Article

Electric Vehicle Charging Load Forecasting Method Based on

Improved Long Short-Term Memory Model with Particle
Swarm Optimization
Xiaomeng Yang 1 , Lidong Zhang 1, * and Xiangyun Han 2

1 School of Transportation and Logistics Engineering, Shandong Jiaotong University, Jinan 250023, China
2 Department of Traffic Management and Engineering, Chongqing Police College, Chongqing 401331, China;
[email protected]
* Correspondence: [email protected]; Tel.: +86-13791038708

Abstract: With the rapid global proliferation of electric vehicles (EVs), their integration
as a significant load component within power systems increasingly influences the stable
operation and planning of electrical grids. However, the high uncertainty and randomness
inherent in EV users’ charging behaviors render accurate load forecasting a challenging
task. In this context, the present study proposes a Particle Swarm Optimization (PSO)-
enhanced Long Short-Term Memory (LSTM) network forecasting model. By combining the
global search capability of the PSO algorithm with the advantages of LSTM networks in
time-series modeling, a PSO-LSTM hybrid framework optimized for seasonal variations is
developed. The results confirm that the PSO-LSTM model effectively captures seasonal
load variations, providing a high-precision, adaptive solution for dynamic grid scheduling
and charging infrastructure planning. This model supports the optimization of power
resource allocation and the enhancement of energy storage efficiency. Specifically, during
winter, the Mean Absolute Error (MAE) is 3.896, a reduction of 6.57% compared to the
LSTM model and 10.13% compared to the Gated Recurrent Unit (GRU) model. During the
winter–spring transition, the MAE is 3.806, which is 6.03% lower than that of the LSTM
Academic Editor: Grzegorz Sierpiński model and 12.81% lower than that of the GRU model. In the spring, the MAE is 3.910,
Received: 11 February 2025
showing a 2.71% improvement over the LSTM model and a 7.32% reduction compared to
Revised: 26 February 2025 the GRU model.
Accepted: 3 March 2025
Published: 5 March 2025 Keywords: electric vehicles (EVs); load forecasting; long short-term memory network;
Citation: Yang, X.; Zhang, L.; Han, X. particle swarm optimization; deep learning
Electric Vehicle Charging Load
Forecasting Method Based on
Improved Long Short-Term Memory
Model with Particle Swarm 1. Introduction
Optimization. World Electr. Veh. J.
2025, 16, 150. https://doi.org/
Amid the escalating global energy crisis and increasing environmental pollution
10.3390/wevj16030150 concerns, electric vehicles (EVs) have rapidly emerged as a sustainable transportation
solution. However, the widespread adoption of EVs has posed new challenges for power
Copyright: © 2025 by the authors.
Published by MDPI on behalf of the
grids. The charging load of EVs is inherently random and volatile, potentially disrupting
World Electric Vehicle Association. the stable operation of power grids. Therefore, accurately forecasting fluctuations in EV
Licensee MDPI, Basel, Switzerland. load [1] is essential for effective grid planning and operation.
This article is an open access article Traditional forecasting techniques encompass various methods, including Monte
distributed under the terms and Carlo simulation and Kalman filtering. Reference [2] developed a Monte Carlo-based
conditions of the Creative Commons
model that incorporates EV types, using the predicted number of EVs in a region as a
Attribution (CC BY) license
(https://creativecommons.org/
basis while considering other factors as model parameters. Reference [3] proposed a
licenses/by/4.0/). hybrid algorithm combining time-series analysis and Kalman filtering, which improved

World Electr. Veh. J. 2025, 16, 150 https://doi.org/10.3390/wevj16030150

World Electr. Veh. J. 2025, 16, 150 2 of 19

accuracy and facilitated the derivation of state and observation equations, effectively
underscoring the benefits of hybrid algorithms in improving short-term load forecasting
for power systems.
Electric load forecasting is inherently nonlinear and influenced by multiple factors.
Machine learning, with its robust nonlinear mapping capabilities, has demonstrated efficacy
in addressing nonlinear problems in load forecasting. Traditional machine learning tech-
niques, such as support vector machines, decision trees, and random forests, are generally
suited for smaller datasets and are effective in addressing nonlinear problems. Reference [4]
addressed issues of low accuracy and inadequate consideration of seasonality in traditional
EV load forecasting by proposing a seasonal EV charging load prediction model based on
random forests. Reference [5] introduced an adaptive improvement method for Particle
Swarm Optimization to solve high-dimensional EV charging load prediction models.
Deep learning-based forecasting methods utilize neural networks as parameter struc-
tures for optimization. Neural networks, also known as artificial neural networks, are
mathematical models inspired by the way biological neural networks process and trans-
mit information. Commonly used neural networks for power load forecasting include
backpropagation (BP) networks, convolutional neural networks (CNNs), recurrent neural
networks (RNNs), and emerging Transformer models. For example, Reference [6] proposed
a spatiotemporal graph convolutional network (GCN+LSTM) that integrates graph convo-
lutional networks (GCNs) with Long Short-Term Memory (LSTM) networks to improve
the accuracy of electric vehicle (EV) charging demand predictions and alleviate traffic
congestion in high-demand areas. Reference [7] integrated Kolmogorov-Arnold Networks
(KANs) into traditional machine learning frameworks, specifically Convolutional Neural
Networks (CNNs). Reference [8] developed a short-term forecasting method based on a
bidirectional LSTM (BiLSTM) neural network optimized by a Sparrow Search Algorithm
(SSA) and variational mode decomposition (VMD).
Moreover, the accuracy of EV charging load forecasting is significantly influenced by
external factors, among which weather conditions play an essential role. Weather variables,
such as temperature, humidity, wind speed, and solar radiation, have a significant impact
on EV charging demand. Therefore, incorporating weather factors is crucial for enhancing
the accuracy and reliability of forecasting models. Reference [9] proposed an EV charging
load forecasting method that considers multiple influencing factors, including weather.
Reference [10] explored the effects of meteorological conditions on the spatiotemporal
distribution of EV charging loads at highway service areas, presenting a weather-inclusive
forecasting model for these areas. Reference [11] proposed a hybrid forecasting frame-
work that integrates Partial Least Squares Regression (PLSR) with a lightweight Gradient
Boosting Machine (LightGBM), combined with Bayesian hyperparameter optimization and
a second-order cone optimal scheduling scheme. This approach significantly improves
short-term load forecasting accuracy, with the Mean Absolute Error (MAE) reduced by
12.7%. Reference [12] addressed the high randomness and spatial heterogeneity of elec-
tric vehicle charging load by proposing a machine learning forecasting framework based
on a dual perspective of industrial parks and charging stations. By combining the MLP
and LSTM algorithms, the framework achieves high-precision forecasting for weekdays
(LSTM R2 = 0.9283), holidays (R2 = 0.9154), and weekends (MLP R2 = 0.9586).
As the world’s largest automobile market and leading producer of electric vehicles
(EVs), China is experiencing rapid growth in the number of EVs in use. In China, there are
significant regional differences in electric vehicle usage patterns and charging demands.
Taking first-tier cities such as Beijing, Shanghai, and Shenzhen as examples, due to their
dense populations and developed public transportation systems, electric vehicles are
primarily used for urban commuting, and charging demand is concentrated during the
World Electr. Veh. J. 2025, 16, 150 3 of 19

