Professional Documents
Culture Documents
Zkac 057
Zkac 057
5, 762–775
https://doi.org/10.1093/ce/zkac057
Advance access publication 28 October 2022
Research Article
Abstract
This work presents a machine-learning (ML) algorithm for maximum power point tracking (MPPT) of an isolated photovoltaic (PV)
system. Due to the dynamic nature of weather conditions, the energy generation of PV systems is non-linear. Since there is no specific
method for effectively dealing with the non-linear data, the use of ML methods to operate the PV system at its maximum power point
(MPP) is desirable. A strategy based on the decision-tree (DT) regression ML algorithm is proposed in this work to determine the MPP of
a PV system. The data were gleaned from the technical specifications of the PV module and were used to train and test the DT. These
algorithms predict the maximum power available and the associated voltage of the module for a defined amount of irradiance and
temperature. The boost converter duty cycle was determined using predicted values. The simulation was carried out for a 10-W solar
panel with a short-circuit current of 0.62 A and an open-circuit voltage of 21.50 V at 1000 W/m2 irradiance and a temperature of 25°C.
The simulation findings demonstrate that the proposed method compelled the PV panel to work at the MPP predicted by DTs com-
pared to the existing topologies such as β-MPPT, cuckoo search and artificial neural network results. From the proposed algorithm,
efficiency has been improved by >93.93% in the steady state despite erratic irradiance and temperatures.
Graphical Abstract
Solar panel
Merits:
Pm Tracking speed
Boost Electrical
Efficiency
converter load
Losses
Exact MPP
No periodic tuning
Duty ratio
Irriadance,
temperature
Validate
Keywords: boost converter; decision tree; maximum power point tracking; photovoltaic system; regression machine learning
the country’s independence, coal and lignite have been India’s history for the country’s growth and reaching new heights in the
primary sources of power supply. However, due to the massive Indian power industry.
emissions of heavy metals and polluting chemicals from the The sun is the most easily available energy source on Earth
stack of these traditional power generators, humans are harmed among all renewable energy sources. Solar energy is converted
by them. into electrical energy using photovoltaic (PV) cells. The PV panel
Fig. 1 illustrates the sources of installed electrical energy in should run at its MPP, or the greatest power possible, to gener-
India as a percentage. As of 30 June 2022, India’s total installed ate as much energy as possible under the specified operating
generating capacity was 403.76 GW, with thermal accounting for parameters. A maximum power point tracking (MPPT) method is
58.47%, renewable accounting for 28.25%, nuclear accounting for required since the MPP varies with the temperature and solar ra-
1.68% and hydro accounting for 11.6% [2, 3]. Coal-based power diation of the environment. An MPPT algorithm should ideally en-
accounted for 52.18% of the energy supply in the Indian electrical hance PV panel output power independently of solar irradiation,
grid. The hybridization of diverse renewable energy sources with ambient temperature and shifting weather conditions. As a re-
energy-storage technologies, which play a vital role in addressing sult, both the amount of energy created and the efficiency of the
ns 2
(yt,k − yp,k )
R2 = 1 − k=1
ns 2
(2) k=1 (yt,k − yag )
Note: share of sourses are in percentage
where the true value is yt, the predicted value is yp, the total
Fig. 1: Installed capacity of electricity in India by source.Source: Central count of samples is ns and the true values average is yag. The R2
Electricity Authority, India. Note: shares of sources are in percentages. value is in between zero and 1, which provides the prediction
strength of the model and for the best-fitted model, the R2 value
is close to 1. Similarly, the RMSE values are measures of the error
Table 1: Renewable-energy installed capacity between yp and yt. Therefore, the value of the RMSE that is close
to zero represents the model with greater strength of prediction.
