Particle Swarm Optimization based Adaptive
Neuro-Fuzzy Inference System for MPPT Control
of a Three-Phase Grid-Connected Photovoltaic
System
Jeffrey Andrew-Cotter1, M. Nasir Uddin1, and Ifte Khairul Amin1
1
Electrical Engineering, Lakehead University, Thunder Bay, Canada. E-mail: jsandrew@lakeheadu.ca
Abstract— This paper presents the analysis and operation of a
grid connected photovoltaic (PV) energy conversion system with
an Adaptive Neuro-Fuzzy Inference System (ANFIS) based
maximum power point tracking (MPPT) algorithm. Particle
swarm optimization is used to train the membership functions
while the least squares algorithm is used to update the
consequent parameters of the ANFIS with changing operating
condition of PV solar system. The MPPT algorithm maximizes
conversion efficiency by adjusting the duty cycle of the buck
boost converter to change the output voltage of the solar panel
and hence achieving the maximum panel output power for a
given set of environmental conditions. The ANFIS is trained by
using a hybrid algorithm implementing least squares estimator
and particle swarm optimization with data obtained by operating
the system using the Perturb and Observe (P&O) MPPT
algorithm. The performance of the proposed ANFIS based
MPPT algorithm is validated in simulation using
MATLAB/Simulink at different operating conditions. It is
proven that the designed ANFIS based MPPT scheme achieves a
very fast response with little oscillations while transferring
maximum power from solar panel to the grid line as compared to
the conventional P&O based MPPT scheme.
Keywords— particle swarm optimization (PSO), photovoltaic
(PV) system, maximum power point tracking, adaptive neuro-fuzzy
inference system, buck boost converter, inverter, energy conversion
I. INTRODUCTION
There is a significant demand for renewable energy systems
to ensure sustainable and environmentally friendly living.
According to Environment and Climate Change Canada
(ECCC), Canada has committed, along with other leading
countries in GHG emissions, to maintain the total global
temperature increase below 2°C. Solar photovoltaic (PV)
energy systems are of particular interest due to the availability
and portability of such systems. According to Natural
Resources Canada (NRC), Canada’s reliance on solar energy as
a renewable energy source is rapidly growing. NRC claims that
Canada’s total quantity of installed solar energy reached 1834
MW (megawatts) in 2014.
978-1-5386-9350-6/19/$31.00 ©2019 IEEE
To generate energy, solar PV modules rely on weather
conditions such as sun irradiance, ambient temperature,
shading etc. The PV curves for a solar panel show that there is
one optimal point on the curve corresponding to the maximum
power under no shading condition. Since the MPP changes
with a change in weather conditions, an algorithm/controller is
required to track this maximum power point and force the solar
panel to operate at that point at all times. There are various
types of MPPT methods [7]. The classical methods are Perturb
and Observe and Incremental Conductance. Since the PV
characteristics are highly nonlinear, these two methods are not
stable when the MPP is reached and large oscillations occur
and hence a decrease in power results.
Artificial intelligence (AI) based methods are becoming
increasingly popular due to the flexibility and more efficient
operation. These methods can be classified into three
categories: artificial neural networks (ANN), fuzzy logic (FL)
and ANFIS. ANFIS is a combination of both FL and ANN and
as such, combines benefitting features from both. ANFIS is a
fuzzy logic system that uses neural network training methods
to tune various linear and nonlinear parameters in the fuzzy
system which removes the need for previous system
knowledge that fuzzy logic inherently requires [1-2]. For the
proposed system, hybrid learning algorithm [1] is used which
implements the recursive least squares estimator (RLSE) for
linear consequent parameter training and particle swarm
optimization (PSO) for nonlinear membership function
parameter training. This proposed learning algorithm is a
derivative-free algorithm so any discontinuities in the system
do not affect the training. The algorithm is also able to search
for the global optimum and hence achieves the best possible
performance for the system.
Previous work on ANFIS based MPPT has been done for
DC-DC converters [1, 3, 6, 8, 9]. The majority of these papers
have focused on the basic hybrid learning algorithm
introduced by Jang [1] and not any advanced learning
algorithms. Further, these previous works have stopped at the
DC link and have connected a resistive load directly to the