morning and evening peak hours on weekdays, as well as at residential charging stations
at night. In some second- and third-tier cities, the use of electric vehicles is more diverse,
including operational vehicles such as taxis and ride-hailing services, resulting in more
scattered and uncertain charging demand.
In order to accurately predict the electric vehicle charging load for rational grid capac-
ity planning, optimized charging facility layout, and maintaining power system stability,
this study proposes a PSO-LSTM model that combines Particle Swarm Optimization (PSO)
with the LSTM network to improve the accuracy of EV charging load prediction. Com-
parative experimental case studies between the PSO-LSTM model and traditional models
demonstrate that the PSO-LSTM model significantly outperforms others in prediction accu-
racy, providing effective data support for grid and power system operations. This research
not only validates the correctness and effectiveness of the PSO-LSTM model but also offers
new perspectives and solutions for the field of electric vehicle charging load forecasting.

2. Model Description
2.1. Parameters and Data
To ensure the scientific rigor and reproducibility of this study, this research systemati-
cally set and rigorously validated key parameters during the construction and optimization
of the electric vehicle charging load prediction model. These parameters cover core aspects
such as data preprocessing, model architecture design, and optimization algorithm config-
uration. Table 1 provides a complete list of the definitions and settings of all parameters,
allowing readers to fully understand the implementation details and key design logic of
this study.

Table 1. Key parameter definitions in the study.

Parameter Definition Unit

D The search space dimension of the PSO algorithm
N The number of particles in the PSO algorithm
ω The inertia weight of the PSO algorithm
num The number of iterations for the PSO algorithm
c1 The acceleration coefficient for personal best position in PSO
c2 The acceleration coefficient for the global best position in PSO
r1 , r2 Random numbers within the range [0, 1]
P Total energy consumption of EV charging stations kWh
Hidden Units1 The number of neurons in the first hidden layer
Hidden Units2 The number of neurons in the second hidden layer
Dropout The dropout rate
Batch Size Batch size
Epochs The number of training iterations

To ensure the academic rigor of the expressions and consistency in reader compre-
hension, this study provides systematic definitions and standardized explanations of the
technical term abbreviations used. Table 2 provides a complete list of all abbreviations
used in the paper, along with their corresponding full forms, definitions, and application
contexts, in order to eliminate potential ambiguities.
World Electr. Veh. J. 2025, 16, 150 4 of 19

Table 2. Abbreviations and their full forms cited in the study.

Abbreviation Full Form

PSO Particle Swarm Optimization
LSTM Long Short-Term Memory
PSO-LSTM Particle Swarm Optimization Long Short-Term Memory
GRU Gated Recurrent Unit
EV Electric vehicle
BP Backpropagation
CNN Convolutional neural network
KAN Kolmogorov-Arnold network
RNN Recurrent neural network
GCN Graph convolutional network
GNN Graph neural networks
BiLSTM Bidirectional LSTM
SSA Sparrow Search Algorithm
VMD Variational mode decomposition
MAE Mean Absolute Error
MSE Mean Squared Error
RMSE Root Mean Squared Error

