Machine learning based maximum power point tracking in

solar energy conversion systems

SR.Jeevaa,S.Mathivannan and P.Akash

Abstract
Extraction of solar energy from photovoltaic cells has different efficiencies corresponding to different algorithms. In
the paper, a power efficient algorithm is suggested for tracking of maximum power point (MPP) in solar energy
conversion systems by implementing machine learning (ML) in the pre-existing perturb and observe (P&O)
methodology. P&O works on the principle of varying duty cycles step by step in the direction of the MPP and is the
most feasible and accurate algorithm. However, the speed of convergence to the MPP is usually less in this method
and it varies in different climatic conditions. This paper describes the application of ML in decreasing the
perturbation time significantly leading to the significant increase in the efficiency to predict the MPP. The suggested
algorithm predicts an MPP based on instantaneous values of solar irradiation, solar cell temperature and humidity as
input features to the localized multivariate regression ML model and is used to fetch maximum available power
(MAP). It is a self-learning algorithm and as the time progresses, the estimation becomes much closer to the
theoretically available power. The simulation was done in python and yielded an average efficiency of 99.8% in
estimating the MPP after 83 hours of training.

1. Introduction

To date, fossil fuels are responsible for fulfilling eighty percent of the world’s total energy demands.
Fossil fuels being a non-renewable resource exist in limited amounts and their unchecked use will lead to
complete consumption within the next few decades. Moreover, global warming is caused by greenhouse
gases released as by-products as a result of generation of energy from fossil fuels [1]. Thus, it is high time
to start looking for alternatives to fossil fuels. Solar energy being the earth’s most readily available source
of energy can turn out to be an excellent substitute and can help solve the energy crisis faced by the world.
Photovoltaic cells are used to generate electrical energy from solar energy but being non-linear in
characteristics, they are highly inefficient. Therefore, it is difficult to extract maximum power from solar
PV cells in varying climatic conditions. The process of generating maximum possible power from the PV
cells is known as maximum power point tracking (MPPT). A variety of MPPT algorithms like the method
of incremental conductance [2], the method of fractional voltage [3], perturb and observe [4], fuzzy logic
control [5], etc., have been developed and are being used in the industry over the years.
There is a need for an effective mechanism to quickly estimate the maximum power point due to
continuously changing intensity of solar radiation. A solar energy conversion system is usually made up
of a solar PV array, charger controllers and an interconnection framework to supply the power generated
for further distribution. When the solar radiation falls on the PV array, the cells are excited leading to
generation of photocurrents which are further directed to load impedances for power consumption. The
photocurrent and power versus voltage characteristics of a solar cell for different intensities of irradiation
are shown in diagram.

Photocurrent and power versus voltage characteristics for different intensities of sunlight.

A charger controller fed with an efficient algorithm is responsible for carrying out MPPT. The
generated power is optimized and controlled via its design. Commonly used MPPT algorithms have been
compared in Table 1[6] based on their efficiency and complexity.
Table 1. Comparison of P&O with other MPPT techniques
MPPT Technique Speed of Convergence Implementation Periodic Tuning Sensed Parameters
Complexity
Perturb and Observe Varies Low No Voltage
Incremental Conductance Varies Medium No Voltage, Current
Fractional VOC Medium Low Yes Voltage
Fractional ISC Medium Medium Yes Current
Fuzzy Logic Control Fast High Yes Varies
Neural Network Fast High Yes Varies

In this paper, an alternative method is described to overcome the limitations of the existing P&O
methodology and other MPPT algorithms. The given model uses machine ML before the conventional
P&O algorithm to estimate MPP. ML is a type of artificial intelligence that trains the system to learn and
doesn’t require manual programming. Its main focus is on the improvement of systems that can learn on
their own and vary accordingly on exposure to new information and data sets. Results have been analyzed
and compared with the pre-existing methods on the basis of performance and efficiency.