output of the converter. The system proposed in this paper is a
2089
complete system with a three phase inverter to obtain AC
signals for grid connection with an advanced MPPT
controller. As a result, the proposed system is more practical
than the aforementioned works.
II. MAXIMUM POWER POINT TRACKING ALGORITHM
In order to transfer maximum power it is hard to match the
load impedance to the impedance of the PV panel with
changing environmental conditions. Therefore, a buck boost
converter with a MPPT algorithm is used to match the
impedance of the load to the solar panel. Changing the duty
cycle of the converter changes the impedance apparent to the
panel. Thus, the buck boost converter can be used as an
impedance matching circuit to guarantee maximum power
transfer from the solar panel to the load by operating the solar
panel at its MPP at all times regardless of weather conditions.
The irradiance significantly influences the PV output
current/short circuit current and hence the maximum power
point of the PV panel. Typical P-V and I-V curves of solar
panel are shown in Fig. 1 for various irradiance conditions.
The temperature of the panel influences the PV output
voltage/open circuit voltage and does not have a significant
effect on the MPP as compared to changing irradiance
conditions. Since the temperature of the panel does not
significantly affect the power output of the panel, a constant
temperature of 25 °C is considered while the irradiance is
considered to be in the range of 300-1000 W/m2.
Fig. 1. PV curves at various irradiances.
2090
A. ANFIS
The proposed MPPT algorithm implements a first order
Takagi-Sugeno-Kang (TSK) system as the 5 layer ANFIS [1].
The inputs are chosen to be PV voltage and PV current and the
output is chosen to be the duty cycle for the converter. Two
Gaussian membership functions per input are used to create
four rules. The rules for the ANFIS are:
where,
Ai (i=1, 2, 3, 4) are the input membership functions
containing nonlinear parameters;
(pi, qi, ri) (i=1, 2, 3, 4) are the linear parameters for the
consequents;
and are the voltage and current of the photovoltaic
cell, respectively.
Fig. 2. Structure of the proposed five layer ANFIS.
Instead of the general fuzzy system representation, the
fuzzy system can be transformed into a five layer network
representation as shown in Fig. 2. For this representation,
there are two inputs with two membership functions per input
and one output. There are also four rules used in this ANFIS.
Here, the output of layer 1 is the membership function’s value
at the specified input (fuzzification). The output of layer 2 is
the firing strengths corresponding to each rule. The output of
layer 3 is the normalized firing strengths. The output of layer 4
is the final output of each rule. The output of layer 5 is the
defuzzified overall system output.
To design a fuzzy logic system, no mathematical model of
the system is required. This feature is beneficial because,
many physical systems are complex and difficult to model.
Further, fuzzy logic is superior at handling system
nonlinearities. If the fuzzy logic system is used as a controller,
it is more advantageous than the classical linear controllers
such as the proportional-integral-derivative (PID) controller
Since the PID controller cannot effectively control highly
nonlinear systems since a linear approximation around the
operating point is usually performed to design a PID controller
for a nonlinear system. Also, classical controllers have the
downfall of being designed to control a system at a specific
operating point. If the operating point is changes, the
controller needs to be re-tuned or a different PID controller is
required. For a fuzzy logic controller, there is a range of
operating points defined by the universe of discourse for a
specific application. Therefore, if the operating point changes,
the fuzzy logic controller will still be effective at controlling
the system under the condition that the new operating point is
within the designed universe of discourse. Since an ANFIS is
a fuzzy logic system, ANFIS benefits from the above
mentioned features of fuzzy logic.
based on their previous position as well as better positions
found by other members of the flock [5]. As a result, the
swarm (or flock) moves towards the best solution for a given
problem. Therefore, by using PSO, the global minimum of
the objective function can easily be found depending on the
initial positions of the members in the population and
improves the tracking of the maximum power point as
compared to conventional MPPT methods. An explanation
and analysis of PSO can be found in [4]. The flow chart for
the proposed PSO method is shown in Fig. 3.

