10 1109@it48810 2020 9070299

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

24th International Conference on Information Technology (IT)

Zabljak, 18 – 22 February 2020

MPPT controller based on the neural network


model of the photovoltaic panel
Maja Rolevski, and Žarko Zečević, Member, IEEE

Abstract—In this paper, we propose the MPPT algorithm great importance. The solar panel output power has a
that is based on the single-layer neural network model of the nonlinear characteristic, i.e. it nonlinearly depends on the
photovoltaic panel. It has been shown that the neural network voltage changes. The voltage that corresponds to the maximal
can be employed to accurately model the relationship between
power is called maximal power point (MPP). Since the MPP
the photovoltaic panel current, voltage, solar irradiance, and
temperature. Unlike the equivalent circuit model, the neural varies with temperature and irradiance changes, it is
network model can be used to calculate the gradient of the P-V necessary to implement algorithms that will track the MPP in
curve, thus enabling the design of the simple MPPT technique the real-time. In the literature, these algorithms are referred to
that relies on the steepest ascent method. Simulation results as MPP tracking (MPPT) algorithms.
show that the proposed algorithm exhibits a faster convergence A number of MPPT algorithms have been proposed in the
speed and a smaller steady-state error than the conventional
literature. These algorithms can be divided into two groups:
P&O algorithm.
conventional algorithms and soft computing approaches.
I. INTRODUCTION Conventional MPPT methods are based on the current and
voltage measurements at the output of the photovoltaic panel.
Renewable sources play a significant role in electricity Among them, the most popular algorithms are Perturb and
production nowadays. Among them, solar energy is Observe Algorithm (P&O) and Incremental Conductance
distinguished by its many advantages: reliable production of [12]. Recently, MPPT algorithms based on soft computing
electricity without pollution, simplicity of the solar system (SC) have received increased attention. SC-MPPT methods
and its easy installation, long-term exploitation, etc. [1]. Also, have the capability to track maximal power in sudden
the widespread use of solar systems has led to a decrease in changes of atmospheric parameters [13]. The disadvantages
price over time, as well as an improvement in their energy of these methods are that they are computationally intensive
efficiency [1]. and require additional sensors for implementation.
Various equivalent electrical circuits for solar cell In this paper, first, we describe a procedure of modeling
modeling have been proposed in the literature, among which the I-V characteristics of a photovoltaic cell using a neural
a single-diode model is mostly used. The reason for this is its network (NN model). It will be shown that the relationship
simplicity and satisfactory accuracy [2-4]. A single-diode between the cell current, voltage, temperature, and solar
solar cell model has five unknown parameters, which are irradiance, can be explicitly modeled by the one-hidden-layer
estimated according to the information provided by the neural network. Furthermore, we propose the MPPT
manufacturer, or by using experimentally obtained data [5]. algorithm, which is based on the NN model of a photovoltaic
However, the current-voltage equation of the solar cell is panel (NMPPT). The proposed algorithm utilizes the steepest
given in the implicit form and, therefore, an iterative ascent method to track the maximal power point. Simulation
Newton-Raphson method is used to solve it [6]. In [7], the results show that the NMPPT has a faster convergence speed
Lambert W function is proposed in order to obtain an explicit and a smaller steady-state error compared to the P&O.
dependence between the solar cell output current and voltage.
Nowadays, soft computing methods are also widely used for II. EQUIVALENT ELECTRICAL CIRCUIT OF A SOLAR CELL
solar cell modeling. In [8], the parameters of an equivalent
electrical circuit are determined by using artificial neural The main component of a solar panel is a solar cell. The
solar cell can be modeled with a single-diode circuit shown
networks, whereas in [9, 10] multilayer neural networks are
in Fig. 1 [2, 3]. The model consists of a current source Iph,
directly applied to model the I-V cell characteristics. A hybrid
diode D, a series resistance Rs and a parallel resistance Rp. In
approach to solar cell modeling, which combines the benefits
this case, the series resistance represents the resistance
of neural networks and fuzzy logic, is proposed in [11]. between the P and N layers of the solar cell, while the
In addition to determining an adequate solar cell model, parallel resistance describes the impurities and micro defects
research focused on maximizing the panel power output is of within the solar cell, which depend on its properties [1 - 3].
The solar cell output current, which is described by this
model, can be expressed by Kirchhoff's law as follows:
The authors are with Faculty of Electrical Engineering, University of
Montenegro, Dzordza Vasingotona bb, Podgorica (e-mails: I  I ph  I d  I p , (1)
maja.rolevski@gmail.com, zarkoz@ucg.ac.me).