2. System Model

The system consists of a solar panel coupled with a buck converter. A buck converter works on the
concept of duty cycles using pulse width modulation (PWM) signal to control the ratio of input to output
voltages thereby acting as a step-down DC to DC voltage converter. The PWM signal is managed by the
MPPT charger controller. Fig. 2 describes the proposed solar energy conversion model and its functioning.
 Machine Learning based Maximum Power Point Tracking in Solar Energy Conversion Systems

Proposed solar energy conversion model and its configuration

2.1. The solar photovoltaic cell model

Photocurrents in the solar cells are generated by the photovoltaic cells and they act as variable current
sources. A solar cell is nothing but a PN junction diode which generates current on exposure to sunlight
[7]. The generated current depends linearly on the solar illumination/irradiance. Fig. 3 shows the basic
circuit analogy for a solar PV cell.

Fig. 3. Functional circuit of a singular solar PV cell

Here, IL is the photocurrent (in Amps),

ID is the diode current (in Amps),
Rsh is the shunt resistance (in ohms),
Ish is the current passing through the shunt resistance (in Amps),
RS is the series multiplier resistance (in ohms),
I and V are the output voltage and current.
Following sets of equations give the current-voltage characteristic of the solar PV circuit:
ID = ISS[e(q(V+IRS)/KT)-1] (1)
Here, the output current I is given by:
I = IL – ID – Ish (2)
thus,
I = IL – ISS[exp(q(V+IRS)/KT)-1] – (V+IRS)/Rsh (3)
Here, q is the charge on an electron (in coulombs), K is the Boltzmann constant (in Joule/Kelvin) and
T is the solar cell temperature (in Kelvin).

2.2. Buck converters

Buck converters are essentially step-down DC-to-DC voltage converters that work on the principle of
varying the output current using energy storage elements [8]. They work in the switched mode power
supply (SMPS) and typically contain at least two semiconductors (a diode and a transistor), energy
storage elements like an inductor, a capacitor or a combination of two for the step-down operation. Buck
converters are remarkably efficient (up to 90%) and are therefore useful for a variety of computational
operations. They control the duty cycles for the stepping down operation using a pulse width modulator
(PWM) signal. The On and Off states responsible for changing the direction of duty cycles are controlled
via SMPS as depicted in Fig. 4.

Figure depicting On and Off states of a buck converter

2.3. The Perturb and Observe (P&O) Algorithm

P&O is one of the most simplistic and accurate algorithms for MPPT in solar energy conversion
systems. A charger controller regulates the output supply of the solar PV array by controlling the pulse
width modulation (PWM) based duty cycles of the buck converter and keeps a track of fall and rise in
power constantly. If the power is incremented upon increasing the buck converter duty cycle, the duty
cycle is increased in the same direction further till the MPP is achieved. Otherwise, if there is a fall in
power, the direction of the duty cycle should be reversed and the same procedure has to be repeated. This
point is the desired MPP and the characteristic parameters of this point are recorded and used for
generating optimal power. This methodology is termed as P&O and is the most commonly used method
in the electrical industry. A flow diagram demonstrating perturb an observe along with ML is shown in
Fig 5.

Flowchart demonstrating P&O

 Machine Learning based Maximum Power Point Tracking in Solar Energy Conversion Systems

2.4. MPPT algorithm

In this model, ML is used to estimate a power point (Pref) close to the theoretical MAP based on a
localized multivariate regression model [9] with the characteristic inputs (Xik in Eq.5). The estimation is
done to decrease the time of perturbation in P&O algorithm. At the end of every iteration, the inputs solar
cell temperature, solar irradiance and humidity (X k in Eq. 5) of the previously detected MPP(s) are fed
into the ML module as training data along the MAP. This model learns by itself based on regression and
pattern recognition of the MPP(s)from the characteristic parameters (b k and Xk). The P&O upgrades the
estimation dataset at the end of every iteration thereby reducing the error (ϵiin Eq. 5). After every iteration,
this model goes undergoes continuous refining using gradient descent algorithm as depicted in Fig. 5.

ϕt = Ft(ϕ0,…,ϕt-1) (4)

Here ϕt is the MPP at tth iteration. The ML algorithm uses the Ft(Xil,…,Xik) = Yt for estimation using
the dataset of previous MPP(s) (ϕ0,…,ϕt-1). The calculated result is further taken for the next set of
iterations as Pref = ϕt after which the algorithm undergoes a regular P&O using a buck converter for
learning and updating the algorithm. A localized regression model can be used to compute by Equation 5.

