Implementing ML and DL based controllers for

improving efficiency in EV charging application

Grisham Garg Dhruv Talan Himanshu
Dept. of Electrical Engineering Dept. of Electrical Engineering Dept. of Electrical Engineering
Delhi Technological University Delhi Technological University Delhi Technological University
New Delhi, India New Delhi, India New Delhi, India
grishamgarg [email protected] dhruvtalan ee20a13 [email protected] himanshu ee20a13 [email protected]

Gaurav Yadav Mukhtiar Singh

Dept. of Electrical Engineering Dept. of Electrical Engineering
Delhi Technological University Delhi Technological University
New Delhi, India New Delhi, India
gauravyadav [email protected] smukhtiar [email protected]

Abstract—This paper analyzes the potential of various Ma- They require specialized electrical infrastructure and cater to
chine Learning and Deep Learning algorithms for replacement the demand for rapid charging, playing a crucial role in reduc-
of the traditional Proportional Integral (PI) Controller in a ing charging downtime for EVs and improving the efficiency
single-phase two-stage on-board Electric Vehicle (EV) charging
system to improve its adaptability and efficiency. Traditional of charging networks. These chargers are integral components
PI controllers, though reliable, often do not perform optimally of EV charging systems, which can utilize either single-
under the dynamic demands of EV charging, leading to in- phase or three-phase power supplies depending on application
efficiencies. Our investigation aims to understand how these needs. Furthermore, chargers can be categorized as onboard
ML/DL methodologies can better handle the variable nature or offboard, with onboard chargers integrated into the EV and
of EV charging through adaptability, accuracy, and efficiency
in managing charging operations. Experimental results from responsible for AC to DC conversion, while offboard chargers,
both simulations and real-world data reveal the effectiveness of located externally, supply power to the vehicle’s onboard
these ML/DL models in enhancing charging system performance charger. Fast-charging technologies, such as DC fast charging,
over standard PI controllers. Decision Trees provide clear, rule- are becoming increasingly prevalent, offering EV owners the
based control paths; Neural networks excel in modeling complex, convenience of rapid charging during long journeys.
non-linear relationships for real-time predictions; and Linear
Regression offers a straightforward approach for linear charging A Proportional-Integral (PI) controller is a fundamental
dynamics. feedback control system utilized in engineering and automa-
Index Terms—PI Controllers, Decision Trees, Artificial Neural tion. It functions by continuously computing an error signal,
Networks, Regression representing the disparity between the desired setpoint and
the actual value of the process variable. The controller then
I. I NTRODUCTION adjusts the system’s output to minimize this error, combining
Electric vehicles (EVs) rely on chargers to replenish their proportional and integral control actions. Traditional PI con-
batteries, with single-phase and three-phase chargers being trollers face several challenges in addressing the complexities
primary options. single-phase chargers, utilizing a single alter- of modern systems. These challenges include difficulties in ac-
nating current (AC) waveform, offer moderate charging speeds curately modeling nonlinear dynamics, handling disturbances
suitable for residential settings and light-duty commercial use, and uncertainties, and adapting to changing operating condi-
boasting ease of installation and compatibility with most EVs. tions. Linear Regression (LR) models offer a potential solution
They are commonly found in home charging setups, providing by providing a simple yet effective means of capturing linear
convenient overnight charging for EV owners. Single-phase relationships within the system, allowing for better prediction
chargers typically operate at lower power outputs, ranging of system behavior. Artificial Neural Networks (ANNs) offer
from 3 kW to 7.4 kW, making them cost-effective solutions. a more flexible approach by leveraging their ability to learn
