparing with the conventional methods.
phenomenon may rise in power grids that some mobile
DERs and electric vehicles (EVs) are connected to the grid 4.2. ANN Model for Classification
as DGs with a feature of changing the operating modes
among generators, loads and compensation equipment on As a hybrid approach, the ANN-SVM estimation
some locations sometime. The above mentions require an model is critical in the proposed scheme. The training and
alternative approach to recognize the state of downstream test of ANN are implemented in MATLAB while SVM is
network and make decision automatically for a computer- performed in LIBSVM. Out of the data available, 30% are
based control system. As the huge historical and (or) randomly selected as the test set while the remaining data
simulation data have become possible in power system, the are used for the model training for both models.
paper utilizes the methodologies in machine learning to train ANNs are a class of powerful non-linear machine
the mathematical model such that the protection and control learning predicted class from the ANN model and each row
can adapt to the variable operating states. Furthermore, an represents models [27], [28]. Among ANN models, the feed-
integrated rule-based scheme is adopted, which has forward neural network, also known as the
distinctive advantages of applying on modern industrial multilayerperceptron (MLP), is most successful. As can be
computers. The simulation data is proposed as a partial seen in Fig.10, each layer of MLP contains numbers of
source of the machine learning based model to implement neurons that take inputs from the previous layer neurons.
smart grid type protection and control. The implementation The excitation signals are fed into the input layer. Each
may consist of a DMS installation at distribution substation neuron is a logistic regression model that acts as a non-
which combines SCADA system and bus/feeder linear function of its inputs (xi) and generates an output (y)
measurements. as an input for the next layer or as a final output, which can
The flow chart of the proposed scheme is given in be found in Fig.11. The neurons are connected by links, and
Fig.9. The processor/computer in the control center (or the each link has a weight (w) associated with it. The link
intermediate level) processes the proposed protection weights and the biases of neurons are adjusted according to
7
Table 3
The 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%

y1 yk yc
Accuracy of ANN
Output ... ... f ()
100%
wkj 80%
z1 zj zn
Hidden ... ... h()
60%
w ji
40%
20%
Input ... ... 0%
x1 xi xd
Numbers of Neurons
Fig. 10. A simple ANN illustration.
Fig. 12. The accuracy of ANN only with measurements
Dendrites on bus1 and bus6.
(Inputs) Synapses Inputs
(Outputs) x1
x2 Table 4
u Output y
+ f (u )
...

Activation Function The Error of ANN with full measurements


Cell Body dx w
Axon
(Summation of Layers error for test error for all
(Activation Function)
Weighted Inputs)
cases cases
Fig. 11. A single neuron model.
2 0 0
the training data set during the training process. After 3 0 0
training, the ANN is ready to be used for the testing data or 4 0 0
data from real-world applications. The ANN model is 5 0 0
determined by its network architecture, transfer function and
learning rules [29]. The output neuron model (Fig. 11) as a where h(.) is the differentiable, nonlinear activation
function of d inputs can be expressed as: functions. D and M are the numbers of neurons on different
layers.
d
Back propagation learning rule is adopted to adjust
y (x, w)  f (u )  f ( w j j (x)) (11) the weights and biases [29]. The problem of finding a
j 1
suitable architecture and the corresponding weights of the
where y is the output, φj (x) is the basis function and f (.) is network is of central importance in the area of training
the neuron activation function. u is the summation output ANNs.
signal: Used as a classification model, the ANN model in the
u  wT x (12) scheme has seven outputs with relevant labels to represent
where w=[w0 , w1 ,⋯ wd ] , x=[x0 , x1 ,⋯ xd ] . The logistic- the six power lines (line12, line23, line34, line45, line56
sigmoid transfer function is used for the signal delivering line16) and no fault. The input neurons for ANN Model are
between the neurons at two layers: the measurements (voltage and current) on different buses.
1 The number of hidden neurons is set as 64, while the
f (u )  (13) number of training epochs is set as 1000.