Yi = b0 + b1X1k + . . . + bkXik + ϵi (5)

Here, bk are coefficients of regression, ϵi is the error at ith iteration and Xik are regressor/input variables.
Y’ is the mean MPP of all the MPP(s), Yi using variable inputs Xik(temperature, irradiance, humidity) and
k predictor variables. The bk values or the regression weights are compared to minimize the squared
deviation sum as depicted in Equation 6. Here, N is the number of iterations.

(6)

The Gradient Descent algorithm [10] minimizes the sum of squared deviations as Cost Function J(θ)
which allows the model to adapt to the dataset and give efficient and accurate outputs in further iterations.

Fig. 6. Figure depicting Gradient Descent algorithm used by ML to minimize the Cost Function J(θ) with two cost
parameters leading to reduced errors in further predictions

3. Experimental Results

Graph between estimated power point (blue) and maximum available power (red) is illustrated in Fig.
7 and Fig. 8. It can be clearly seen in the diagrams that the estimated power points are remarkably close to
the actual available power in iterations ranging from 900 to 1000 as compared to iterations ranging from 0
to 100. These figures justify the learning accuracy of the algorithm. These results were obtained after 83
hours of training the ML model with all iterations being 5 minutes apart each.
Graph showing variation of MAP (P : Red) and Estimated MPPs (φ : Blue) with respect to time (0-100
iterations).

Fig. 8. Graph showing variation of MAP (Pt: Red) and Estimated MPPs (φt: Blue) with respect to time (900-1000
iterations).

∆ = 100*|Pt −φt| / | Pt| % (7)

The percentage error ∆ can be computed by using Equation 7. Average error for iterations 950-1000 is
computed to be 0.2%. Fig. 9 shows the error percentage in the iteration set of 950-1000. This graph
verifies the increase in efficiency after every iteration. Similarly, average efficiency can be calculated by
using Equation 8.

ηt = (100−∆t[avg]) % (8)
The average efficiency of the model is at 99.8% after around 1000 iterations corresponding to 83 hours
of training. Thus, Fig. 9 clearly depicts that more the training, lesser the likelihood of an error in
prediction and the error threshold never crossed 0.5%. The training set included real time data from BITS
Pilani K.K. Birla Goa Campus.
 Machine Learning based Maximum Power Point Tracking in Solar Energy Conversion Systems

Percentage error between MAP and Estimated MPP with respect to time (iteration 950-1000).

4. Advantages of Using Machine Learning

P&O combined with ML converges much faster to the MPP compared to the conventional P&O. The
proposed model learns quickly and is much more efficient compared to similar AI algorithms such as
artificial neural networks and deep learning. Moreover, they do not provide high performance and
increased accuracy. Misclassification of input data by simple addition of small perturbations can easily
confuse such algorithms resulting in prediction of inaccurate values. These perturbations occur practically
all the time due to continuously changing climatic conditions. But in the proposed method, the learning is
not affected because of the inclusion of supervised learning. The results’ complexity with respect to the
time taken for estimation is also directly affected by neural networks work on added hidden layers. The
various features of the proposed algorithm are mentioned in Table 2.
Table 2. Comparison of MPPT Techniques with proposed model
Technique Learning Phase Complexity Accuracy Time Taken
ML into P&O Yes (Supervised) High High Low
P&O No Low High High
Fuzzy Logic Control No High Medium Medium
Neural Networks Yes High Medium Low

5. Conclusion

An efficient method to track maximum power point under varying climatic conditions is described in
this paper. A python simulation of a solar energy conversion system has been carried out to validate the
proposed MPPT method. The analysis showed that the proposed MPPT method estimated the MPP with
higher frequency. Observations show that the performance and the accuracy of the proposed method is
not affected by the variations in input data or small fluctuations and neither during nasty and rapid
weather changes. The perturbation time decreases with each iteration. The proposed algorithm gradually
learns and incorporates to the new data at the end of each iteration. Machine Learning in the proposed
model overcomes overfitting that usually occurs in other AI based MPPT algorithms such as artificial
neural networks and other deep learning algorithms. The main advantages of the proposed MPPT control
method are faster convergence to the MAP, higher efficiency than its AI counterparts (ANN), robustness,
ease of implementation and its ability to learn from previous data irrespective of the season of the year.
Thus, much faster convergence speeds are obtained from the proposed algorithm which being a modified
version of P&O can be directly cascaded to the existing P&O equipment with convenient setup.
References

