Islanding Detection Grid-Photovoltaic System

Using Artificial Neural Network

Tirta Samuel Mehang Dedet Candra Riawan Vita Lystianingrum B. Putri
Departmen of Electrical Engineering Departmen of Electrical Engineering Departmen of Electrical Engineering
Sepuluh Nopember Institute of Sepuluh Nopember Institute of Sepuluh Nopember Institute of
Technology Technology Technology
Surabaya, Indonesia Surabaya, Indonesia Surabaya, Indonesia
[email protected] [email protected] [email protected]

Abstract—used photovoltaic grid distributed aims to Si, the energy is absorbed by hv, or so-called photon. Large
improve the reliability of a grid system. However, the enough photon energy to release electrons in the charge-free
application of photovoltaic grid can cause problems when part areas, resulting in free electrons and holes[7].
of the electrical distribution system is isolated from the main
supply and still be energized by Distributed Generation (DG). The power generated by the solar cells is strongly
The condition is called as a condition islanding. If the islanding influenced by the amount of light received by the strong
condition is not detected, the bus load will experience more solar cells. On Figure 1 shows the strong influence of the
stress disorder (over voltage) and a decrease in power quality. light generated power.

In this study, conducted islanding detection system based

methods Artificial Neural Network (ANN). The system will be Ic
subjected to some load conditions, including the conditions 1000 W/m2
powermatch, overvoltage and undervoltage with different
MPPT
cosphi value as training data ANN. The signal voltage at node Im - 2
750 W/m
bus load of PCC classified by ANN to distinguish the condition
of the system, whether or not the islanding islanding condition.
From the simulation, the ANN method can detect the condition 500 W/m2
or not islanding system.

Um U(V)
Keywords— Islanding detection,, photovoltaic, Artificial
Neural Network. Figure 1 The strong influence of solar radiation to power
cell.
I. INTRODUCTION
In the event of termination of the main plant located on B. Inverter
the grid resulted in load will be supplied entirely by The basic principle of the use of inverters intended to
photovoltaic. Given that photovoltaic is a current source so it convert the DC voltage into AC or vice versa. In the case of
would be very dangerous to the load with unstable voltage. switching inverter topology switch to the manifold.
This condition can be called islanding condition. To be able Topology used in the system is Voltage Sourde Inverter
to detect the islanding condition parameters that can be used Control Current (VSICC) can generate a sinusoidal signal by
is a voltage parameter[1], When there is a failure to detect applying a positive / negative DC link[8].
the islanding then lead to poor quality of power on the
system, damaging equipment and hazardous for the workers C. Artificial Neural Network (ANN)
so that required quick process to detect islanding[2],
According to previous studies allowable time to decide Artificial Neural Network (ANN) is a network created by
photovoltaic when islanding condition is 100-2000 ms. To the human neural network. Human neural networks are
reach the time limit then used some method known as arranged on billions of billions of neurons that are connected
passive and active methods[3], In the active method, to other neurons. In the structure of neurons ANN is shown
disruption locally injected into the system for example, is by the number of nodes. Node function is to Figure out the
Active Frequency Drift[4] Sandia voltage and frequency pattern of a given input to output. Because of its nature as
shift of voltage variation methods and DG[5], While passive described previously, ANN referred to as adaptive systems.
methods to detect the islanding condition is read changes to ANN can be used to model complex relationships between
the system voltage amplitude[6], islanding detection must inputs and outputs to find data patterns[8], Here is the basic
distinguish events from other disorders, such as load form of a neuron can be seen in.
switching and short circuit. In this context, islanding
detection method is used Artificial Neural Network (ANN). Weights

Input Output
II. ISLANDING METHOD
Processing
A. Photovoltaic
unit
Photovotaicis a tool that works to convert solar energy
into electrical energy. photovotaic is a DC current source. To Figure 2 Basic forms of Neurons
convert the DC output, use inverter. The working principle of
solar cells is that when sunlight falls on the surface of the N-

