Вы находитесь на странице: 1из 29

Low Complexity Neural Network for Maximum

Power Point Tracking of Photovoltaic System


in Rapidly Changing Weather Conditions

Outline
Introduction
Proposed PV system
PV array modeling & characteristics
Proposed techniques
Computationally efficient FLANN (CEFLANN)
Trigonometric FLANN (TFLANN)
Result analysis
Conclusion and future work
References

Introduction
PV systems are most preferred because:

Provides clean green energy


Less operating & maintenance cost
Smart distributed generation

MPP

MPP

Operate at MPP to maximize the efficiency of system.

MPPT
A technique used to get the maximum possible power from solar

panels.
Maximum power point tracker (MPPT) tracks the MPP &

connected between the PV array and boost converter.

Proposed PV system
Solar Irradiance
Temperature

PV array

Boost
converter

VPV

Duty
cycle

MPPT
controller
(NN)

Vmpp
or Vref

Control
unit

Fig.1 Block diagram of PV system


MPPT controller generates the reference voltage with input as irradiance and
temperature.
Control unit to generate duty cycle for the boost converter.

PV array modeling & characteristics


Basic building block of PV arrays is solar cell.
A P-N junction that converts light energy into electricity directly.
RS IPV

P
ID
I ph

RJ

Ish
Rsh

+
VPV

Fig. 2. Single diode model to model single solar cell

Contd.
Applying Kirchhoffs law to the node P we get

I pv = I ph - I sh - I D
The mathematical model of PV array is represented by the
equation:

q V pv I pv Rs

I pv N p I ph N p I rs exp

Ns

kTA
N s N p V pv I pv RS


Rsh

(1)
(2)

Contd.

Table.1 Key specifications of ELDORA-40 module (1kW/m2, 250C)


Parameters

Values

Maximum Power (Pmpp)

40W

Voltage at maximum power (Vmpp)

17.4V

Current at maximum power (Impp)

2.3A

Short circuit current (Isc)

2.3A

Open-circuit voltage (Voc)

21.9V

Temperature coefficient of short

0.0017

circuit (ki)
Cell reverse saturation current (Irr)

1.2*10-7A

Number of series cells (Ns)

36

Contd.
Ip v (A m p )

I-V characteristics of PV array


1000 W/m2

100

MPP

500 W/m2

MPP

50
0
0

MPP

300 W/m2

50

100

P p v (W a tt)

Vpv (Volt)
(a
)
P-V characteristics
of PV array

10000

MPP

1000 W/m2
MPP

500 W/m2

5000

MPP

0
0

300 W/m2

50

100

Vpv (Volt)

(b)

Fig.3. (a) I-V characteristics and (b) P-V characteristics of PV array

Proposed Techniques
Functional Link Artificial Neural Network (FLANN) is used

to track MPP.
Because it is
Very efficient
Nature to find the non-linearity
Has two basic components:
Functional expansion component or block (FEB)
Learning component.
FEB expands the input space to higher dimensional space.
FEB produces nonlinearity through non-linear functions .

Contd.

Fig.4. the structure of neural network for MPPT

Two low complexity FLANNs are discussed here.


Computational efficient FLANN (CEFLANN)
Trigonometric FLANN (TFLANN)

Computationally Efficient FLANN (CEFLANN)


Single layer network
All the inputs of the input pattern pass through FEB to produce

the expanded input pattern


The structure of CEFLANN is shown in Fig.5.

Fig.5. Architecture of Computationally Efficient FLANN

Contd.
The output of FEB is described by the

equation:
s

FEBi tanh(ai 0

i p, j n

i 1, j 1

Ai j X j )

Here i=1, 2=P(P=2) and j=1, 2n(n=2)


Where, p = number of output of FEB,
n = number of input to FEB for one output,
a = input bias weight matrix,
s = current number of sample,
S = total number of samples or patterns
And A = input weight matrix.

(3)

Contd.
After the expansion the output is calculated by
M

y {W s ( k ) X s (k )}
s

(4)

k 1

Where,

X s [ X s ( k ), ........., X s ( M )]T ,

W [W ( k ), ........., W ( M )]
s

(5)

X s=input matrix for sth sample,


W s =output weight matrix for sth sample,
s
y =predicted output for sth sample, M = total number of input to
the summation block after expansion.

Trigonometric FLANN (TFLANN)


Single layer neural network
Trigonometric functions are used in the FEB
Each xi in input pattern is expanded using trigonometric functions

with order p as :
{ sin( *xi), cos( *xi), sin (2 *xi), cos(2 *xi),sin(p *xi),
cos(p *xi)}.

Contd.
The structure of the TFLANN is shown in Fig.6.

Fig.6. Architecture of Trigonometric FLANN

Contd.
The trigonometric expansion after the FEB is given below:

X 1, X irradiance, X X
s
1

s
2

s
4

s 2
2

X sin( pi X ), X cos( pi X ),
s
5

s
2

s
6

X 3s temperature, X 7s X

s
2

s 2
3

(6)

X sin( pi X ), X cos( pi X )
s
8

s
3

s
9

s
3

the output is calculated by:


M

y {W s (k ) X s ( k )}
s

k=1, 2.M(M=9)

(7)

k 1

Where,

X s [ X s (k ), ........., X s ( M )]T ,

W [W (k ), ........., W ( M )]
s

(8)

Adaptation Of Weights And Performance


Evaluation
The error for each sample is expressed as:

edy
The cost function is :

1 2
E e (k )
2

(9)
(10)

The cost function minimized by gradient descent algorithm by

training the weights.