Traffic flow studies have shown that travelers’ travel patterns exhibit distinct regular-
ities, which, in turn, determine the patterns of urban road traffic conditions. As a result,
the traffic flow fluctuations generated by travelers on the same road segment also follow
a regular pattern, including periodic trends with yearly, quarterly, monthly, weekly, and
daily cycles. Collecting data with these time spans will form long-term, medium-term, and
short-term traffic flow patterns with strong regularity and high similarity in the fluctuation
curves. Based on the basic characteristics of urban road traffic flow, it is known that traffic
flow data are time-series data, making time similarity analysis essential.
The modeling data used in this study were sourced from a single charging pile. The
dataset includes the start and end times for each charging event, as well as the total energy
consumed [13]. The raw dataset was converted into a corresponding dataset containing the
hourly average charging load in kW—P(kWh). The data selected span from 1 January 2023
to 29 April 2023, with a prediction time period of one week, with an interval of one hour,
totaling 2856 data points.
To thoroughly analyze the seasonal variation characteristics of the charging load, this
study performed initial data preprocessing, including missing value imputation, outlier
removal, and normalization, to ensure data completeness and consistency. Subsequently, the
data were divided into the following three periods based on common seasonal classifications:
• Winter Period (1 January–3 February): During this period, low temperatures may lead
to reduced battery performance and changes in user charging behavior, potentially
affecting the charging load.
• Winter–Spring Transition Period (4 February–4 March): As temperatures gradually
rise, the charging load may exhibit transitional characteristics.
• Spring Period (5 March–29 April): In this period, temperatures are moderate, and the
charging load is likely to stabilize, reflecting typical spring user behavior patterns.
To assess the predictive performance of the model, the data for each period were split
into training and testing sets in an 80:20 ratio. The training set was used for model training
and hyperparameter optimization, while the testing set was employed to evaluate the
final prediction accuracy and generalization capability of the model. This study, through
seasonal partitioning and visualization analysis, comprehensively captured the seasonal
variation patterns of the charging load. By training the model independently for each
season, the influence of seasonal factors on the model was effectively reduced, allowing the
 variation patterns of the charging load. By training the model independently for each sea-
World Electr. Veh. J. 2025, 16, 150 variation patterns of the charging load. By training the model independently for 5each of 19 sea-
son, the influence of seasonal factors on the model was effectively reduced, allowing the
son, the influence of seasonal factors on the model was effectively reduced, allowing the
seasonal characteristics to be more accurately modeled, thereby enhancing the accuracy
seasonal characteristics to be more accurately modeled, thereby enhancing the accuracy
ofseasonal
the prediction results.to be more accurately modeled, thereby enhancing the accuracy of
characteristics
of the prediction results.
Figures 1–3
the prediction present the three-dimensional visualization results of the charging load
results.
Figures 1–3 present the three-dimensional visualization results of the charging load
data, Figures
segmented by season.
1–3 present In the figures, thevisualization
the three-dimensional x-axis represents 24ofhthe
results of charging
a day, theload
y-axis
data, segmented by season. In the figures, the x-axis represents 24 h of a day, the y-axis
represents the days,
data, segmented and theInz-axis
by season. represents
the figures, the charging
the x-axis representspower (ina kWh).
24 h of day, theThe three-
y-axis
represents the days, and the z-axis represents the charging power (in kWh). The three-
dimensional
represents theplots provide
days, and theanz-axis
intuitive observation
represents of the power
the charging spatial–temporal
(in kWh). The distribution
three-
dimensional plots provide an intuitive observation of the spatial–temporal distribution
dimensional plots provide an intuitive observation of
characteristics of the charging load across different seasons. the spatial–temporal distribution
characteristics of the charging load across different seasons.
characteristics of the charging load across different seasons.

Figure
Figure1.1.The three-dimensional spatial–temporaldistribution
distributionof of charging load in the winter season.
Figure 1. The
The three-dimensional
three-dimensionalspatial–temporal
spatial–temporal distribution charging load
of charging in the
load in winter season.
the winter season.

Figure 2. The three-dimensional spatial–temporal distribution of charging load during the winter–
Figure 2. The three-dimensional spatial–temporal distribution of charging load during the winter–
spring transition
Figure period.
2. The three-dimensional spatial–temporal distribution of charging load during the winter–
spring transition period.
spring transition period.
WorldWorld
Electr. Veh.
Electr. J. 2025,
Veh. J. 2025,16,
16, xx FOR PEER
FOR PEER REVIEW
REVIEW 6 of 19
World Electr. Veh. J. 2025, 16, 150 6 of 19

Figure 3. The three-dimensional spatial–temporal distribution of charging load in the spring season.

2.2. Improved LSTM Prediction Model

Figure The three-dimensional
The3.Long Short-Term Memoryspatial–temporal distribution
(LSTM) network of charging
is a distinct load in
variant ofthe spring season.
recurrent neu-
Figure 3. The three-dimensional spatial–temporal distribution of charging load in the spring
ral networks (RNNs) specifically designed to address the vanishing and exploding gradi-
2.2. Improved LSTM Prediction Model
ent problems typically encountered by conventional RNNs when processing long se-
2.2. Improved LSTM Prediction
The Long Short-Term Memory (LSTM) Model network is a distinct variant of recurrent neural
quences of data. The core innovation of LSTM lies in its unique gating mechanism, which
networks (RNNs) specifically designed to address the vanishing and exploding gradient
allows The Long Short-Term
the network to dynamically Memory
regulate(LSTM) the flow of network is a distinct
information. varianten-
This mechanism of recurre
problems typically encountered by conventional RNNs when processing long sequences
ables
ral the retention
networks and
(RNNs) transmission
specifically of critical
designed temporal
to address dependencies
the vanishing over long
and se-
exploding
of data. The core innovation of LSTM lies in its unique gating mechanism, which allows
quences [14].
ent problems
the network typically regulate
to dynamically encountered the flowby conventional
of information. This RNNsmechanismwhen processing
enables the lo
The core structure of an LSTM unit is illustrated in Figure 4. It consists of three gating
quences of data.
retention and The core
transmission innovation
of critical temporal of dependencies
LSTM lies inover
mechanisms—the forget gate, input gate, and output gate—along with a cell state update
its long
unique gating[14].
sequences mechanism,
allows Thethe
corenetwork
structuretoof an LSTM unit regulate is illustrated in Figure 4. It consists of three
module. The forget gate ( f tdynamically
), governed by a sigmoidthe flow
function, of information.
determines which Thisinfor-mechani
gating mechanisms—the forget gate, input gate, and output gate—along with a cell state
ables tothe retentiontheand transmission of critical temporal dependencies over lo
mation discard from
update module. The forgetcell state
gate ( f t(),Cgoverned
t −1 ). The input gate ( it ) and
by a sigmoid the candidate
function, determines cellwhich
state
quences [14].
 ) collaboratively t−1 ).( C
t ), while the output gate ( ot ) regu-
(Cinformation to discard fromthe
update the cell state
current cell (C
state The input gate (it ) and the candidate
t
cell state (Ct ) collaboratively update the current cell state (Cin
The core structure of an LSTM unit is illustrated t ), Figure
while the4. Itoutput
consists
gateof three
lates the output of the hidden state ( ht ). The arrows indicate the direction of information
e
mechanisms—the
(ot ) regulates the output forget
of thegate,
hidden input gate,
state (ht ).andTheoutput
arrowsgate—along with a cell
indicate the direction of state
flow, while the Tanh function serves as the activation
information flow, while the Tanhf function serves as the activation function. function.
module. The forget gate ( t ), governed by a sigmoid function, determines which
mation to discard from the cell state ( Ct −1 ). The input gate ( it ) and the candidate ce
( C t ) collaboratively update the current cell state ( Ct ), while the output gate ( ot
lates the output of the hidden state ( ht ). The arrows indicate the direction of infor
flow, while the Tanh function serves as the activation function.