While they may not offer the rapid charging speeds of three- complex nonlinear mappings from data, enabling them to
phase chargers, single-phase chargers are sufficient for meeting handle nonlinear dynamics and adapt to changing conditions.
the daily charging needs of many EV drivers, especially Decision Trees (DTs) provide interpretable models that can
in residential environments. In contrast, three-phase chargers capture complex decision-making processes, aiding in un-
leverage three AC waveforms, providing significantly higher derstanding system behavior and identifying optimal control
power output and faster charging rates, making them ideal for strategies.
commercial and public charging stations. By incorporating these advanced modeling techniques, we
can enhance the performance and robustness of control sys- In decision tree analysis, a fitness function evaluates the
tems, ultimately overcoming the limitations associated with effectiveness of divisions within the tree, generally employ-
traditional PI controllers and better addressing the challenges ing criteria like Gini impurity or information gain. Consider
faced in modern applications. a training set L, where L = (x1 , c1 ), (x2 , c2 ), ..., (xi , cj ),
with (x1 , x2 , ..., xi ) being a set of input feature vectors and
II. M ATERIAL AND M ETHOD
(c1 , c2 , ..., cj ) representing corresponding class labels. In this
A. Linear Regression context, xi represents a vector comprising input attributes, and
Linear regression is a fundamental statistical method used the conditions for splitting are determined based on one of
to model the relationship between a dependent variable Y and these attributes. Let pi denote the likelihood that a random
one or more independent variables X1 , X2 , X3 , X4 , ..., Xn ,It sample falls into class ci, then pi can be calculated as
aims to find the best-fitting linear equation that describes how
ci
the independent variables X predict the dependent variable Y pi = (3)
.The equation for a simple linear regression model with one L
predictor variable X is represented as:- 1) Entropy: Entropy is a metric used to quantify the disor-
der or variability within a dataset, foundational for computing
Y = β0 + β1 X + ϵ (1) information gain. In situations where subsets of data are
• Y is the dependent variable. consistent, or uniform, the dataset is considered to have no
• X is the independent variable. disorder or unpredictability. A dataset whose subsets all pertain
• β0 is the slope coefficient, representing the change in Y to a single classification will exhibit an entropy value of zero.
for a one-unit change in X The formula for entropy is the aggregate of the product of
• ϵ is the error term, capturing the difference between the the probability of each class label and the logarithm of this
observed and predicted values. probability.
In multiple linear regression, the equation extends to include j
X
multiple predictor variables Entropy(L) = − pi log2 (pi ) (4)
i=1
Y = β0 + β1 X1 + β2 X2 + · · · + βp Xp + ε (2)
2) Information Gain: Information gain measures how much
where X1 , X2 , ..., Xp are the predictor variables and a feature f reduces uncertainty or entropy in a dataset before
β0 , β1 , β2 , ..., βp are their respective coefficients. and after it’s divided based on f . It indicates the effectiveness
One of the key strengths of linear regression lies in its of f in separating the dataset L into subsets |Lv |, where each
simplicity and interpretability. The linear relationship between subset represents data with a distinct value of f . A higher
the variables can be easily visualized using plots and un- information gain suggests a feature is better at improving
derstood through the coefficients of the regression equation. classification clarity. The best feature for splitting is the one
This transparency makes linear regression particularly useful with the highest information gain, implying it offers the most
for exploratory data analysis and communicating findings to significant reduction in uncertainty.
stakeholders.
B. Decision Trees V
X |Lv |
IG(L, f ) = Entropy(L) − Entropy(Lv ) (5)
Decision trees are a kind of analytical model used in both |L|
v=1
machine learning and statistical analysis to aid in decision-