The learning rate () is taken in the range of (0.1-0.2).

Contd.
Differential Error: e d y
Root Mean Square Error (RMSE)=

(11)

( e ) / S
2

(12)

Mean Absolute Percentage Error (MAPE)

{ (| e | /d) }/ S 100

(13)

RESULT
RESULT ANALYSIS
ANALYSIS

DATASET GENERATION
Generated in the MATLAB/SCRIPT environment
The reference voltage (Vmpp) for randomly taken irradiance and

temperature for the discussed PV system


Irradiance is varied in the range of (100-1000) W/m 2 in a step of 25 W/m2.
The temperature is varied in the range of (25-75) 0C in a step of 20C.
80% data for training, next 10 % data for testing and last 10% is for

validation from the 900 samples

Contd.
Case: 1 (CEFLANN)

magnitude(p.u)

target & predicted of testing

target
predicted

0.5
0
720

740

760

780

No. of samples

Fig.7. Target and Predicted using CEFLANN

800

Contd.
Case: 2 (TFLANN)

magnitude (p.u)

target & predicted of testing

1
0.5
0
720

target
predicted

740

760

780

No. of samples

800

Fig.8. Target and Predicted using TFLANN

Contd.
Case: 3 Comparison of different errors between CEFLANN and
TRFLANN

magnitude(p.u)

error during testing

0.1
0.05
0
-0.05
0

CEFLANN
TFLANN

20

40

60

No. of samples

80

Fig.9. error comparison between CEFLANN and TFLANN

Contd.

magnitude (p.u)

Case: 3 Comparison of different errors between CEFLANN and


TRFLANN

x 10

-3

RMSE During testing


CEFLANN
TFLANN

4
2
0
0

20

40

60

No. of samples

80

Fig.10. RMSE comparison between CEFLANN and TFLANN

Contd.

magnitude (p.u)

Case: 3 Comparison of different errors between CEFLANN and


TRFLANN

0.2

MAPE During testing


CEFLANN
TFLANN

0.1
0
0

20

40

60

No. of sample

80

Fig.11. MAPE comparison between CEFLANN and TFLANN

Contd.

COMPARISON BETWEEN CEFLANN AND TFLANN


DURING THE
TRAINING PROCESS
Type of
network

Network
structure

CEFLAN
2-6-1
N
TRFLAN 2-9-1
N

No. of
iteration
in
training

Executio
n time
(ms)

720

30.4

720

27.2

error
more
less

No. of
Weights
to be
updated
12
9

CONCLUSION AND FUTURE WORK


Compared to CEFLANN, TFLANN is:
Very good technique to predict the output.
More efficient
Less erroneous
Less computational complexity.
More accurate
The future works are:
I) Implementation of control unit to find the duty cycle for the
boost converter
II) Hardware implementation of this algorithm
III) MPPT controller for partially shaded condition using
intelligence techniques.

REFERENCES
S. Premrudeepreechacharn, and N. Patanapirom. "Solar-array modelling

and maximum power point tracking using neural networks.", Power Tech
Conference Proceedings, 2003 IEEE Bologna. Vol. 2. IEEE, 2003.
Dezso Sera, Tamas Kerekes, Remus Teodorescu and Frede Blaabjerg.
"Improved MPPT algorithms for rapidly changing environmental
conditions." Power Electronics and Motion Control Conference, 2006.
EPE-PEMC 2006. 12th International, pp. 1614-1619. IEEE, 2006.
Fangrui Liu, Yong Kang, Yu Zhang and Shanxu Duan ,"Comparison of
P&O and hill climbing MPPT methods for grid-connected PV converter.",
Industrial Electronics and Applications, 2008. ICIEA 2008. 3rd IEEE
Conference on. IEEE, 2008.
Zhou Xuesong,Song Daichun,Ma Youjie, Cheng Deshu. "The simulation
and design for MPPT of PV system Based on Incremental Conductance
Method." information engineering (ICIE), WASE international conference
on 2010. Vol. 2, IEEE, 2010.

Trishan Esram and Patrick L. Chapman, "Comparison of photovoltaic array

maximum power point tracking techniques." IEEE transactions on energy


conversion ec 22, no. 2 (2007): 439.
Mummadi Veerachary, Tomonobu Senjyu and Katsumi Uezato, "Neuralnetwork-based maximum-power-point tracking of coupled-inductor
interleaved-boost-converter-supplied PV system using fuzzy controller."
Industrial Electronics, IEEE Transactions on 50, no. 4 (2003): 749-758.
Tsai-Fu Wu, Chien-Hsuan Chang, and Yu-Kai Chen. "A fuzzy-logiccontrolled single-stage converter for PV-powered lighting system
applications." Industrial Electronics, IEEE Transactions on 47, no. 2 (2000):
287-296.
Huan-Liang Tsai, Ci-Siang Tu, and Yi-Jie Su. "Development of generalized
photovoltaic model using MATLAB/SIMULINK." In Proceedings of the
world congress on Engineering and computer science, vol. 2008, pp. 1-6.
2008.
K.H. Hussein, I. Muta, T. Hoshino and M. Osakada,"Maximum photovoltaic
power tracking: an algorithm for rapidly changing atmospheric

THANK YOU

Вам также может понравиться