Renewable energy sources Installation capacity (MW) Share (%)
This research article is organized as follows. Section 1
Wind 40 788 35.76 describes the system, which includes the specifications and
Solar 57 706 50.59 characteristics of the PV panel, the boost converter and the
Biomass 10 206 8.95 regression-tree (RT) algorithm. Section 2 describes the method-
Waste to energy 477 0.42
ology for data collection, model preparation and the MPPT RT
control approach to the solar panel. Section 3 provides the re-
Small hydro 4888 4.29
sults of the simulation and discussions. Finally, Section 4 of the
Total 114 065 100.00
study provides a comparative evaluation of the different MPPT
764 | Clean Energy, 2022, Vol. 6, No. 5
control mechanisms provided in the literature with the proposed where q is the electron charge, K is Boltzmann’s constant, cell
method, along with a conclusion in Section 5. temperature is T and the diode ideal factor is n (1 ≤ n ≤ 2). The
output current I is given by Equation (5):
q(V+IRs ) V + IR
s
1 System description I = IPH − I0 e Ns nKT − 1 −
Rsh
(5)
1.1 PV panel, characteristics and mathematical When the PV cells are connected in series–parallel and the out-
model put current I is determined, Equation (5) may be changed and is
A PV cell converts solar energy into DC electrical energy through shown in Equation (6) as follows:
a physical process called the photoelectric effect. A PV array is a q(V+IRs ) N ÅV Rs I
ã
p
collection of PV cells connected in parallel and in series to increase I = Np × IPH − Np × I0 e Ns nKT − 1 − +
(6) Rs Ns Rp
the voltage and current within the array. Fig. 2 depicts a more ac-
curate circuit model of a solar PV cell. Series resistance (Rs) and where Np and Ns indicate the number of parallel and series con-
M. Shixun Mayfly Studied under uniform and non-uniform illuminations. Can track Compared with only the PSO algorithm. 2022
et al. [24] intelligent MPP by two populations Temperature fixed at 25°C
optimization
S.R. Adaptive PV array of five Canadian Solar CS5C-80M modules serially Limited shading patterns considered. 2022
Revathy neuro-fuzzy connected to a boost converter. 231 data sets acquired from the I–V Complex rule base and system with five
et al. [25] inference and P–V characteristics layers
system
K.R. Deep learning Deep learning with back propagation neural network is employed. A Just the methodology is stated not 2022
Ahmed et two-stage control is employed simulated
al. [26]
H. Islam Proportional Results in fast-tracking and reduced power oscillations. The PV Effect of temperature is not considered. 2021
et al. [27] integral (PI) system is connected to the grid through an H-bridge inverter Proper tuning is required for PI controller
control
R.M. Asif Modified fuzzy- MPPT power loss reduction. Consider severe climatic drifts. Charging The effect of temperature on battery 2021
et al. [20] logic algorithm management control is implemented on the lead-acid battery bank charging/discharging is not considered.
to store photovoltaic energy for backup use. To maintain the state of The lifetime of the battery is not
charge (SOC) of the batteries, the charging and discharging time are considered
calculated using a C-rate scale
K. Sudoku MPPT under partial shading condition. Performance analysis carried Wiring complexity increases 2020
Rajani, T. reconfiguration out in terms of global maximum power (GMP), fill factor, mismatch
Ramesh technique losses and efficiency
[28]
S. Dorji et P&O and fuzzy For boost converter and quadratic boost converter is carried and Simulated under only standard test 2020
al. [6] logic control compared conditions. Dynamic weather conditions
(FLC) not considered
X. Li et al. β parameter- A novel β-parameter-based FLC is coined with three inputs and one The temperature effect is not considered. 2019
[22] based FLC output, and this β method will reduce the number of membership The sensitivity of V and I sensing devices
functions are not taken into account
Maximum power point tracking using decision-tree machine-learning algorithm for photovoltaic systems | 765
Rs
I A
0.8
Ish + 0.6
Current (A)
ID
25 °C
0.4
IPH Rsh V
0.2 35 °C
45 °C
0
0 5 10 15 20 25
–
Voltage (V)
B
Power (W)
A 35 °C
0.8 5 45 °C
1 kW/m2
0.6
0.8 kW/m2
Current (A)
0
0.4 0.6 kW/m2 0 5 10 15 20 25
0.4 kW/m2 Voltage (V)