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE

 input variable used both during lessons and in III. EXPERIMENTAL SETUP
recognizing an object. Weight is a load / weight are always Artificial Neural Network doing the learning on the
changing every time the given input as a learning process. system to get the weights and biases. Both grades that have
Processing Unit is the venue for the introduction of an object been obtained are used to perform calculations on a
by loading given. Output is output from the introduction of computing process that will be explained Figure 5:
an object.
Multiple input neuron is a condition of entry (input) is Start
more than one. Each of the existing input has a weight value
W shown in Figure 3. Identification and problem
Input Layer of Neuron
formulation

W1.1
P1 Modeling a grid-
P2 n a
ƒ
Photovoltaic system
P3
P4

PR b
W1.R Load combinations

a = ƒ (Wp+b)
PCC measurement
Figure 3 Multiple Input Neuron (Voltage)
On Figure 3 looks neurons have bias values and the
weights that when added together will form a network,
equation obtained from the multiple input neurons obtained Training ANN
as follows[8]:
, or (1)
no
(2)
(3) minimum MSE

D. Backpropagation
Backpropagation is a supervised learning algorithm and
is typically used by perceptron with many layers to change yes
the weights connected to existing neurons in the hidden
Weight and bias
layer. Output error backpropagation algorithm change to
change the value of the weight-weight in a backward Figure 5 Flow Chart ANN Modeling System islanding
direction. to get this error, advanced propagation stage
should be done first. On Figure 5 visible identification and problem
formulation process is part of the system design. One of the
Architecture backpropagation network as shown in problems posed when connecting photovoltaic and grid are
Figure 4 consists of three neurons in the input layer is X1, islanding condition. For that we need a method to be able to
X2 and X3. 2 hidden layers, namely Z1 and Z2 and one recognize the condition. The next process is to model grid
output layer unit is Y. activation function used in the hidden photovoltaic system in order to test the planned system. Load
layer to the output layer is the activation function logsig[8]. combinations used in planning. Measurements at the point of
point of comment coupling (PCC) will generate data that will
Input Layer
be used as input ANN.
Hidden Layer
X1 v11
v12
v13 Z1
W1
v21
Output Layer
x2 v22
v23
v31 Z2 Y
W2
v32
v33
x3
b12

b11
b2

b1
b2

Figure 4 Architecture Artificial Neural Network

 Point of Common Coupling
(PCC)
CB
L

Point of Common Coupling Voltage

(PCC) C Source
DC
CB Inverter
L

Liner
Voltage Current
DC Source C Load RLC Islanding Phase Control
Inverter detection Locked
Loop

Liner Islanding
Current detection AC
Control Phase Load RLC
Locked
Loop Load RLC

GRID

Figure 6 Grid-Photovoltaic System Modeling the Distributed

The training process is the process of recognition of data Readings were taken per cycle adapted to the training data
patterns that will result in the weights and biases. The that has been described Figure 5.
training process is done to get the value of Mean Square
Error (MSE) of the smallest. Having obtained the weights A. Load configuration
and biases with the smallest MSE, will be conducted in the load conditions will be simulated on the system when the
ANN computational process which will be explained on islanding condition, the output voltage signal generated at the
Figure 7: point of the PCC will be passed as input ANN. There are
Start
several conditions that are planned with cosphi value which
varies the conditions Powermatch, Overvoltage and
Undervoltage. Powermatch condition is the amount of power
generated by photovoltaic power needed is equal to the load.
PCC measurement In the overvoltage condition, when the condition of islanding
(Voltage) magnitude of the amplitude measured on PCC point will
increase. In undervoltage conditions, when the conditions
islanding magnitude of the amplitude measured at the point
Data per cycle
of the PCC will be reduced.
-1 = no
t = 1/50 islanding
B. Modeling-Grid Photovoltaic Systems
Modeling aims to simulate the conditions islanding on-
grid photovoltaic system are distributed. It is expected from
no the results of the simulation, the system can detect the
ANN Computing
islanding condition. Figure 6 Modeling shows photovoltaic
system of grid-distributed.