1  e  k (u  u0 )
Concerning the input neurons, two cases are
where u0 is the value of the midpoint of sigmoid, k is the considered for comparison (the overdetermined set of
steepness of the curve. The ANN’s capability of dealing measurements is available or not): only measurements from
with nonlinear and complex problems is realized by making bus1 and bus6 are available, and full measurements from the
the basis function φj (x) depend on parameters and then to six buses (bus1, bus2, bus3, bus4, bus5 and bus6) are
allow the parameters to be adjusted, along with the available. If the measurements from bus1 and bus6 are
coefficients {wj}, during training: available for training (the worst condition), the average
accuracy of ANN for all test cases is 90.3%, while the
M D
yk (x, w)  f ( wkj(2) h( w(1) xi )) (14) confusion matrix (error matrix) is given in Table III. Each
j 0
ji
i 0 column of the confusion matrix represents the instances in
8
an actual class. The accuracy of ANN is shown in Fig.12.
Error of SVM
From Table III and Fig.12, it can be found that the
accuracies of Line23, Line34 and Line45 are around 80%, 0.6%
while they are higher for Line12, Line 56 and Line16. In the
first scenario, the diagnosis model can not identify all the 0.4%
faults in the microgrid. It indicates that measurements on
bus2, bus3 or bus4, which is inside the microgrid, may be 0.2%
necessary to improve the estimation accuracy. Therefore, the
enhanced decision making is necessary for the system to 0.0%
main the reliable operation. For comparison, the ANN line12 line23 line34 line45 line56 line16
model is trained again with full measurements from all the Fig. 13. The accuracy of SVM Model.
six buses. The accuracy is shown in Table IV. It can be
found that the accuracy keeps at 100% (with 0 error) for Table 5
different hidden layers since the measurements (from all The Error of SVM
buses) used in this case implicate the direction for a fault on
Line Line Line Line Line Line
each feeder, which is similar as the absolute selectivity of
12 23 34 45 56 16
the differential protection. Moreover, the comparison results
of the test show that the performance of the learning-based Test
0.24% 0.31% 0.57% 0.44% 0.16% 0.1%
models can keep being improved with the increasing error
measurements and the relevant data.
In order to solve the training task, Lagrangian function and
4.3. SVM Model for Fault Location Lagrange multipliers are introduced so that (18) is
transformed into the dual problem:
n n
Used as a regression model in the scheme, the SVM
model outputs the estimated fault location on the power line. max{  i (i  i )    (i  i )
i 1 i 1
SVM is a supervised learning method for a linear model, (20)
1 n n
while it performs well for nonlinear problems by using   (i  i )( j   j )K (x i , x j )}
kernel trick. With the basis function 𝜑(x), the input vector x 2 i 1 j 1
in SVM is mapped into a higher D-dimensional feature subject to:
space [25]. The output y(x) in the high dimensional space is n

defined as (15):  (    )  0
i 1
i i
D
y (x)  w T  (x)   w j j (x) (15) 0  i  C (21)
j 0
0  i  C
where w=[w0 ,w1 ,⋯,wd ]T is the weight vector, ‖w‖2 <𝛾. 𝛾 is 𝑇
a constant. φ(x)=[φ0 (x),φ1 (x),⋯,φd (x)]T is the basis where K(xi , xj ) = φ (xi )φ(xj ) is an inner-product kernel.
Equation (21) can be finally solved as a quadratic
function vector.
programming optimization problem. The choice of the
Introducing the ε-insensitive loss function Lε (y(x),β):
coefficients ε and C is the critical problem during training
0, for   y(x)  
L ( y (x),  )   (16) SVM. Through adjusting and testing, the final parameters
   y (x)   , for   y (x)   used are C=1000 and ε=0.05.
where ε is a free parameter that serves as the threshold of the In the SVM model, there are six sub-models built
loss function. x is the input vector, β is the target value. The separately for the six power lines using the same
main task of training a SVM model is minimizing the error measurements in ANN. One of six the sub-models in the
function: SVM model will be activated once the fault is confirmed on
the corresponding power line in ANN. Full measurements
1 n
E   L ( y (x i ),  i ) (17) on each bus are used according to the test results in the
n i 1 section above. The following metric is used to assess the
where n is the number of learning data pairs (xi, βi). performance of SVM for fault location:
Considering the slack variables ζ and ζ' for the outliers in d  da
the data, the training task can be treated as: err  e 100% (22)
n
l
1
min{ w T w+C[ ( i   i )]} (18) where de is the estimated value, da is the actual value and 𝑙
2 i 1 is the length of the power line.
subject to: The test results for each sub-model in SVM Model
w T  (x i )  i     i are given in Table 5 and Fig.13. It can be found the error is
always within 1% for all the tests, which are acceptable for
i  w T  (x i )     i the requirements on the estimation tasks.