Figure4.4.AAstructural
Figure structuraldiagram
diagramofofthe
theLSTM
LSTMnetwork.
network.
World Electr. Veh. J. 2025, 16, 150 7 of 19

2.2.1. Basic Structure of LSTM

The LSTM unit consists of three gates: the forget gate, the input gate, and the output
gate. Each gate operates as a neural network layer, collaborating to determine which
information should be retained, updated, or discarded.
1. Forget Gate:
It determines which information should be discarded from the cell state.
 
f t = σ W f × [ h t −1 , x t ] + b f (1)

where f t represents the output of the forget gate, σ is the sigmoid function, W f denotes the
weight matrix, ht−1 is the hidden state from the previous time step, xt is the current input,
and b f represents the bias term.
2. Input Gate:
It controls the storage of new information.

it = σ (Wi · [ht−1 , xt ] + bi ) (2)

et = tanh(WC · [ht−1 , xt ] + bC )
C (3)

Ct = f t · Ct−1 + it · C
et (4)

where it represents the output of the input gate, C

et is the new candidate value vector, and
Ct is the current cell state.
3. Output Gate:
It determines the output of the hidden state.

ot = σ (Wo · [ht−1 , xt ] + bo ) (5)

ht = ot · tanh(Ct ) (6)

where ot represents the output of the output gate, and ht is the hidden state at the current
time step.
The information propagation process in an LSTM neural network is described as
follows [15]:
• Forgetting and Memory: The input information and stored information are multiplied
by weight matrices, and after adding the bias term, they pass through a sigmoid
function for normalization to obtain the final input information.
• New Information Input: During the input phase, the data are processed by passing
them through the weight matrix and multiplying them with the activation matrix,
producing the information that will be transferred to the memory unit.
• Cell State Update and Information Output: The results of the first two steps are
combined to compute the current cell state. This cell state is then multiplied by the
output matrix to generate the final output.

2.2.2. Improvements to the LSTM Model

To enhance the model’s ability to capture patterns in time-series data, two key improve-
ments have been made to the traditional LSTM model: the introduction of bidirectional
LSTM and the stacking of multiple LSTM layers.
The standard LSTM model can only learn the features of the current time step from
past time steps. In contrast, the bidirectional LSTM processes sequence data in both
World Electr. Veh. J. 2025, 16, x FOR PEER REVIEW 8 of 19

World Electr. Veh. J. 2025, 16, 150

The standard LSTM model can only learn the features of the current time step 8from of 19

past time steps. In contrast, the bidirectional LSTM processes sequence data in both for-
ward and backward directions simultaneously, enabling it to capture comprehensive con-
forward and backward directions simultaneously, enabling it to capture comprehensive
textual information from both past and future time steps [16]. This design is particularly
contextual information from both past and future time steps [16]. This design is particularly
suitable for scenarios where there is a strong correlation between future and past time
suitable for scenarios where there is a strong correlation between future and past time
steps, such as in periodic load data. By integrating information from both time directions,
steps, such as in periodic load data. By integrating information from both time directions,
bidirectional LSTM significantly improves the model’s ability to understand the latent
bidirectional LSTM significantly improves the model’s ability to understand the latent
patterns in the sequence [17]. Stacking multiple LSTM layers allows the model to gradu-
patterns in the sequence [17]. Stacking multiple LSTM layers allows the model to gradually
ally extract higher-dimensional patterns, from shallow features to deeper ones.
extract higher-dimensional patterns, from shallow features to deeper ones.
This stacking approach mimics the structure of deep neural networks, progressively
This stacking approach mimics the structure of deep neural networks, progressively
extracting more abstract features. During implementation, to ensure smooth information
extracting more abstract features. During implementation, to ensure smooth information
transmission across multiple layers, the first LSTM layer is set with “return_sequences =
transmission across multiple layers, the first LSTM layer is set with “return_sequences = True”
True” to output the complete time series for the next layer. The final LSTM layer is con-
to output the complete time series for the next layer. The final LSTM layer is configured
figured with “return_sequences = False” to produce a fixed-length vector as output, which
with “return_sequences = False” to produce a fixed-length vector as output, which is then
is then connected to a dense layer for subsequent stages of prediction tasks.
connected to a dense layer for subsequent stages of prediction tasks.
The architecture of the improved LSTM model is shown in Figure 5. In this design,
The architecture of the improved LSTM model is shown in Figure 5. In this design, the
the bidirectional LSTM processes both forward and backward sequence data simultane-
bidirectional LSTM processes both forward and backward sequence data simultaneously,
ously, enabling the model to capture global contextual information between past and fu-
enabling the model to capture global contextual information between past and future time
ture time steps. The stacking of LSTM layers is as follows: The first layer consists of a
steps. The stacking of LSTM layers is as follows: The first layer consists of a bidirectional
bidirectional LSTM layer, with a dropout layer inserted in between to randomly deacti-
LSTM layer, with a dropout layer inserted in between to randomly deactivate neurons
vate neurons and reduce overfitting. The final LSTM layer outputs a fixed-length vector,
and reduce overfitting. The final LSTM layer outputs a fixed-length vector, which is then
which is then connected to a fully connected layer for the final load prediction. The input
connected to a fully connected layer for the final load prediction. The input layer receives
layer receives sequential data, with “x0, x1, x2, ..., xi” representing the sequence elements.
sequential data, with “x0, x1 , x2 , . . ., xi ” representing the sequence elements. “A” and “A‘“
“A” and “A‘“ represent LSTM units processing in opposite directions, with red arrows
represent LSTM units processing in opposite directions, with red arrows indicating forward
indicating forward time steps and blue arrows denoting backward time steps. The se-
time steps and blue arrows denoting backward time steps. The sequence “h0 , h1 , h2 , . . ., hi ”
quence “h0, h1, h2, ..., hi” represents the output hidden states.
represents the output hidden states.