0.2 0.2 kW/m2
Fig. 4: Characteristics of PV panel with provided temperatures at
1000 W/m2.(a) V vs I; b) V vs P.
0
0 5 10 15 20 25
Voltage (V) Ipv L Diode
IL Io
B
+
10 1 kW/m2 +
+
Pm ICi IC0
PV panel
8 0.8 kW/m2
Co Ro Vo
Vpv
Power (W)
6 0.6 kW/m2 Ci
MOSFET –
4 0.4 kW/m2 –
–
2 0.2 kW/m2
0 Fig. 5: The PV supplied boost converter circuit.
0 5 10 15 20 25
apacitive filter smoothes the switching action’s pulsing current
c
Voltage (V)
while supplying a DC voltage to the load.
Fig. 3: Characteristics of PV panel with provided irradiances at 25˚C.(a)
V vs I; (b) V vs P.
2 DT ML algorithm
A DT is a ML algorithm generally used for regression and classifi-
In this work, a PV panel whose specifications are a maximum cation problems. The output of a classification-type tree is a class
power capacity of 10 W with a short-circuit current of 0.62 A and or label of the data, whereas the RT output is a real value. A RT
an open-circuit voltage of 21.50 V, the voltage and current at MPP is a DT meant to approximate real-valued functions rather than
are 17.50 V and 0.57 A, respectively, and with 36 solar cells used be utilized for categorization. The input variables may include
for simulation. Figs 3 and 4 show the I–V and P–V characteristics both continuous and categorical data when using the standard
of the panel for a range of temperatures and irradiances. RT building method. The tree is referred to as a DT when each
decision node tests the value of an input variable. The terminal
1.2 DC–DC boost converter nodes of the tree contain the predicted output variable values.
Fig. 5 shows a DC–DC boost converter powered by a PV panel and Binary recursive partitioning is an iterative approach that splits
controlled by pulse width modulation (PWM). The metal oxide data into branches or partitions and divides each branch into
field effect transistor duty ratio (D) influences the power trans- smaller units as the procedure continues along each partition.
ferred from the panel to the load. The circuit increases the PV The training set’s records (pre-classified data used to define the
voltage to the expected output level by an inductor (L). The input tree’s architecture) are first sorted into the same division. The
capacitor (Ci) and output capacitor (Co) minimize the ripple con- approach then separates the data into two segments or branches,
tent of the output voltages. The boost inductor’s current grows using each field’s possible binary split.
linearly while the switch is in the on position and the diode is The method chooses a division that reduces the sum of squares
off. The inductor’s stored energy is released when the switch is of the deviations from the mean of the two independent divisions.
off, passing via the diode and onto the output RoCo circuit. The Then this rule is used to split each new branch. This operation
766 | Clean Energy, 2022, Vol. 6, No. 5
(12)
The reduction in MSE is (if xi has missing values, with an as-
sumption that there are randomly missing observations) given by
No Missing Yes
Equation (13):
indices?
(13) ∆IU = P (Tt − TU ) εt − P (TL ) εtL − P (TR ) εtR
Compute ∆I Compute ∆IU where Tt – TU is the collection of all valid observation indices for
node t.
3 Methodology
Split the predictor variable at the cut point where the The proposed approach consists of two steps and involves build-
MSE reduction is greatest. ing the RT ML model using the data obtained based on the PV
panel parameters and using the prepared model for MPPT. First,
since irradiance (Ir) and T are functions of the maximum power
Stopping No (Pm) and the respective voltage (Vm) at MPP, Ir and T are used as
criteria? features to predict the data for Pm and Vm. Therefore, the created
models are utilized for the specified Ir and T to predict the PV
Yes panel’s Pm and Vm. Then, the predicted values are used to deter-
End mine the duty cycle (D) to drive the PV panel at MPP.