978-1-7281-5136-6/20/$31.00 ©2020 IEEE


Iph I Rs

Id Rp Ip V

Fig. 1. Equivalent electrical circuit of a solar cell

where Iph is the photovoltaic current generated by the current Fig. 2. P-V curve for different values of the irradiance and temperature
source, Id is the current through the diode, and Ip is the
current through the resistance Rp. The current Id is defined as: There is often a situation where the resistance effect of Rp
can be neglected. In practice, it has high values and causes a
  q(V  IRs )   low current Ip. Assuming Rp = ∞, the equation (7) takes the
I d  I s  exp    1 , (2)
  nkTc   following form:
where Is is the saturation current of the diode [A], q is the   q(V  IRs )  
elementary charge [= 1.602 ∙ 10-19 C], V is the output I  I sc  I s exp    1 . (8)
  nkTc  
voltage of the cell [V], n is the diode ideality factor, k is the
Boltzmann constant [= 1.381 ∙ 10-23 J / K], and Tc is the Furthermore, it could be observed that the I-V
actual temperature of the solar cell [K] [1]. characteristic given by (8), is defined in implicit form. For
The Ip current is defined as: this reason, an iterative Newton-Raphson method is usually
V  IRs used to solve it.
Ip  . (3) Different definitions of the solar cell array, modules and
Rp
panels can be encountered in the literature. In this paper, it is
In particular, the current generated by the current source,
considered that multiple solar cells connected in series define
Iph, exhibits a linear dependence on the solar irradiance and
a module, as well as that serial or parallel connection of
ambient temperature, which is described by the following
multiple modules defines a panel. The equivalent resistance
relation:
of the solar panel depends on the number of series and
G parallel connected modules. Similarly to (7), the output
I ph   I ph, n  K i ΔT  , (4) current of the panel can be written as follows:
Gn
  N   N 
where Iph,n is the current generated under Standard Test   V  IRs  s    V  IRs  s 
Conditions (STC, defined at ambient temperature 25 °C and  
I  I sc N p  I s N p  exp   Np    1   Np  , (9)
irradiance 1000 W/ m2), Ki short circuit current coefficient, nV N  
  t s    N 
ΔT=Tc - Tc,n (Tc,n is the nominal temperature defined in  Rp  s 
    Np 
Kelvin), G is the solar irradiance on the panel surface, and Gn      
is the nominal irradiance [4]. where N represents the number of serially connected cells in
By using the equations (1), (2) and (3), the expression for one module, Vt=NkTc/q the thermal voltage of the diode, Ns
the output current of the cell is obtained: the number of serially connected modules, and Np the number
 of parallel connected modules [4].
 q(V  IRs )   V  IRs
I  I ph  I s exp    1  . (5) The P-V characteristic of photovoltaic module, for
  nkTc   Rp different values of the irradiance and temperature, is shown
in Fig. 2. It can be seen that maximal power point depends on
In order to simplify the equation (4), the following S and T. For this reason, the MPPT algorithms are employed
approximation is often used in practice: to track the MPP and keep the output power at constant
I ph  I sc , (6) value. One of the most commonly used MPPT methods in
practice is the Perturb and Observe algorithm (P&O) [12].
where Isc is the short circuit current of the cell, specified by This algorithm changes the output voltage towards its optimal
the cell manufacturer. Considering equation (6), equation (5) value in discrete steps. The direction of change (positive or
can now be expressed as: negative) is determined according to the change of the output
power (the voltage and current need to be measured). The
  q(V  IRs )   V  IRs disadvantages of the P&O algorithm are slow convergence
I  I sc  I s exp    1  . (7)
  nkTc   Rp and oscillations about the MPP in the steady-state [12, 14].
III. PROPOSED MPPT ALGORITHM in order to reduce the cost of the implementation, but that is
beyond the scope of this paper. It is worth noting that MPPT
A. Proposed neural network model
algorithm implemented in this way will have faster
In this paper, the Feedforward Neural Network (FFNN) is convergence than the P&O algorithm, in which the gradient
used to approximate the panel equation, given by (9). The is calculated heuristically, by observing the current and
solar irradiance G, temperature Tc and voltage V are used as voltage measurements.
inputs to the neural network, while the panel current I is set
to its output. The neural network model with one hidden IV. SIMULATION RESULTS
layer and three neurons is used to model the I-V curve (Fig.
3). The neural network should be trained for a wide range of The characteristics of the solar cell that we used in
parameters, that is, for those values of (G, Tc) that can occur simulations are given in Table I. The photovoltaic panel
in practical scenarios. More details on modeling the I-V consists of a total of 90 solar cells (5 serially connected
curve with a neural network can be found in [15]. modules, consisting of 18 serially connected cells). The
neural network is trained for 50 random values of G and Tc
form the range G ϵ {50, 1000} and Tc ϵ {0, 75}. In both the
T P&O and NMPPT algorithm we used the fixed step size: 0.1
and 0.5, respectively.