Figure 5.
Figure A diagram
5. A diagram of
of the
the improved
improved model
model structure.
structure.
World Electr. Veh. J. 2025, 16, 150 9 of 19

2.3. LSTM Core Parameter Optimization Based on PSO Algorithm

The Particle Swarm Optimization (PSO) algorithm, introduced by Dr. Eberhart and
Dr. Kennedy in 1995 [18], is a stochastic search method based on collective cooperation.
It is inspired by the foraging behaviors of bird flocks. The algorithm’s initial operational
framework was designed as a simplified model of social behavior, particularly how birds
forage in groups.
Researchers have found that individuals within a bird flock can interact and influence
one another, a phenomenon that exists in biological groups as a mechanism for information
sharing. The PSO algorithm leverages this information-sharing mechanism, allowing
both individual particles and the entire swarm to synchronize their knowledge [19]. In
the PSO algorithm, the term “particle” is used as an intermediary selection because the
members of the group exist in actual states rather than abstract descriptions, such as mass,
volume, speed, and acceleration. Therefore, particles are used as replacements. Particles
communicate and exchange information, collectively forming a swarm.
The objective of the PSO algorithm is to guide multiple particles in finding the optimal
solution within a multi-dimensional hypervolume, also known as the solution space, which
serves as the search space for the optimization problem. In this framework, the optimal
solution for each problem is represented as a particle within this space, and all particles are
mapped into a D-dimensional space. Each particle has a fitness value, determined by an
optimization function, which assesses the quality of its current position. Additionally, each
particle is associated with a velocity vector that governs its movement and position. The
particle swarm then searches for the current optimal particle within the solution space.
In the Particle Swarm Optimization (PSO) algorithm, each unit is referred to as a
particle. A swarm consists of N particles, which are randomly initialized within a D-
dimensional search space [20]. In the search process, each particle i is represented by two
vectors: the velocity vector Vi and the position vector Xi . Each particle i updates its velocity
and position using its personal best position Pbesti and the global best position Gbesti
found so far. The update equations for Xi and Vi are as follows:

Vidt+1 = ω ∗ Vidt + c1 · r1 · Pbestid

t t
+ c2 · r2 · Gbesttd − Xid
t
   
− Xid (7)

t +1 t
Xid = Xid + Vidt+1 (8)
t and V t represent the velocity and position of particle i in dimension d during
Here, Xid id
iteration t, respectively. Pbestidt is the best position found by particle i in dimension d up

to iteration t, while Gbesttd is the best position found by the entire swarm in dimension d
during iteration t. t and t + 1 represent the current and next iterations, respectively [21]. ω
denotes the inertia weight, and c1 and c2 are the acceleration coefficients for the cognitive
and social components, respectively. r1 and r2 are random numbers uniformly distributed
in the range [0, 1]. The personal best position and the global best position of each particle
are updated in each iteration using the following equations:
  
 X t +1 t +1 t


t +1 id , f Xid ≤ f Pbestid
Pbestid =   (9)
 Pbestt t +1 t


id , f Xid > f Pbestid

   
 Gbesttd t +1
, f Gbesttd ≤ min f Pbestid


Gbestdt+1 = i    (10)
 Pbestt+1 t +1
, f Gbesttd > min f Pbestid


id i

Figure 6 illustrates the concept of particle optimization, depicting the Particle Swarm
Optimization (PSO) algorithm. The particle velocity update equation in the PSO algorithm
 Gbestd =  (10)

t +1
 Pbestid
i
(
, f ( Gbestdt ) > min f ( Pbestidt +1 ) )
World Electr. Veh. J. 2025, 16, 150 Figure 6 illustrates the concept of particle optimization, depicting the Particle Swarm 10 of 19
Optimization (PSO) algorithm. The particle velocity update equation in the PSO algorithm
consists of three components. The first component represents the particle’s inertia, reflect-
ingconsists of three components.
its “memory” Thevelocity,
of the previous first component represents the
and is controlled by particle’s
the inertiainertia,
weight ( ω ).
reflecting
its “memory” of the previous velocity, and is controlled by the inertia
The second component captures the particle’s self-awareness, driving it to move toward weight (ω). The
t t +1
itssecond component
own historical bestcaptures
position the particle’s
( Pbest self-awareness,
), which driving it to move
reflects “self-awareness.” Here,toward
“ X ”its
own historical best position (Pbestt ), whicht +1 reflects “self-awareness”. Here, “X t+1 ” denotes
denotes the adjusted position, tand “ V ” represents the velocity increment. The third
the adjusted position, and “V +1 ” represents the velocity increment. The third component
component represents the exchange of information and cooperation between particles,
represents the exchange of information and cooperation between particles, referred to
referred to as “socialization” [22], guiding particles toward the global best position (
as “socialization”
t [22], guiding particles toward the global best position (Gbestt ) and
Gbest ) and group
facilitating facilitating group information
information sharing. sharing.

Figure
Figure 6. Diagram
6. Diagram of of particle
particle global
global andand historical
historical optimal
optimal solutions,
solutions, velocity,
velocity, and
and position.
position.