yes C. Inverter
2 = islanding inverter is a tool that can convert input DC voltage into
AC sinusoidal. To generate a sinusoidal ac output then the
process of switching on the inverter must be set. The
CB Trip conversion result is used as a supplier of the load connected
Figure 7 ANN computing system flowchart to the grid. Inverter topology used Voltage Source Inverter
(VSI) will be explained in Figure 8.
On Figure 7 looks to detect the system under normal
conditions or islanding computation performed on the input
signal continuously. readings that are applied are reading the
signals from the measurement results at the point PCC.
 (12)
S1 S2 Another parameter that is used for the calculation of the
+ value of the filter is the source of DC link voltage of 1000
- volts, 50Hz grid frequency, frequency 10.000. From these
DC
parameters obtained value of the LC filter on the system is L
44.19 14.48 mH and C μF. The following is Figure 11 LC
S1- S2- filter design the inverter circuit.
Figure 8 Topology Switching Inverter
The following is Figure 9 comparator Phase Locked DC PWM Load
Loop (PLL) and signal modulation of the inverter switching
process. Filter LC
S1 Figure 11 Single Phase LC filter on the output Inverter
+
_ S1-

S2 F. Active Frequency Drift Method

+ S2- islanding occurs when the main power plant of
_
photovoltaic system is impaired but continued to supply the
load. This condition must be detected aims to keep electrical
equipment from damage that can be caused. Use of Active
Frequency Drift (AFD) is one of the methods used to detect
Figure 9 Comparator PLL and signal modulation islanding. Selection of active methods performed to detect
the condition powermatch system. The following is Figure
Here is the calculation of modulation and triangular 12 wave scheme using AFD.
voltage as a reference for the process of switching the
inverter compiled the PLL output signal as seen in Figure 9
are as follows. Grid Voltage

Grid Current

(4)
0
Tz
(5)
T grid
(6)
(7)
(8) Time (s)
Figure 12 Using Wave scheme AFD
D. Phase Locked Loop
Phase Locked Loop (PLL) is a mechanism that is The working principle of active detection frequency drift
contained in the grid connected inverter to synchronize the is to provide disorder by injecting an active signal at the
phase of the phase measured at the PCC with phases at the output of the inverter. By keeping the THD of less than 5%.
inverter output. PLL also generates output signals with the Equation currents on active methods used are as follows.
same frequency as the input signal. The following is Figure (13)
10 basic diagram of the PLL.
Where Cf is the Chopping fraction, Tx is Tgrid frequency
Phase Loop
Vaoltage shift is half cycle of the voltage signal. The amount of
Control
Reference
Detector Filter
Oscillator Reference interference on the AFD method is determined by the value
Signal PLL of the frequency shift.

Figure 10 Basic Chart Phase Locked Loop (PLL) IV. SIMULATION RESULT
Discussion of results of simulations carried out using
E. LC Filter Design several conditions of load on the system and islanding
The use of filter aims to minimize the ripple that is a detection process using Artificial Neural Network (ANN).
result of the process of switching on the inverter. Calculation Some carried the load configuration intended that the
of the value filter is used based on the following equation. condition of the method used to detect the islanding
condition at the time of reading a voltage signal in Point
(9) commen Coupling (PCC).

(10)

(11)
A. Modeling-Grid Photovoltaic Systems
The following is Figure 13 design modeling system is
used to simulate the conditions in accordance islanding
discussed in the previous chapter.

Figure 13 Modeling-Grid Photovoltaic Systems

Can be seen on the entire circuit of the system.

Photovoltaic which is assumed to use a DC source that is B. Artificial Neural Network Method for Detection
converted into AC voltage using an inverter. To pull out a islanding
good voltage on the inverter used LC filter. The output of the Islanding detection process performed by ANN. ANN
inverter is connected to the grid and load. The process of reading output voltage of the PCC point. The readings done
merging the grid is done using a PLL to synchronize the every cycle continuously. When ANN ANN detects the
phase, amplitude and frequency of the voltage. The output of islanding conditions will instruct CB which connects
the inverter can produce power equal to 4400 VA. between the inverter and the load should be disconnected. It
aims to avoid damage to the equipment. Previous ANN will
learn the pattern of the normal signal and islanding by
training overall load condition data that is powermatch,
overvoltage and undervoltage with varying grades cosphi. So
expect any change in the load will still be detected using
ANN. The following is Figure 14 chart data used in the ANN
training.