G I TABLE I
SOLAR CELL CHARACTERISTICS

Short-circuit current, Isc [A] 7.34


Open-circuit voltage, Voc [V] 0.6
V
Irradiance used for measurements,
Input Hidden Output 1000
G0 [W/m2]
layer layer layer
Quality factor, n 1.5
Series resistance, Rs [Ω] 0.001
Fig. 3. The structure of proposed neural network model

B. NMPPT algorithm
In general, the photovoltaic output power is defined as:
P  VI . (10)
The aim of the MPPT algorithm is to find the point (Vmp,
Imp) which corresponds to maximal output power Pmp:
Pmp  P (Vmp , I mp ) P (V , I ) , V  [0, Vmp ] & I  [0, I sc ]. (11)

Observe from Fig. 2 that the P-V curve is concave. Thus,


the ascend gradient technique can be applied to find the (a)
maximal power point:
P 
Vn 1  Vn   Vn  μ Vn  I n 
Vn Vn
(12)
 I 
 Vn  μ  I n  Vn  n 
 Vn 

It is worth considering that the photovoltaic equation (9) is


given in implicit form and, therefore, it could not be used to
calculate (12). On the other side, the NN model of the (b)
photovoltaic proved us with direct relationship between the Fig. 4. (a) P-V curve for changing cell temperature, (b) P-V curve for
changing solar irradiance
current and voltage. The derivative of In with respect to Vn
can be approximated as follows: A. NN model PV of the photovoltaic panel
I n I n (Vn  V , S , T )  I n (Vn , S , T ) In this example, we examine the capability of the neural
 . (13) 1-3-1 NN to model a considered photovoltaic panel. In the
Vn V first scenario, the solar irradiance has a constant value of
G=800W/m2, while the cell temperature has the following
Note that this algorithm requires the knowledge of S, T, V
values Tc ϵ [10˚C, 50˚C, 60˚C, 100˚C]. In the second
and I for implementation. For that reason, we will consider
scenario, the cell temperature has a constant value Tc=25˚C,
all necessary measurements known in simulations.
while the solar irradiance has the following values G ϵ
Moreover, it is possible to design the estimator of irradiance
[200W/m2, 600W/m2, 1000W/m2, 1400W/m2]. Fig. 4 shows this is that the NMPPT uses the true gradient to search for
P-V characteristics generated by Eq. (9) (solid lines) and by maximal power point.
NN model (dot lines). It can be observed that, for the data Fig. 6. shows simulation results for the time-varying
values for which the network is trained, the characteristics of irradiance. Similar conclusions as in the first example can be
the network model coincide very well with those of the drawn here.
simulated model. However, when the cell temperature is V. CONCLUSION
100˚C, as well as when the irradiance is 1400W/m2, there is
a slight deviation from the actual values, as expected. In this paper, we propose the simple MPPT algorithm that
is based on steepest ascend method. The neural network
B. NMPPT controller photovoltaic model is used to calculate the gradient of the P-
In this example, we made comparison between the P&O V curve with respect to the voltage. Simulation results show
and the NMPPT algorithm. Since the NMPPT algorithm that proposed method exhibits the faster convergence and
requires knowledge of irradiance and temperature, it has smaller steady-state error than conventional P&O algorithm.
been assumed that these values are measured or estimated In the proposed method, it is assumed that the irradiance
without errors. Fig. 5 shows the output power of the panel and temperature are measured without errors. The future
for the irradiance G=1000W/m2 and the temperature of the work will be focused on the design of the irradiance
panel T=25°C. estimator, as well as on the design of the hybrid algorithm
that will be immune to the irradiance estimation errors. Also,
the proposed method will be verified by using the real
photovoltaic measurements.