[1] Jaccard M. Sustainable Fossil Fuels: The Unusual Suspect in the Quest for Clean and Enduring Energy. Cambridge
University Press; 2006.
[2] Safari A, Mekhilef S. Simulation and hardware implementation of incremental conductance MPPT with direct control method
using Cuk converter. In: IEEE Trans. Ind. Edition, Vol. 58, no. 4, 2011:1154-1161.
[3] Ahmad J. A fractional open circuit voltage based maximum power point tracker for photovoltaic arrays. Presented at: 2010
2nd International Conference on Software Technology and Engineering.
[4] Nedumgatt J J, Jayakrishnan K B, Umashankar S, Vijayakumar D, Kothari D P. Perturb and observe MPPT algorithm for
solar PV systems-modeling and simulation. Presented at: 2011 Annual IEEE India conference.
[5] Alajmi BN, Ahmed KH, Finney SJ, Williams BW. Fuzzy-logic-control approach of a modified hill-climbing method for
maximum power point in Microgrid standalone photovoltaic system. In: IEEE Trans. Power Electron, vol. 26 no. 4, 2011:
1022-30.
[6] Esram T, Chapman PL. Comparison of photovoltaic array maximum power point tracking techniques. In: IEEE Trans.
Energy Convers., 2007; 22(2): 439–49.
[7] Jha AR. Solar Cell Technology and Applications. CRC Press; 2009.
[8] Thandayithabani MP, Voltage Mode Control for Buck Converter. 2012.
[9] Olive DJ. Multivariate Linear Regression. In: Linear Regression, 2017:343-387.
[10] Cristianini N. Gradient Descent (Steepest-Descent Method). In: Dictionary of Bioinformatics and Computational Biology.
2004.
NEURAL NETWORK AND FUZZY
CONTROL

JEEVA SR 20BEE0238
MATHIVANNAN S 20BEE0243
AKASH P 20BEE0023
MAXIMUM POWER POINT TRACKING
Abstract :
To extract Maximum Power from PV system, many MPPT
techniques have been employed by the researchers. The Duty Cycle
of the DC-DC Boost converter is controlled by the proposed Fuzzy
MPPT control Technique and hence Maximum power is obtained
from PV system. Simulations were carried out and the proposed
MPPT technique is tested under different environmental
conditions. The proposed method is effectively tracking the
maximum power for all environmental conditions. The output
voltage of the DC-DC Converter is varying one with respect to
Maximum Power Point (MPP). One more DC-DC Buck-boost
converter with PI control is used in order to make the DC Bus
voltage constant. A PWM Inverter with L-filter is used to generate
three phase supply required to operate three phase loads and
single phase loads. The results are presented for different
irradiation and temperature conditions.
MAXIMUM POWER POINT TRACKING
Introduction:
The solar PV system consists of PV module, DC-DC
Converters, MPPT control, Fuzzy Logic Control and a
load. The single solar PV module consists of series and
parallel solar cells, which is used to increasing the
voltage and current. The output of solar PV module
current is given to the input of a current controlled
source. The single solar cell does not provide the
maximum power, so the numbers of solar cells are
connected in series and parallel and improve the output
power. The output power system is connected to boost
converter and to track the maximum power using Fuzzy
logic controller.
MAXIMUM POWER POINT TRACKING
 Maximum power point tracking is a technique used
commonly with wind turbines and photovoltaic solar system
to maximize power extraction under all conditions.
 The objective of the project is to design a Maximum Power
Point Tracking (MPPT) charge collector which operate with
photovoltaic module and produce maximum power to solar
power collector. This component optimized the amount of
power obtained from the photovoltaic array and is given to a
load.
TYPES OF PERFORMING MPPT:
There are many different methods to perform MPPT.
Some of them are:
 Perturb and Observe

 Incremental conductance

 Current sweep

 Constant voltage
 CIRCUIT DIAGRAM