In In applied
applied research,the
research, theeffectiveness
effectivenessofofaamodel
modelisisprimarily
primarilydetermined
determined byby the
the care-
careful
selection of key hyperparameters. However, traditional manual tuning
ful selection of key hyperparameters. However, traditional manual tuning methods are methods are not
only
not onlytime-consuming
time-consumingand andlabor-intensive
labor-intensivebut but also prone
also pronetoto
getting stuck
getting stuckinin
local optima,
local op-
preventing the model from fully realizing its potential. To address this issue,
tima, preventing the model from fully realizing its potential. To address this issue, this this study
uses the Particle Swarm Optimization (PSO) algorithm to optimize five key parameters
study uses the Particle Swarm Optimization (PSO) algorithm to optimize five key param-
in the LSTM model: the number of neurons in the first hidden layer (Hidden Units1), the
eters in the LSTM model: the number of neurons in the first hidden layer (Hidden Units1),
number of neurons in the second hidden layer (Hidden Units2), the dropout rate (Dropout),
the number of neurons in the second hidden layer (Hidden Units2), the dropout rate
the batch size (Batch Size), and the number of training epochs (Epochs). Through the global
(Dropout), the batch size (Batch Size), and the number of training epochs (Epochs).
search capabilities and efficient parameter adjustment strategy of the PSO algorithm, this
Through the global search capabilities and efficient parameter adjustment strategy of the
study aims to identify the optimal combination of parameters to maximize the predictive
PSO algorithm, this study aims to identify the optimal combination of parameters to max-
performance of the LSTM model while minimizing resource consumption during the
imize the predictive performance of the LSTM model while minimizing resource con-
training process.
sumption during the training process.
2.4. Development of the Electric Vehicle Load Forecasting Model
2.4. Development of the Electric Vehicle Load Forecasting Model
Identifying the optimal parameters for a predictive model is inherently challenging,
Identifying the optimal parameters for a predictive model is inherently challenging,
and the training process is both time-consuming and computationally intensive. The
and the training process is both time-consuming and computationally intensive. The op-
optimal values obtained through manual tuning are often only locally optimal rather than
timal values obtained through manual tuning are often only locally optimal rather than
globally optimal. To mitigate the errors and randomness introduced by manual tuning
globally optimal. To mitigate the errors and randomness introduced by manual tuning
and to achieve automatic parameter optimization, this study employs a swarm intelligence
and to achieve automatic parameter optimization, this study employs a swarm intelli-
optimization algorithm for parameter adjustment and optimization.
gence optimization algorithm for parameter adjustment and optimization.
The predictive model in this study is based on the LSTM architecture, where key
hyperparameters—such as the number of neurons, dropout rate, batch size, and number of
training epochs—significantly influence both prediction accuracy and convergence speed.
To enhance the model’s performance, the Particle Swarm Optimization (PSO) algorithm is
used to optimize these hyperparameters [23].
Figure 7 presents the overall framework of the PSO-LSTM model. Starting from
the data preprocessing stage, the purple box section performs cleaning, missing value
imputation, and normalization on the raw charging pile load data to ensure the quality
 speed. To enhance the model’s performance, the Particle Swarm Optimization (PSO) al-
gorithm is used to optimize these hyperparameters [23].
Figure 7 presents the overall framework of the PSO-LSTM model. Starting from the
World Electr. Veh. J. 2025, 16, 150 data preprocessing stage, the purple box section performs cleaning, missing value impu- 11 of 19
tation, and normalization on the raw charging pile load data to ensure the quality and
consistency of the input data. Hyperparameter optimization plays a critical role in en-
and consistency
hancing of the input
model performance. In data.
the blueHyperparameter
box section, the optimization
PSO algorithm playsis aused
critical role in
to opti-
mize the hyperparameters of the LSTM model, ensuring that the model identifies the glob- to
enhancing model performance. In the blue box section, the PSO algorithm is used
optimize
ally optimalthe hyperparameters
parameter of the
combination LSTMcomplex
within model, ensuring that the
data patterns, model improving
thereby identifies the
globally optimal parameter combination within complex data patterns,
prediction accuracy and stability. The model construction phase is represented within thereby improving
the
green box, where, based on the optimized parameters, a prediction model is built thatthe
prediction accuracy and stability. The model construction phase is represented within
green box,
includes where, based
a bidirectional LSTMon the
layer,optimized
a dropout parameters,
layer, and aaprediction modellayer.
fully connected is builtThe that
includes a LSTM
bidirectional bidirectional LSTM layer,
layer captures a dropout
both forward andlayer, and atemporal
backward fully connected layer. in
dependencies The
the data, while the dropout layer effectively prevents overfitting by randomly deactivat- in
bidirectional LSTM layer captures both forward and backward temporal dependencies
ingthe data, while
neurons. the dropout
The fully layer
connected effectively
layer prevents
integrates overfitting
the extracted by randomly deactivating
high-dimensional features
and outputs the final prediction results. This framework not only significantly features
neurons. The fully connected layer integrates the extracted high-dimensional enhances and
theoutputs
model’sthe final prediction
predictive results.
performance butThis
alsoframework not only
provides reliable significantly
technical support enhances
for accu-the
model’s predictive performance but also provides
rate load forecasting of electric vehicle charging stations. reliable technical support for accurate
load forecasting of electric vehicle charging stations.

Figure 7. The overall framework of the PSO-LSTM model.

Figure 7. The overall framework of the PSO-LSTM model.
In the LSTM-based hybrid model, the network is structured with two layers: the
In the LSTM-based
first layer consists of hybrid model,
two LSTM the network
layers, is structured
while the second layerwith two layers:
contains the first
a single LSTM
layer
layer. Consequently, in the Particle Swarm Optimization (PSO) algorithm, the layer.
consists of two LSTM layers, while the second layer contains a single LSTM particle
Consequently,
dimension isinsetthe to Particle Swarm Optimization
five, corresponding (PSO) algorithm,the
to five hyperparameters: thenumber
particleofdimen-
neurons
sion is set to five, corresponding to five hyperparameters: the number
in the first hidden layer (Hidden Units1), the number of neurons in the of neurons
secondinhidden
the
first hidden
layer layerUnits2),
(Hidden (Hiddenthe Units1),
dropout therate
number of neurons
(Dropout), in the
the batch second
size (Batchhidden layerthe
Size), and
number of training epochs (Epochs).
Initially, the position range of the particles is defined by establishing the upper and
lower bounds for the five dimensions. The Mean Squared Error (MSE) of the GRU-RNN
model on the validation set is used as the fitness function. Subsequently, the fitness value
of each particle is calculated, and the particles’ velocity and position are updated according
to the PSO update mechanism. The number of particles is set to 10, and the maximum
number of iterations is set to 30, with both learning factors c1 and c2 set to 2.05.
World Electr. Veh. J. 2025, 16, 150 12 of 19