Fig. 6: Flowchart of node-splitting procedure for CART. 3.1 Data collection and model preparation
Ir, T, P and V make up the data needed to train and test the model.
Fig. 8 shows a flowchart of the conceptual process of collecting
continues until each node reaches the user-specified minimum data and creating a ML model.
node size and becomes a terminal node. Even if a node does not
meet the minimum size requirement, if its sum of squared devi-
3.2 MPPT with boost converter and RT model
ations from the mean is zero, it is considered a terminal node. The The prepared ML model predicts the available maximum power
stepwise procedure for standard classification and regression tree in the PV panel (Pm) and the respective voltage (Vm) for the fea-
(CART) with all predictors xi (i = 1,2,..., p) to split node t is shown in tures Ir and T. Using the predicted values Pm, Vm as in Equation
Fig. 6. The basic structure of a DT is shown in Fig. 7. (14), the resistance Rmp that corresponds to MPP is computed. By
The weighted mean squared error (MSE) of the response in adjusting the D of the converter, the Rmp will be reproduced be-
node t is given by Equation (10): tween node-n1 and node-n2 in Fig. 9. In Equation (15), the load
resistance (R0) and Rmp provide the converter’s D:
2
εt = wj (yj − yt )
V2
(10) j∈ Tt
Rmp = m
(14)Pm
where wj is jth observation weight and the set of all observation
indices in node t is denoted by Tt. If no weights are specified, use Rmp
wj = 1/n and n as the sample size. Probability estimation is given D=1−
(15) R0
by Equation (11):
The approach suggested by Ayop et al. [32] can be used to es-
P (Tt ) = wj timate the highest and lowest values of the resistance for the
(11) j∈ Tt load. The process outlined by Rashid [33] was used to design
Maximum power point tracking using decision-tree machine-learning algorithm for photovoltaic systems | 767
Use equations (5) to (9) and Vip and Vop are the input and output voltages in volts, re-
spectively. Fig. 9 shows the block diagram of the PV panel asso-
The simulation was run at four intervals of 0.5 sec each for a
using MATLAB
Prepare the regression tree using the DATA total of 2 sec to determine the tracking accuracy of the RT under
various irradiances and temperatures. For each interval, the val-
ues of Ir or T are changed while keeping the other fixed. Table 3
Train
displays this change.
RT Model
1000
800
Irradiance
600
400
200
50
40
35
30
25
17.5
17.0
Max voltage
16.5
16.0
15.5
10
8
Max power
0
250 500 750 1000 30 40 50 15.5 16.0 16.5 17.0 17.5 0.0 2.5 5.0 7.5 10.0
Irradiance Temperature Max voltage Max power
5 Comparative analysis where ipv and vpv, respectively, stand for the output current and
voltage of the PV module. The diode constant is C, q is the elec-
The developed control strategy’s outcomes are compared to those
tron charge (1.6 × 10–19 coulomb), n is the ideal diode factor and the
of the β-MPPT method [8], which incorporates the P&O algorithm
Boltzmann constant is K, 1.38 × 10−23 J/K, the p–n junction tempera-
[5], the CS optimization [12] approach and a perceptron artificial
ture is T in Kelvin and N is the number of PV cells in the module.
neural network (PANN) algorithm [18, 19].
This approach’s transient and steady-state stages use variable
and fixed steps, respectively. Fig. 18 shows the flow chart for this
5.1 β-MPPT method strategy. Before continually calculating the values of β, the cur-
The fundamental idea behind the β-MPPT approach is to monitor rent and voltage must first be monitored. The Beta technique en-
an intermediate variable called β instead of the change in power, ters the steady-state stage if the β is inside the bounding range of
which is denoted by Equations (18) and (19): the (βmin, βmax). Otherwise, it enters the temporary stage, in which
Å ã the P&O method is applied. In the interim stage, the variable step
ipv
β = ln − C × vpv size ΔD is calculated using a guiding parameter βg, which may be
(18) vpv
written as Equation (20):
q
C= ∆D = F × (β − βg )
NnKT
(19) (20)
Maximum power point tracking using decision-tree machine-learning algorithm for photovoltaic systems | 769
where F is the scaling factor. the PV system from working at MPP. The system can be operated
The temperature and irradiance affect the β parameter’s range. at MPP in a steady state with more precision using the proposed
In this study, βmin = 15.45 and βmax = 19.02 are used with a scaling RT approach compared with the β-MPPT method.
factor F = 0.01. The average value of βmin and βmax was used as
βg = 17.24.