(19)
i  0 As is shown in Fig.9, after the state diagnosis in the
scheme, the corresponding decisions will be followed to
 i  0 decide whether reconfigure or not. Meanwhile, alarm is sent
C is an adjustable parameter that determines the to the utility operators. Some additional decisions can be
compromise between the model’s complexity and accuracy. considered further by the utility. Finally, the actuators
9
perform the reconfiguration automatically, while the related Bus 8
Bus 7 230 kV Bus 9
events are recorded with excite time. Bus 2 230 kV Load C 230 kV Bus 3
18 kV 13.8 kV
This proposed adaptive protection scheme is practical T2 T3
G2 G G G3
for smart grids when the penetration of DGs and microgrids 163.2MW Line 3 Line 4
108.8 MW
are increasing in distribution system, which requires more
intermediate layers controlling the downstream network via
Line 2 Line 5
various computers. Due to its flexible integration in the
protection and control system, the rule-based protection Bus 5 Bus 6
scheme reduces the complexity of fault retrieve for 230 kV 230 kV
Line 1 Line 6
distribution system operator (DSO) and enhances the Load A Load B
reliability of operation. With the help of SNTP server and/or Bus 4
230 kV
IRIG time code generator, synchronous phasor measurement T1
can be adopted for microgrids in distribution automation.
Bus 1
16.5 kV
4.4. Test on IEEE 9 Bus Model
G
G1
To verify the feasibility of the recognition model, the 247.5MW
test is performed on IEEE 9 bus standard model for its Fig. 14. The IEEE 9 bus Model.
modest configuration. The data is obtained with a smaller
step as following: Table 6
 A three-phase faults with different locations on one line The Error of ANN with different measurements
segment; (change from 0 to 100% of the rated value with Measurements Accuracy for test Accuracy for all
a variation of 10% every time, 11 cases in all) cases cases
 Several faults with different resistance; (change from 0 to Bus 5, 6, 8 7.18% 7.14%
10Ω with a variation of 1Ω every time, 11 cases in all) Bus 4, 7, 9 0.74% 0.62%
 Variation on Load1; (change from 0 to 100% of the rated All buses 0 0
value with an interval of 25% every time, 5 cases in all)
 Variation on Load2; (change from 0 to 100% of the rated Table 7
value with an interval of 25% every time, 5 cases in all) The Error of SVM
 Variation on Load3; (change from 0 to 100% of the rated Line1 Line2 Line3 Line4 Line5 Line6
value with an interval of 25% every time, 5 cases in all) Bus
The collected data (11×11×5×5×5×6lines=90750) are 0.25% 0.11% 0.04% 0.23% 0.14% 0.25%
5,6, 8
used to train the ANN and SVM model. 144 hidden layers
Bus
and 500 training epochs are set for ANN. There are seven 0.12% 0.2% 0.23% 0.03% 0.12% 0.02%
4,7, 9
outputs in ANN, which are fault on line1, line2, line3, line4,
line5, line6 and no fault. The error of ANN models with All
0.25% 0.21% 0.25% 0.23% 0.13% 0.22%
different collected data as the inputs is given in Table 6. It buses
can be found that the accuracy of ANN used data only from
Bus4-7-9 is better than Bus5-6-8 due to the connection of requires the original data (or prior knowledge) to hold the
branch line and transformer on Bus4-7-9. It implies the accuracy to avoid the subsequent influence on the final
measurements from Bus4-7-9 contain more information for results. Some data processes may be required before the
the state recognition. The error of ANN for case “all busses” practical application. In case of data corruption, some
is 0, because the used data here is ideal without corruption additional countermeasures should be considered. For other
from simulation cases for training and testing ANN. The diverse fault or event types, the data can be obtained in a
inherent characteristics of learning-based model, such as similar way from precise simulation or historical database in
ANN and SVM, should be taken into account for the a system. It should be emphasized that the machine learning
practical application during training and testing. On the based model can work together with the other mathematical
other hand, C=1000, ε=0.02 is set for SVMs in this test, models (LSE etc.) to improve the accuracy and
while the output is the estimated fault location for the compatibility. In the specific situations, the operators can
specific lines. As is shown in Table 7, all SVM models with also intervene the automatic process to maintain the reliable
different measurements data work well with small errors. operation. The performance of ANN-SVM can keep being
From the test results, it can be found the ANN-SVM model improved with various new coming data and knowledge,
can also recognize the states accurately for transmission which is also the excellent advantage of machine learning.