REFERENCES
[1] G. M. Masters, “Renewable and efficient electric power systems”,
Wiley-IEEE Press, Second edition, Jun 2013.
[2] M. F. Nayan, S. M. S. Ullah, "Modelling of solar cell characteristics
considering the effect of electrical and environmental parameters”,
Third International Conference on Green Energy and Technology
(ICGET), IEEE, Dhaka, Bangladesh, November 2015, pp. 1-6.
(a) [3] W. D. Soto, S. A. Klein, "Improvement and validation of a model for
photovoltaic array performance", Solar Energy, Vol. 80, Issue 1,
January 2006, pp. 78-88.
[4] M. G. Villalva, J. R. Gazoli, E. R. Filho, "Modeling and circuit-based
simulation of photovoltaic arrays," 2009 Brazilian Power Electronics
Conference, Bonito-Mato Grosso do Sul, 2009, pp. 1244-1254.
[5] E. Moshksar and T. Ghanbari, "Adaptive Estimation Approach for
Parameter Identification of Photovoltaic Modules," in IEEE Journal of
Photovoltaics, March 2017, vol. 7, no. 2, pp. 614-623.
[6] V. Quaschning, R. Hanitsch, “Numerical simulation of current-voltage
characteristics of photovoltaic systems with shaded solar cells”, Solar
Energy, Vol. 56, Issue 6, Jun 1996, pp. 513-520
[7] A. Jain, A. Kapoor, “Exact analytical solutions of the parameters of
real solar cells using Lambert W-function“. Solar Energy Mater. Solar
(b) Cells, 2004, vo. 81, pp. 269–277.
Fig. 5. Comparison between the P&O and NMPPT algorithm, a) transient, [8] Karatepe, E., Boztepe, M., Çolak, M.: "Neural network based solar
b) steady-state cell model", Energy Convers. Manage., 2006, vo 47, pp. 1159–1178.
[9] M. Karamirad, M. Omid, R. Alimardani, H. Mousazadeh, S. N.
Heidari, “ANN based simulation and experimental verification of
analytical four- and five-parameters models of PV modules”,
Simulation Modelling Practice and Theory, Vol. 34, 2013, pp. 86-98.
[10] M. Hadjab, S. Berrah, H. Abid, "Neural network for modeling solar
panel“, International Journal of Energy, Vol. 6, Issue 1,February 2012.
[11] A. Chikh, A. Chandra, “Adaptive neuro-fuzzy based solar cell model”,
IET Renewable Power Generation, Vol. 8, Issue 6, August 2014, pp.
679-686.
[12] Dezso Sera, Laszlo Mathe, Tamas Kerekes, Sergiu Viorel Spataru,
Remus Teodorescu, On the Perturb-and-Observe and Incremental
Conductance MPPT Methods for PV Systems, IEEE Journal of
Photovoltaics, Volume 3 (Issue 3), Pages 1070–1078, July 2013.
Fig. 6. Comparison between the P&O and NMPPT algorithm in the time- [13] Dileep G., S.N.Singh, Application of soft computing techniques for
varying irradiance scenario maximum power point tracking of SPV system. Solar Energy, Volume
141, Pages 182–202, 1 January 2017
Observe from Fig. 5(a) that the NMPPT needs 9 iterations [14] A. Kapić, Ž. Zečević and B. Krstajić, "An efficient MPPT algorithm
for PV modules under partial shading and sudden change in
to reach the steady-state, whereas the P&O algorithm irradiance," 2018 23rd International Scientific-Professional
requires 57 iterations. Beside the faster convergence, the Conference on Information Technology (IT), Zabljak, 2018, pp. 1-4.
NMPPT exhibits a smaller steady-state error, without [15] M. Rolevski, Ž. Zečević, "Modeling of photovoltaic panel using
oscillations, which can be seen from Fig. 5(b). The reason for neural networks", 6th CG KO CIGRE Session, Montenegro, 2019.

You might also like