Fig. 19 compares the power response of the β-MPPT (Pbeta), the 5.2 CS method
maximum power predicted by RT (Pm) and the control technique The CS method is one of the rapid-converging swarm optimiza-
using the RT algorithm (PRT). For low irradiance levels, the β-MPPT tion techniques. The CS approach is implemented as a flow chart
method operates at MPP, but for high irradiance values, this tech- in Fig. 20 and creates four new D samples using a series of equa-
nique fails and exhibits a significant amount of error, preventing tions from Equations (21)–(24):
(t+1) (t)
(21) Di = Di + α ⊕ l evy(λ); i = 1, 2, 3, 4.
1.00 where α = αo (Dbest − Di ).
tion [12]:
1 0.52 –0.13 1 0.75 Ç å
u
αo (Dbest − Di ) ⊕ levy(λ) ≈ k1 × 1/β
(Dbest − Di )
(22) |v|
0.50
where β = 1.5 and k1 (the Levy multiplication coefficient) are equal
Temperature
(Pcs) and the PRT. The CS method has dynamic behaviour. Initially,
1 0.46 –0.064 1
the CS response has an undershot, which is undesirable. For
–0.75 small irradiance values, the CS gives good results in tracking MPP.
For high irradiance levels, such as 800 W/m2, the CS technique
Irradiance Temperature Max voltage Max power overshoots and fails to operate the system on MPP, leading to a
severe steady-state error. Compared with the CS approach, the
Fig. 11: Heat map of data correlation.
Table 3: Predicted values by RT models, D and percentage mean efficiency for all intervals
x1 < 162.5 x1 >= 162.5 x1 < 437.5 x1 >= 437.5 x1 < 687.5 x1 >= 687.5 x1 < 912.5 x1 >= 912.5
9.6536 9.231
17.57 17.59
1
Estimates
Predicted data
X: 1
5
Y: 1.142
0.5
0 X: 2
50 Y: 0.001464
40 1000
Tem 30 600 800 0
pera
tu 20 200 400 2) 1 2
re (°C 0 e (w/m
) Irridanc Predictors (x1,x2)
X: 2
Estimates
16 Y: 0.0161
Actual data
0.01 X: 1
14 Predicted data Y: 0.005333
50 1000 0.005
40
Tem 500
pera 30 m2 ) 0
ture a n c e (w/ 1 2
(°C) 20 0 Irrid
Predictors (x1,x2)
Fig. 14: Actual vs predicted data.(a) RT-1 model on the Pm plane; (b) RT-2
model on the Vm plane. Fig. 15: Predictor importance estimates for (a) RT-1 and (b) RT-2.
50 Vload Begin
40 VPV
Voltage (V)
30
20 Measure v(k), i(k)
10
0
Calculate β = In[i(k)/v(k)] – C × v(k)
0 0.5 1 1.5 2
0.5
Iload Yes
0.4 βmin < β < βmax
IPV
Current (A)
0.3 Steady
Time (sec)
Power (W)
6
Fig. 16: V, I and P waveforms of the solar panel and load with RT
models.
4 Pm
PRT
100 2
Pbeta
% Mean efficiency
80
0
60 0 0.5 1 1.5 2
Time (sec)
40
Fig. 19: Power response comparison of Pm, PRT and Pbeta.
20
6 Pm node is Wj,o, the error value at the output node o is eo, the weighted
PPV sum of the hidden nodes together with the bias bo value is vo and
4
the derivative value of the purelin function is φ’(vj).