systems. Moreover, it shows that the ANN-SVM model can In practical application, massive variables and scenarios can
still work well with enhanced decision making when the be performed with certain accuracy in computers/machines
overdetermined equations are not available for LSE. for self-learning and/or self-training. With the development
The paper utilizes data mining technology first to of technology, the proposed methodology may play a role in
extract the meaningful information and knowledge in the protection and control systems in future smart grid.
electrical system, then the learning-based model is trained
and integrated into the automatic protection and control 5. Conclusion
system. It undergoes a process from huge data to meaningful In this paper, we proposed a rule-based adaptive
information to final intelligence. The proposed methodology protection scheme combined with machine learning
10
methodology for microgrids. The correlation between the [11] G. D. Rockefeller, C. L. Wagner, J. R. Linders, K. L.
uncertain elements and the short-circuit measurements was Hicks, and D. T. Rizy, “Adaptive transmission relaying
analyzed quantitatively by Pearson correlation coefficients concepts for improved performance,” IEEE Trans.
from a statistic view. The results reveal that the operating Power Delivery, vol. 3, pp. 1446-1458, 1988.
states and topology influence the short-circuit measurements [12] S. M. Brahma and A. A. Girgis, “Development of
and stability of the system significantly. An ANN-SVM adaptive protection scheme for distribution systems
diagnosis model was integrated into the adaptive protection with high penetration of distributed generation,” IEEE
scheme for state recognition in distribution automation to Trans. Power Delivery, vol. 19, pp. 56-63, 2004.
extract and utilize the information and knowledge from the [13] Snehamoy Dhar, Pradipta Kishore Dash, “Differential
growing massive data streams in power system. For reliable current-based fault protection with adaptive threshold
operation and system resilience, reconfiguration can be for multiple PV-based DC microgrid,” IET Renew.
performed automatically and intelligently to change the Power Gener., vol. 11, pp. 778-790, 2017.
topology of the network and update the protective settings. [14] Susmita Kar, Subhransu Rajan Samantaray, “Time-
With the assistance of machine learning, this paper has frequency transform-based differential scheme for
showed that the adaptive protection scheme can modify the microgrid protection,” IET Gener. Transm. Distrib.,
settings properly and intelligently for different operating vol. 8, pp. 310–320, 2014.
situations. At the same time, the integrated adaptive [15] A. Oudalov and A. Fidigatti “Adaptive network
protection system enables accurate information for further protection in microgrids,” International Journal of
intelligent operation/control in smart grid and energy Distributed Energy Resources, pp. 201-226, 2009.
internet, which has a “self-learning” and “self-training” [16] P. Mahat, C. Zhe, B. Bak-Jensen, and C. L. Bak, “A
capability. Simple Adaptive Overcurrent Protection of Distribution
Systems with Distributed Generation,” IEEE Trans.
6. References Smart Grid, vol. 2, pp. 428-437, 2011.
[17] Haughton D A, Heydt G “A linear state estimation
[1] R. H. Lasseter, “MicroGrids,” in Power Engineering formulation for smart distribution systems,” IEEE
Society Winter Meeting, vol.1, 2002. R. H. Lasseter, Trans. Power System, vol. 28, no. 2, pp. 1187-1195.
“MicroGrids,” in Power Engineering Society Winter 2013.
Meeting, vol.1, 2002. [18] Sohrab Mirsaeidi, Dalila Mat Said, “A protection
[2] L. Hengwei, L. Chengxi, J. M. Guerrero, J. C. Vasquez, strategy for micro-grids based on positive-sequence
and T. Dragicevic, “Modular power architectures for component,” IET Renew. Power Gener, vol.9, pp. 600-
microgrid clusters,” in Green Energy, pp. 199-206, 609, 2015.