Load

Powermatch Overvoltage Undervoltage

Condition

Resistive Capacitive Inductive Resistive Capacitive Inductive Resistive Capacitive Inductive

Load Character

Figure 14 Data Training On ANN

On seen the load is divided into three conditions, namely to 1. Thus, the total signal data voltage ANN is trained on the
powermatch, overvoltage and undervoltage. Each of these data 39. The data has been obtained, each one their
conditions is divided into three properties, namely load sampling data. For before the islanding condition and after
resistive, capacitive and inductive. Each of these properties is islanding, each condition is taken 10 cycles. So each
modeled load to 13 data at each condition. By changing the condition load on the input data obtained at 260 ANN cycle
value cosphi of the load. In cosphi system used is cosphi 0.65 data, a total of three data obtained for 780 data.
 Of the sampling results, conducted training using ANN.
The training randomly formed with the composition of 80%
for the training process, 10% and 10% for the test of
validation (including data trainning and data testing). ANN
used consisted of three layers consisting of one input layer,
one hidden layer and one output layer. To get the value of
Mean Squared Error (MSE) the minimum training will be
done several times by changing the number of neurons in the
hidden layer up to get his small MSE value to the fewest
number of neurons. Attached training and testing data in the
attachment sheet.

Figure 15 Target Suitability curve ANN

On Figure 15 looks curve fit between the target data and
training data, validation and testing. With an overall
correlation of 0.98.

Figure 16 1 Simulation Results and islanding detection ANN

From the simulation results in, the nature of which is
used is a resistive load with cosphi 1, the value of the resistor
that is used by 11 Ω. looks ANN continued to work until 0.2
seconds, 0.2 seconds before the normal system conditions. REFERENCES
After 0.2 seconds on the cycle to 8 ANN detect islanding [1] “Intelligent-Based Approach to Islanding Detection in Distributed
condition of the system. The simulations have been Generation - IEEE Explore Document.” [Online]. Available:
performed on all load conditions. ANN system capable of http://ieeexplore.ieee.org/abstract/document/4141116/. [Accessed:
detecting the islanding condition. ANN read signals for 6 07-May-2017].
[2] P. K. Dash, M. Padhee, and S. K. Barik, “Estimation of power quality
consecutive cycles detected islanding. ANN then will issue indices in distributed generation systems during power islanding
an order to trip the circuit breaker. This is done to avoid conditions,” Int. J. Electr. Power Energy Syst., vol. 36, no. 1, pp. 18–
ANN read other conditions of the system, or when there is 30, Mar. 2012.
example transient load changes. [3] IEEE Standard for Interconnecting Distributed Resources with
Electric Power Systems. The Institute of Electrical and Electronics
Engineers, Inc., 2003.
[4] J. Stevens and G. Ginn, “Development and testing of approach to
anti-islanding in utility-interconnected photovoltaic systems,” 2000.
V. CONCLUSSION [5] A. Y. Hatata, E.-H. Abd-Raboh, and B. E. Sedhom, “Proposed
Method Artificial Neural Network (ANN) can detect Sandia frequency shift for anti-islanding detection method based on
artificial immune system,” Alex. Eng. J., vol. 57, no. 1, pp. 235–245,
islanding condition system. ANN identify the pattern of the Mar. 2018.
input signal at the point PCC. The correlation of the results [6] F. Wang and Z. Mi, “Passive Islanding Detection Method for Grid
of the ANN training and target of 98%. Connected PV System,” 2009, pp. 409–412.
[7] teguh utomo, “Kajian kelayakan sistem photovoltaik sebgai
pembangkit daya listrik skala rumah tanggga (studi kasus di gedung
ACKNOWLEDGMENT VEDC malang),” EECCIS, pp. 13–14, 2009.
Authors expect criticism and comments from readers to [8] S. Kusumadewi, Membangun Jaringan Syaraf Tiruan Menggunakan
MATLAB & EXCEL LINK. Graha Ilmu, 2013.
improve quality of this paper. thank you [9] Y. Jung, J. Choi, B. Yu, J. So, G. Yu, and J. Choi, “A Novel Active
Frequency Drift Method of Islanding Prevention for the grid-
connected Photovoltaic Inverter,” 2005, pp. 1915–1921.