B. ANFIS TRAINING USING PSO-RLSE

Gaussian membership functions have two parameters that
need to be chosen: center and standard deviation. ANFIS uses
neural network training algorithms to tune all parameters of
the system. A hybrid training algorithm is used in this work in
which RLSE is used to train the linear consequent parameters
and PSO is used to train the nonlinear membership function
parameters. The recursive least squares estimator is defined
by the equations below:

where,
Өk+1 contains an estimate of the optimal linear parameter
values;
ak+1 is a vector containing the coefficients of the linear
parameters to be optimized;
yk+1 is the desired system output;
Po=αI, where α is a large number;
Өo is usually initialized to a vector of zeros;
k=1, 2, …, N;
N is the number of training data pairs.

Traditionally, the hybrid algorithm uses LSE and gradient

descent (GD) for training. GD uses the derivative of the
objective function to search for the minimum of the objective
function but has a high tendency to get caught in a local
minimum and never find the global optimum [1]. As such, an Fig. 3.Flow nchart of the PSO.
algorithm that does not use the derivative of the objective
function is of particular interest. The proposed ANFIS uses The update of particle velocity and position is achieved by
PSO for nonlinear parameter training due to its inherent ability the following equations:
to search for the global optimum and not get caught in local
optima. By choosing the personal and global learning rates
correctly, the PSO algorithm is stable and will guarantee
convergence. Particle swarm optimization mimics a flock of
birds travelling in which each individual in the flock moves

2091
where,
v is the velocity of each particle;
w is an inertia weight that reduces the effect of the
previous velocity as the search progresses;
c1 and c2 are the local and global learning rates
respectively. Note c1+c2≥4 for stability of the search;
rand(0,1) is a random number between 0 and 1;
gbest is the best position among all members (globally best
position);
pbest is the best position for a given particle (personal best
position);
x is the current position of the particle.

III. RESULTS AND DISCUSSION

A MATLAB/Simulink model is designed to test the

proposed ANFIS based MPPT control of the PV solar energy
conversion system as shown in Fig. 4. The model consists of a
PV array connected to a buck boost converter. The output of
the buck boost converter is connected to a three phase inverter
through a DC link. The output of the inverter is connected to
Fig. 5. Training Error.
the grid and to a three phase resistive load. Control of the
inverter is done in the d-q reference frame with tuned
proportional integral (PI) controllers controlling the d and q
voltages of the three phase output. PWM logic switching
signals are generated through sinusoidal pulse width
modulation (SPWM) and applied to the inverter. The ANFIS
calculates a duty cycle that is applied to the switch of the
converter in order to operate the solar panel at its maximum
power point. This operation allows maximum possible power
to be transferred to the DC link at all times. For training, the
root mean square error tolerance is set at 20% and the ANFIS
is trained for 85 epochs. Training was completed for more
than 85 epochs but no change in the error function was
observed. Since the error converges in 85 epochs, no more
epochs are required for training purposes. Fig. 5 shows a plot
Fig. 6. Input Membership Functions.
of the training error versus epoch number. Although the
training error is around 0.202, the output of the ANFIS Fig. 7. Solar panel output power for the proposed ANFIS based MPPT
matches closely to the training data. scheme.
The resulting membership functions for PV voltage and
PV current are shown in Fig. 6. Figure 7 shows the output
power of the solar panel for the proposed system and Fig. 8
shows the PV output power for the system operated with
conventional P&O based MPPT. Table-I compares the actual
power output of the panel and theoretical power output
(obtained from the PV curves) at various irradiances for the
proposed system. It is shown that the ANFIS is very effective

2092
 Fig. 4. PV solar energy conversion system block diagram. Fig. 9. Three phase output voltage at inverter end.

at forcing the solar panel to operate at its MPP with a fast

response and almost no oscillations at the MPP regardless of
irradiance conditions.
As shown in Fig. 7 and Fig. 8, the proposed ANFIS based
MPPT algorithm outperforms conventional P&O based MPPT
by achieving a faster response and fewer oscillations at each
MPP. These results validate the proposed algorithm and show
that the proposed algorithm has less power loss due to
oscillations at the MPP. Hence, the proposed system is able to
transfer more power to the DC link resulting in a more
efficient system as compared to the system using conventional
P&O MPPT. Fig. 10. Inverter output current
Figures 9 and 10 show the three phase voltage and current,
respectively, at the output of the inverter for the proposed Fig. 11. Buck-boost converter efficiency.
system. It is shown that both the current and voltage are
almost perfectly sinusoidal with abc sequence and a frequency TABLE-I: COMPARASION OF PV PANEL OUTPUT POWER FOR THE PROPOSED
MPPT SCHEME
of 60Hz. The total harmonic distortion for the both the voltage
and current is found less than 1%. Similar results for the THD Irradiance Theoretical Actual Power
are obtained for ‘b’ and ‘c’ phases. Thus, effective grid (W/m2) Power (W) (W)
connection is possible. 300 275 275