2014. [19] L. Hengwei, J. M. Guerrero, J. Chenxi, “Adaptive
[3] P. M. Anderson, Power System Protection, New York: Overcurrent Protection for Microgrids in Extensive
Wiley, 1999, pp. 201–240. Distribution Systems” in 42nd Annual Conference of
[4] M. A. Anthony, Electric Power System Protection and the IEEE Industrial Electronics Society (IECON2016),
Coordination, New York: McGraw-Hill, 1995, pp. 342- Italy, Florence, pp. 4042-4047, Oct. 2016.
346. [20] J. Joe-Air, Y. Jun-Zhe, L. Ying-Hong, L. Chih-Wen,
[5] J. L. Blackburn, Protective Relaying Principles and and M. Jih-Chen, “An adaptive PMU based fault
Applications, New York: Marcel Dekker, 1998, pp. detection/location technique for transmission lines. I.
383–408. Theory and algorithms,” IEEE Trans. Power Delivery,
[6] N. Jayawarna, N. Jenkins, M. Barnes, M. Lorentzou, S. vol. 15, pp. 486-493, 2000.
Papthanassiou, and N. Hatziagyriou, “Safety analysis of [21] J. Joe-Air, L. Ying-Hong, Y. Jun-Zhe, T. Tong-Ming,
a microgrid,” in Future Power Systems, 2005. and L. Chih-Wen, “An adaptive PMU based fault
[7] L. Hengwei, L. Chengxi, J. M. Guerrero, J. C. Vasquez, detection/location technique for transmission lines. II.
“Distance Protection for Microgrids in Distribution PMU implementation and performance evaluation,”
System” in 41st Annual Conference of the IEEE IEEE Trans. Power Delivery, vol. 15, pp. 1136-1146,
Industrial Electronics Society (IECON2015), Japan, 2000.
Yokohama, pp. 731-736, Nov. 2015. [22] L. Hengwei, J. M. Guerrero, J. C. Vasquez, L. Chengxi,
[8] M. Dewadasa, “Protection for distributed generation “Adaptive Distance Protection for Microgrids,” in 41st
interfaced networks,” Ph.D. thesis, Dept. Electrical Annual Conference of the IEEE Industrial Electronics
Engineering, Faculty of Built Environment and Society (IECON2015), Japan, Yokohama, pp. 725-730,
Engineering, Queensland University of Technology, Nov. 2015.
Australia, 2010. [23] Chattopadhyay B, Sachdev M S, Sidhu T S. “An on-line
[9] M. Dewadasa, R. Majumder, A. Ghosh, and G. relay coordination algorithm for adaptive protection
Ledwich, “Control and protection of a microgrid with using linear programming technique,” IEEE
converter interfaced micro sources,” in Third Transactions on Power Delivery, vol. 11, pp. 165-173,
International Conference on Power Systems, pp. 1-6, 1996.
2009. [24] Benesty, Jacob, et al. “Pearson correlation coefficient,”
[10] H. Nikkhajoei, R. H. Lasseter, “Microgrid fault Noise reduction in speech processing. Springer Berlin
protection based on symmetrical and differential current Heidelberg, pp. 1-4, 2009.
components,” Power System Engineering Research [25] R. Salat and S. Osowski, “Accurate fault location in the
Center, 2006. power transmission line using support vector machine

11
approach,” IEEE Trans. Power Systems, vol. 19, pp.
979-986, 2004.
[26] B. Cannas, G. Celli, M. Marchesi, et al. “Neural
networks for power system condition monitoring and
protection,” Neuro computing, vol. 23, pp.111-123,
1998.
[27] Z.–H. Tan, “Hybrid evolutionary approach for
designing neural networks for classification,” IEE
Electronics Letters, vol. 40, no. 15, pp. 955-957, July
2004.
[28] Saha M M, Izykowski J J, Rosolowski E. Fault location
on power networks, Springer Science & Business
Media, pp. 395-399, 2009.
[29] Chris Bishop, Pattern Recognition and Machine
Learning, UK: Springer, pp. 232-240, 2006.

12

View publication stats

You might also like