2
To train, validate and test the model, the Levenberg–Marquardt
0 optimization method was used with the same data set used for
the proposed methodology. Equation (29) contains the weight-
0 0.5 1 1.5 2
updating formula for determining updated weights (Wi,j* and Wj,o*).
Time (sec)
Pm and Vm were predicted using this PANN model. The train, valid-
Fig. 17: Percentage mean efficiency, Pm and Ppv waveforms with RT ate and test regression curves in Fig. 23 demonstrate that R2 is ~1:
models. −1 T
Wi,j ∗ = Wi,j − JT J − lp I J e
−1 T
where xi is the present value at the ith (i = 1, 2) input node, Wi,j Wj,o ∗ = Wj,o − JT J − lp I J e
(29)
stands for the weight value between the ith input node and the jth where J is the Jacobian matrix, I is the identity matrix, e is the cu-
hidden node, the learning rate lr ranges from 0 to 1, the error value mulative error vector and lp is the learning parameter.
at the jth hidden node is ej, vj stands for the sum of the weighted The Pm predicted by the RT model, PANN method (PNN) and PRT is
inputs including the bias bj and the derivative value of the tansig compared in Fig. 24. For low irradiance levels, the PANN response
function is ϕ’(vj): is almost the same as that of the proposed method in this work.
772 | Clean Energy, 2022, Vol. 6, No. 5
Start 10
8
Initialize 4 nest positions (duty cycles)
D1 = 0, D2 = 0.3, D3 = 0.5, D4 = 0.9
set the search limits Dmin = 0, Dmax = 1 6
Power (W)
4
Calculate the fitness function of each nest
(power values P1, P2, P3, P4) Pm
2
PRT
Specify the best duty cycle (Dbest) 0
Pcs
which gets the (Pmax) from the PV panel
–2
5 5
5 5
5 10 15 5 10 15
Target Target
10
4
8
8
6 3.5
Power (W)
0.5 0.6 7
6
4
1 1.05
Pm
2 PRT
PNN
0
0 0.5 1 1.5 2
Time (sec)
Fig. 24: Comparison of the power response of Pm, PRT and PNN.
Table 4: Comparison of numerical MPP tracking response for different methods during 0–0.5 sec
4.5 In: Eltamaly AM, Abdelaziz AY, (eds). Modern Maximum Power
2 Pm Point Tracking Techniques for Photovoltaic Energy Systems. New
4 York: Springer, 2020, 379–400.
PRT
1 [13] Wei L, Li K. Research on the maximum power point tracking
Pcs method of photovoltaic based on Newton interpolation-
0.4 0.45
PNN assisted particle swarm algorithm. Clean Energy, 2022,
0 6:496–502.
[28] Rajani K, Ramesh T. Maximum power enhancement under systems operating under actual environmental conditions.
partial shadings using a modified Sudoku reconfiguration. Computational Intelligence and Neuroscience, 2017, 2017:1673864.
CSEE Journal of Power and Energy Systems, 2020, 7:1187–1201. [32] Ayop R, Tan CW. Design of boost converter based on maximum
[29] Phan BC, Lai YC, Lin CE. A deep reinforcement learning-based power point resistance for photovoltaic applications. Solar
MPPT control for PV systems under partial shading condition. Energy, 2018, 160:322–335.
Sensors, 2020, 20:3039. [33] Rashid MH. Power Electronics: Circuits, Devices & Applications. 4th
[30] Carballo JA, Bonilla J, Berenguel M, et al. Machine learning for edn. London: Pearson, 2004.
solar trackers. AIP Conference Proceedings, 2019, 2126:030012. [34] Kim P. MATLAB Deep Learning With Machine Learning, Neural
[31] Shareef H, Mutlag AH, Mohamed A. Random forest-based ap- Networks and Artificial Intelligence. 1st edn. Berkeley, CA: Apress,
proach for maximum power point tracking of photovoltaic 2017.