The dataset used in this study, after undergoing missing value imputation, reorganiza-
tion, and standardization, consists of a total of 2857 samples. Among them, 2352 samples
are used in the validation set and fed into the model for parameter optimization. Once
the optimization process is complete, the position of the global best particle in the five-
dimensional space represents the optimal hyperparameters for the LSTM model.
The specific process for optimizing LSTM parameters using PSO is as follows:
Step 1: Initialize the parameters of the LSTM algorithm. Define the five hyperpa-
rameters to be optimized as the five-dimensional particle space. Initialize the position X
and velocity V of each particle; set the number of particles, N; and specify the number of
iterations, num.
Step 2: Define the fitness function as the mean squared error (MSE) of the LSTM model
on the validation set. Compute the fitness values of all particles and identify the positions
corresponding to the initial personal best and global best, with the lowest MSE values
serving as the criteria.
Step 3: Update the velocity and position of each particle according to Equations (9)
and (10), and then recalculate the fitness value for each particle.
Step 4: Update the individual best position and the global best position.
Step 5: Check whether the termination condition is met. If yes, output the global best
position; if not, return to Step 3.
The final position of the global best particle in the five-dimensional space at the
conclusion of the optimization process is considered the optimal set of hyperparameters for
the LSTM model. Based on these optimized parameters, the PSO-LSTM predictive model is
constructed. The initialization parameters for the PSO-LSTM model are shown in Table 3.

Table 3. Initialization of PSO-LSTM parameters.

Parameter Initial Range

Population size, N 10
Number of iterations, num 30
Learning factor c1 2.05
Learning factor c2 2.05
Hidden Units1 64
Hidden Units2 32
Dropout 0.3
Batch size 32
Epochs 50

3. Experimental Results and Analysis

3.1. Experimental Setup
3.1.1. Experimental Environment Setup
The experiments were conducted on a Windows 10 operating system, utilizing the
TensorFlow framework within the Anaconda environment. Programming was performed
using the PyCharm editor. Anaconda is an open-source distribution that simplifies package
management and environment management, offering built-in tools that do not require
independent installation. It supports switching between different environments and is
designed for the parallel use of multiple versions of Python [24]. TensorFlow is an open-
source library commonly used for developing and implementing machine learning models,
and it is regarded as one of the most popular frameworks in the field of machine learning.
It includes numerous packages for scientific computing, image processing, and other tasks
and supports automatic differentiation and model construction. PyCharm, developed by
JetBrains, is a powerful and feature-rich Integrated Development Environment (IDE) specif-
World Electr. Veh. J. 2025, 16, 150 13 of 19

ically designed for Python programming. It offers features such as code editing, debugging,
testing, version control, and more, specifically designed for Python development, including
powerful code autocompletion, code inspection, and one-click code navigation.
The specific experimental environment is shown in Table 4.

Table 4. Experimental environment.

Operating System Windows 10

CPU 13th Gen Intel(R) Core (TM) i7-1360P 2.20 GHz
System RAM 32 GB
Learning framework Tensorflow
Editor PyCharm
Programming language Python3.9

3.1.2. Model Parameter Settings

In this study, the PSO algorithm was used to fine-tune the key hyperparameters of
the LSTM network, with the aim of enhancing its forecasting accuracy [25]. A two-layer
LSTM network model was constructed, where the first layer consisted of a bidirectional
LSTM to improve the model’s ability to capture temporal dependencies in sequential data.
The optimization process focused on five key hyperparameters: the number of neurons
in the first hidden layer (Hidden Units1), the number of neurons in the second hidden
layer (Hidden Units2), the dropout rate (Dropout), the batch size (Batch Size), and the total
number of training epochs (Epochs).
To capture complex nonlinear relationships within input sequences, the search range
for the number of neurons in the first hidden layer was set between 32 and 128, while
the range for the second hidden layer was limited to 32 to 64 neurons. The dropout rate
was optimized between 0.1 and 0.5 to balance learning capacity and generalization ability,
aiming to prevent overfitting while maintaining model complexity. For the batch size, a
range of 16 to 64 was selected to balance computational efficiency and model performance.
Finally, the number of training epochs was set within the range of 10 to 100 to ensure
adequate training and model convergence.
During the PSO algorithm’s initialization, the population size, N, was set to 10, with
each particle representing a possible hyperparameter combination. The fitness function
was defined as the MSE of the LSTM model on the validation set, with the goal of mini-
mizing MSE to assess the quality of various hyperparameter combinations. The number of
iterations (num) was set to 30, indicating that the algorithm would run for a maximum of
30 iterations to identify the optimal solution. During the optimization process, the Mean
Squared Error between predicted and actual values was used to guide the selection of the
best hyperparameter combination at each iteration. To improve optimization performance,
the inverse of the RMSE was incorporated into the fitness function, ensuring that the
optimal neural network structure parameters identified during optimization were applied
to construct the LSTM model.
To improve the accuracy of sequence prediction, this study standardized all subse-
quences and then input them into the PSO-LSTM model for training and forecasting. The
training data were divided into three batches based on the seasons: winter, winter–spring
transition, and spring. The data for each season were independently used to optimize the
LSTM model, with the optimal hyperparameters determined through the PSO algorithm.
After multiple iterations, the optimal hyperparameter combinations for each season were
identified. Based on the optimization results from the PSO algorithm, the parameters for
the PSO-LSTM model were configured, while the GRU and LSTM models used their default
settings. Table 5 provides a detailed listing of the hyperparameter settings for all models.
World Electr. Veh. J. 2025, 16, 150 14 of 19

Table 5. Parameter settings for all models.

Model Hidden Units1 Hidden Units2 Dropout Batch Size Epochs

GRU 64 32 0.3 32 50
LSTM 64 32 0.3 32 50
PSO-LSTM (Winter) 115 57 0.2 16 43
PSO-LSTM
110 28 0.1 16 51
(Winter–Spring Transition)
PSO-LSTM (Spring) 128 61 0.1 17 24

3.2. Comparative Analysis of Experimental Results

To comprehensively evaluate the model’s prediction performance, this study em-
ployed three key metrics: Mean Absolute Error (MAE), Mean Squared Error (MSE), and
Root Mean Squared Error (RMSE). These metrics provide distinct perspectives on the
difference between the predicted and actual values.
1. MAE is the average of the absolute errors:
It determines which information should be discarded from the cell state.