making and forecast results from given data. They outline the 3) GINI Index: GINI index determines the purity of a
potential outcomes of decisions, incorporating the probabilities specific class after splitting along a particular attribute. The
of different events, the costs associated with resources, and best split increases the purity of the sets resulting from the
their respective utilities. Structured similarly to a flowchart, split. If L is a dataset with j different class labels, GINI is
each internal node in a decision tree corresponds to an defined as
”evaluation” of a feature (for example, the result of flipping a j
coin), while each branch signifies the result of that evaluation, X
GINI(L) = 1 − p2i (6)
and every leaf node denotes a category label (the resolution
i=1
reached after considering all features). The routes from the
root to each leaf delineate the rules for classification or the Where pi is relative frequency if class iin L. If the dataset
sequences of decisions. A decision tree is a hierarchical model is split on attribute A into two subsets L1 and L2 with sizes
with internal nodes that segment instances based on input N1 and N2 respectively, GINI is calculated as
variables and leaves representing outcomes. In its construction, N1 N2
an algorithm seeks the optimal tree by minimizing a fitness GIN IA (L) = GIN I(L1 ) + GIN I(L2 ) (7)
N N
function that assesses split quality. For regression, it identifies Reduction in impurity is calculated as
split points that reduce the prediction error for the target
variable. ∆GIN I(A) = GIN I(L) − GIN IA (L) (8)
The fitness function is calculated concerning MSE, MAE, and Y as arguments. In our case, the independent variable is the
R2 square input variable VDC . and Pin , the dependent variable is the
output variable Pout .
1X
MSE = (Yi − Ȳt )2 (9) After training the model, additional steps may include data
n i∈n
preprocessing, such as handling missing values, normalizing
1X features, and splitting the dataset into training and testing
MAE = |Yi − Ȳt | (10)
n i∈n sets. It is essential to perform thorough data preprocessing
to ensure the accuracy and reliability of the model.After
(Yi − Ȳt )2
P
fitting or training the linear regression models, we analyze
R = 1 − Pi∈n
2
(11) their variation and behavior with different constant functions.
2
i∈n (Yi − Ỹ )
Initially, we experimented with constant values of 400 and
C. Neural Networks 500. However, these models did not exhibit satisfactory per-
Neural networks, inspired by the structure and function of formance compared to other models such as decision trees
the human brain, are a class of machine learning algorithms and neural networks. The linear regression models showed
capable of learning complex patterns and relationships from limitations in their saturation behavior, failing to accurately
data. Comprising interconnected nodes organized in layers, predict outcomes beyond certain ranges. This comparison
ANNs process input data through a series of mathematical highlights the importance of exploring alternative modeling
operations, known as activation functions, to generate output techniques and considering the limitations of linear regression
predictions. These models have gained prominence due to their in certain scenarios.
ability to tackle a wide range of tasks, including classification,
regression, and pattern recognition, across various domains
such as computer vision, natural language processing, and
predictive analytics.
Neural networks have found extensive application beyond
power grid control. In fields such as computer vision, natural
language processing, finance, and healthcare, neural networks
are utilized for various tasks such as image classification, senti-
ment analysis, financial forecasting, and medical diagnosis.The
versatility and adaptability of neural networks make them
a valuable tool in solving complex problems across diverse
domains.
The fundamental operation of a neural network involves
propagating input data forward through the network to produce
output predictions. Mathematically, this process can be repre-
sented using equations ‘(12)’,‘(13)’,‘(14)’,‘(15)’ that describe
the computations performed at each layer of the network.