450 414 414

500 460 460

700 642 641.5

850 776 775

1000 908 906

The efficiency of the converter for varying irradiances

Fig. 8. Solar panel output power for the conventional PO based MPPT
scheme.
(and hence varying duty cycles) with is found in Simulink.
The buck boost converter is simulated with the proposed
ANFIS-based MPPT controller and connected to a resistive
load. Note that the inverter is not connected for this efficiency
simulation as only the efficiency of the buck boost converter is
considered. A plot of the efficiency under varying irradiance
condition is shown in Fig. 11. Note that the spikes in
efficiency every 0.05 seconds are due to the transition when
irradiance levels change. The converter efficiency is found to
be 0.97 or 97% for all irradiance levels

2093
 IV. CONCLUSION

An ANFIS based maximum power point tracking control

of PV solar system has been presented in this paper. Particle
swarm optimization is used to train the membership functions
and the least square algorithm is used to update the consequent
parameters of the ANFIS to cope with uncertain and variable
weather conditions of PV system. Thus, the proposed ANFIS
based MPPT scheme ensures maximum power transfer to the
DC link at all operating conditions. A PI controlled three
phase inverter is implemented to connect the PV system to the
grid. It is observed that the ANFIS controller is able to
effectively extract maximum power from the panel under
varying irradiance conditions while achieving a very fast
stable response. The performance of the proposed ANFIS
based MPPT scheme is also compared with the conventional
P&O based MPPT scheme. It is found that the proposed
system achieves a faster response with significantly less
oscillations as compared to the conventional MPPT scheme.
As a result, the proposed system is able to transfer more power
from the panel to the DC link and hence, improving the
overall efficiency of the conversion system as compared to
conventional P&O based system.

REFERENCES
[1] J.-S. R. Jang, C.-T. Sun, and E. Mizutani, Neuro-fuzzy and soft
computing: a computational approach to learning and machine
intelligence. Upper Saddle River, NJ: Prentice Hall, 1997.
[2] F. Karray and C. deSilver, Soft computing and intelligent systems
design: theory, tools, and applications. Pearson Publishing Inc., 2004.
[3] F. Belhachat, C. Larbes, “Global maximum power point tracking based
on ANFIS approach for PV array configurations under partial shading
conditions”, Renewable and Sustainable Energy Reviews, vol. 77, pp.
875-889, 2017.
[4] V.G. Gudise, G.K. Venayagamoorthy, “Comparison of particle swarm
optimization and backpropagation as training algorithms for neural
networks”, IEEE Swarm Intelligence Symposium, April 26, pp. 110-
117, 2003.
[5] V.Seydi Ghomsheh, M. Aliyari Shoorehdeli , M. Teshnehlab, “Training
ANFIS structure with modified PSO algorithm”, Mediterranean
Conference on Control and Automation, July 27-29, 2007.
[6] K. Ishaque, Z. Salam, M.Amjad, S. Mekhilef, “An improved particle
swarm optimization (PSO)–based MPPT for PV with reduced steady-
state oscillation”, IEEE Trans. on Power Electronics, vol. 27, no. 8, pp.
3627-3638, August 2012.
[7] T. Esram, P. L. Chapman, “Comparison of photovoltaic array maximum
power point tracking techniques”, IEEE Trans. on Energy Conversion,
vol. 22, no. 2, June 2007.
[8] A. El Khateb, N. Abd Rahim, J. Selvaraj and M. N. Uddin, “Fuzzy-
logic-controller-based SEPIC converter for maximum power point
tracking,” IEEE Trans. Ind. Appl., vol. 50, no. 4, pp. 2349-2358,
Jul./Aug. 2014.
[9] A.K. Abdelsalam, A.M. Massoud, S. Ahmed, and P.N. Enjeti, “High
performance adaptive perturb and observe MPPT technique for
photovoltaic-based microgrids,” IEEE Trans. Power Electronics, vol. 26,
no. 4, pp. 1010-1021, April 2011.

2094