z (1) = W (1) x + b(1) (12)

a(1) = σ(z (1) ) (13)
z (2) = W (2) a(1) + b(2) (14)
(2)
ŷ = σ(z ) (15)
• x is the input vector.
• W(i) and b(i) are the weights and biases of the i-th layer
• σ is the activation function
• ŷ is the predicted output vector
III. I MPLEMENTATION
A. Linear Regression
In the implementation phase, we first train the linear regres-
sion model on our dataset. This involves utilizing the ‘ftrlinear’
function, which is a built-in function in MATLAB designed Fig. 1. Decison Tree Algorithm
for training linear regression models. The ‘ftrlinear’ function
takes the independent variable X and the dependent variable
B. Decision Trees our models were effectively trained and capable of producing
accurate predictions.
In the implementation phase, we first train the decision tree
model using our dataset. MATLAB provides functions and IV. R ESULTS AND D ISCUSSION
tools to create, visualize, and make predictions with decision
In our comparative analysis of machine learning models
trees using the Statistics and Machine Learning Toolbox. Our
applied to voltage reference saturation, Linear Regression
problem involves continuous value handling so the fitrtree
exhibited under-performance, notably struggling to achieve
method is used. The ‘fitrtree’ function takes the independent
saturation at the targeted reference voltage. Conversely, the
variable X and the dependent variable Y as arguments. In our
Decision Tree model demonstrated rapid convergence, achiev-
case, the independent variable is the input variable VDC . and
ing saturation significantly ahead of its counterparts. Mean-
Pin , the dependent variable is the output variable Pout .
while, ANN displayed a balanced performance, reaching the
After training the model, additional steps may include data
saturation threshold within an intermediate timeframe as de-
preprocessing, such as handling missing values, normalizing
picted in Fig. 3. These results highlight the varying dynamics
features, and splitting the dataset into training and testing sets.
and efficiency of machine learning approaches in electrical
It is essential to perform thorough data preprocessing to ensure
parameter optimization.
the accuracy and reliability of the model.After implementing
The waveform of DC Link at a constant value of VDC =
graph is also derived with a (max − depth = 3).
450 V and VDC = 500 V are shown in Figs. 4 and 5 for the
After fitting or training the decision tree models, we analyze
Decision Tree model predictions, respectively. These figures
their variation and behavior with different constant functions.
clearly demonstrate that the prediction value of VDC saturates
Initially, we experimented with constant values of 400 and
at their respective reference voltage.
500. However, these models exhibit satisfactory performance
Fig 6 contains the plots of the quantities : Vg , Ig , Vbat ,
compared to other models such as linear regression and
Ibat and Vdc . Ibat and Vbat refer to the battery current and
neural networks. Unlike other algorithms, the Decision tree
voltage, respectively, in an electric vehicle. Analyzing the Ibat
achieves the (y = 0) value relatively fast after the overshoot
graph reveals that it stabilizes around -5 A approximately
and trough, hence stabilizing at desired output in less time.
0.2 sec after operation begins. This quick saturation time
Different performance metrics are also applied on decision
suggests efficient system functionality and accurate modeling,
tree including MSE , MAE and R2 .
as it indicates a rapid establishment of steady-state battery
current. Similarly, examining the Vbat graph shows saturation
C. Neural Network
occurring around 280 V , which is lower than the voltage
In the implementation phase, we utilize the ‘fitrnet’ function provided by the grid to the electric vehicle. This observation
to train the ANN regression model using our dataset. The ‘fitr- implies that the vehicle’s electrical system operates within
net’ function is a part of MATLAB’s Neural Network Toolbox expected parameters, efficiently utilizing the battery’s power
and is specifically designed for training neural networks for while maintaining a stable voltage level.
regression tasks. It employs a feed-forward neural network During the initial phase of a lithium-ion battery charging cy-
architecture and employs a back-propagation algorithm to cle ,the DC voltage is being maintained at a constant level(400
iteratively adjust the network weights to minimize the mean V ) hence acquiring a DC profile , which is a characteristic of
squared error between the predicted and actual outputs. the constant current (CC) charging phase.This phase is crucial
The ‘fitrnet’ function takes the independent variables X and because it allows for a rapid and efficient intake of energy
the dependent variable Y as input arguments, along with op- without exceeding the battery’s chemical and physical limits.
tional parameters specifying the network architecture, training The State of Charge (SOC) of a battery is a measure of the
algorithm, and regularization techniques. The user can specify current amount of energy stored in the battery compared to its
the number of hidden layers, the number of neurons in each capacity when fully charged, expressed as a percentage. The
layer, activation functions, and regularization parameters to SOC gives an indication of how much longer the battery can
customize the network architecture and optimize performance. continue to provide power under the current conditions.it can
After training the neural network models, we conducted an observed that the current capacity of the battery is increasing
analysis of the behavior of power grid elements such as VDC and hence it is charging.
and Ig across various constant reference values. The results In our analysis, we focus on two crucial parameters: Vg ,
indicate that the neural network models performed better than representing the grid voltage supplied from the grid to EV
linear regression in predicting the behavior of these elements. battery, and Ig , denoting the grid current. These values play
However, the performance of neural networks was not as ro- a pivotal role in our assessment and provide valuable insights
bust as decision trees. Notably, the saturation values achieved into system performance.
by the neural network models closely matched the desired Upon examination of the graph, it is evident that both
target values. For instance, at a constant reference value of Vg and Ig exhibit sinusoidal behavior, aligning closely with
400 V , the neural network accurately predicted saturation our anticipated expectations. The Vg signal demonstrates a
values of 400 V , and similarly, at 500 V , it yielded precise consistent and relatively stable sinusoidal pattern, indicative
saturation values of 500 V . This consistency demonstrates that of robust grid voltage supply. This reliability is fundamental
 Fig. 2. Graph View of the Decision Tree

Fig. 5. Waveform of DC link for a Reference of 500 V

for ensuring the proper functioning and stability of the system

While the Ig signal also follows a sinusoidal trend, notable
Fig. 3. Plot of VDC vs Time for ANN,LR,DT observations include its smaller peak and narrower range com-
pared to Vg . Despite these variations, the overall performance
remains satisfactory, albeit with slight fluctuations.

TABLE I
T RAINING P ERFORMANCE

Model Number Model Type MSE RSquared MAE

1 Tree 0.0039 0.999 0.03
2 Neural Network 0.076 0.997 0.16
3 Linear Regression 0.84 0.97 0.55

TABLE II
T ESTING P ERFORMANCE

Model Number Model Type MSE RSquared MAE

1 Tree 0.003 0.999 0.0379
2 Neural Network 0.171 0.994 0.29
3 Linear Regression 0.844 0.974 0.557

Table 1 and Table 2 shows the expression of the prediction

accuracies of the ML models on the training and testing
Fig. 4. Waveform of DC link for a Reference of 450 V
data,respectively. These results show that the developed ML-
based Decision Tree model exhibit high accuracy compared to
others.
 Fig. 6. Plot of Various quantities for Decision Tree Algorithm

V. C ONCLUSION Electronics, Smart Grid, and Renewable Energy (PESGRE), Trivandrum,

India, 2023, pp. 1-6, doi: 10.1109/PESGRE58662.2023.10404443.
In our comparative analysis, the results achieved by the [4] G. Yadav and M. Singh, ”Robust Control Design for Grid-Tied EV Sys-
Decision Tree Regressor have notably surpassed those attained tem with SOGI-Based Architecture in Wide Voltage Range Scenarios,”
2023 International Conference on Electrical, Electronics, Communica-
by both Linear Regression and ANNs within the regression tion and Computers (ELEXCOM), Roorkee, India, 2023, pp. 1-6, doi:
scenario. The Decision Tree Regressor demonstrated superior 10.1109/ELEXCOM58812.2023.10370088.
performance, particularly in its ability to capture non-linear [5] G. Yadav and M. Singh, ”Unveiling the Superiority: Comparative Analy-
sis of ANFIS, FOPID, and PI Controllers in Grid-Connected EV Systems
relationships and complex patterns in the data, which the for G2V/V2G Applications,” 2023 International Conference on Electri-
Linear Regression model struggles with due to its inherent cal, Electronics, Communication and Computers (ELEXCOM), Roorkee,
linearity. Moreover, while ANNs are typically strong per- India, 2023, pp. 1-6, doi: 10.1109/ELEXCOM58812.2023.10370196.
[6] Suryakanthi Tangirala, “Evaluating the Impact of GINI Index and
formers in capturing intricate data relationships, they require Information Gain on Classification using Decision Tree Classifier Al-
extensive data and hyperparameter tuning to reach optimal gorithm*” International Journal of Advanced Computer Science and
performance. In this case, the simpler, more interpretable Applications(IJACSA), 11(2), 2020.
[7] T. W. Gyeeda, A. J. H. Simons and M. Stannett, ”Regression Analysis of
structure of the Decision Tree led to quicker adjustments and Predictions and Forecasts of Cloud Data Center KPIs Using the Boosted
a more straightforward understanding of the data’s underlying Decision Tree Algorithm,” in IEEE Transactions on Big Data, vol. 9,
patterns, resulting in more accurate predictions. This outcome no. 4, pp. 1071-1085, 1 Aug. 2023
[8] A. Sharma and N. Chaudhary, ”Linear Regression Model for Agile
underscores the importance of choosing the right model based Software Development Effort Estimation,” 2020 5th IEEE Interna-
on the specific characteristics of the dataset and the problem at tional Conference on Recent Advances and Innovations in Engineering
hand, rather than defaulting to more complex models assuming (ICRAIE), Jaipur, India, 2020.
[9] T. Guillod, P. Papamanolis and J. W. Kolar, ”Artificial Neural Network
they will perform better. (ANN) Based Fast and Accurate Inductor Modeling and Design,” in
IEEE Open Journal of Power Electronics, vol. 1, pp. 284-299, 2020
R EFERENCES [10] .S. Kim et al., ”Integration of Neural Network-Based Symbolic Regres-
sion in Deep Learning for Scientific Discovery,” in IEEE Transactions on
[1] M. Ansari, G. Yadav and M. Singh, ”A 3.3kW Modified LLC Resonant Neural Networks and Learning Systems, vol. 32, no. 9, pp. 4166-4177,
Converter for Grid-Tied EV System Under Wide Voltage Range,” 2023 Sept. 2021
3rd International Conference on Intelligent Technologies (CONIT),
Hubli, India, 2023, pp. 1-6, doi: 10.1109/CONIT59222.2023.10205690.
[2] M. Ansari, G. Yadav and M. Singh, ”Modified Bidirectional LLC
Converter for Electric Vehicle Application,” 2023 Second International
Conference on Electrical, Electronics, Information and Communication
Technologies (ICEEICT), Trichirappalli, India, 2023, pp. 1-7, doi:
10.1109/ICEEICT56924.2023.10157520.
[3] Akshay, G. Yadav and M. Singh, ”Validation of a Parameter Design
methodology for Air-core two-coil WPT system with multi-resonant
LC compensation,” 2023 IEEE International Conference on Power