1 n
n i∑
MAE = |yi − ŷi | (11)
=1

where n is the number of samples, yi represents the actual values, and ŷi denotes the
predicted values.
2. MSE is the average of the squared errors:
It controls the storage of new information.

1 n
n i∑
MSE = (yi − ŷi )2 (12)
=1

where n is the number of samples, yi represents the actual values, and ŷi denotes the
predicted values.
3. RMSE is the square root of the Mean Squared Error (MSE):
It determines the output of the hidden state.
s
1 n
n i∑
RMSE = (yi − ŷi )2 (13)
=1

where n is the number of samples, yi represents the actual values, and ŷi denotes the
predicted values.
During the experiment, we independently trained and tested three different time-series
forecasting models—GRU, LSTM, and PSO-LSTM—on the same dataset. This process was
crucial for capturing the models’ performance under different random initialization condi-
tions, allowing for an accurate assessment of their stability [26]. Specifically, the training
process for the GRU and LSTM models involved data normalization, dataset creation, net-
work structure definition, and training loops. In contrast, the LSTM model optimized using
PSO incorporated an additional phase, where the Particle Swarm Optimization algorithm
was applied to fine-tune the parameters.
This study conducted three independent experiments using charging load data from
different seasons, namely, winter, winter–spring transition, and spring, for model training
and testing. Figures 8–10 present comparison curves between the real load values and the
 tion
mizedalgorithm
using PSO wasincorporated
applied to fine-tune the parameters.
an additional phase, where the Particle Swarm Optimiza-
tion This studywas
algorithm conducted
appliedthree independent
to fine-tune experiments using charging load data from
the parameters.
different seasons, namely, winter, winter–spring
This study conducted three independent experiments transition, using
and spring,
charging for model training
load data from
World Electr. Veh. J. 2025, 16, 150 and testing. Figures 8–10 present
different seasons, namely, winter,comparison
winter–spring curves between
transition, andthespring,
real loadfor values and
15 ofthe
model training 19
predicted
and testing.results
Figuresfrom different
8–10 presentmodels (GRU,curves
comparison LSTM,between
PSO-LSTM) for load
the real eachvalues
season.and
In the
the
figures,
predictedtheresults
black from
solid different
line represents
modelsthe real LSTM,
(GRU, load values, the green
PSO-LSTM) for dashed line repre-
each season. In the
predicted
sents results from
thethe
predicted different
results models (GRU, LSTM, PSO-LSTM) for each season.the
Inpre-
the
figures, black solid linefrom the GRU
represents themodel, the values,
real load blue dashed line
the green represents
dashed line repre-
figures,
dicted the black
results solid
from the line represents
original the real load values, the green dashed line represents
sents the predicted results fromLSTM
the GRUmodel, andthe
model, theblue
red dashed
dotted line
line represents
represents thethe pre-
pre-
the predicted
dicted results results
from from
the the GRU model,
PSO-optimized LSTM themodel.
blue dashed line represents the predicted
dicted results from the original LSTM model, and the red dotted line represents the pre-
results from the original LSTM model, and the red dotted line represents the predicted
dicted results from the PSO-optimized LSTM model.
results from the PSO-optimized LSTM model.

Figure 8. Winter charging load prediction comparison (1 January 2023–3 February 2023).
Figure 8. Winter charging load prediction comparison (1 January 2023–3 February 2023).
Figure 8. Winter charging load prediction comparison (1 January 2023–3 February 2023).

Figure 9. Winter–spring transition charging load prediction comparison (4 February 2023–

4 March 2023).

The model output data were organized and summarized for evaluation metric analysis,
with MAE, MSE, and RMSE calculated for each model. MAE provides a straightforward
measure of the average absolute deviation between the model’s predicted values and
the true values. A smaller MAE indicates that the model’s predictions have a smaller
average error, meaning the predictions are closer to the actual values. MSE averages the
squared prediction errors, amplifying the effect of larger errors and offering a more sensitive
reflection of how well the model handles outliers. A smaller MSE suggests better overall
prediction accuracy. RMSE, the square root of MSE, has the same units as the original data,
making it easier to compare with the actual data and providing a more accurate measure of
the average deviation between predicted and true values.
 Figure 9. Winter–spring transition charging load prediction comparison (4 February 2023–4 March
World Electr. Veh. J. 2025, 16, 150 16 of 19
2023).

Figure 10. Spring

Figure 10. Spring charging
charging load
load prediction
predictioncomparison
comparison(5
(5March
March2023–29
2023–29April
April2023).
2023).

Table 6 summarizes
The model output data all were
the evaluation
organized metrics, which helps
and summarized us gain insights
for evaluation metric from
anal-
multiple dimensions into the strengths and weaknesses of each model under
ysis, with MAE, MSE, and RMSE calculated for each model. MAE provides a straightfor- different
seasonal conditions. This will provide a solid and reliable foundation for further model
ward measure of the average absolute deviation between the model’s predicted values
optimization and the development of charging load forecasting strategies.
and the true values. A smaller MAE indicates that the model’s predictions have a smaller
average error, meaning the predictions are closer to the actual values. MSE averages the
Table 6. Comparison of model results.
squared prediction errors, amplifying the effect of larger errors and offering a more sen-
Model sitive reflection
Seasonof how well the modelMAE handles outliers. A smaller MSE suggests
MSE RMSE better
overall prediction
Winter accuracy. RMSE, the square
4.335 root of MSE, has
33.171 the same units as the orig-
5.760
GRU inal data, making
Winter–spring it easier to compare with
transition 4.365the actual data33.382
and providing a more5.778accurate
measure of the average deviation between
Spring predicted and 31.046
4.219 true values. 5.572
TableWinter
6 summarizes all the evaluation4.170 metrics, which helps us gain insights
29.701 5.450 from
LSTM multiple dimensions
Winter–spring into the strengths4.050
transition and weaknesses 30.625
of each model under different
5.534
Spring 4.019 31.292 5.594
seasonal conditions. This will provide a solid and reliable foundation for further model
optimization and the development of charging
Winter 3.896 load forecasting
28.717 strategies. 5.359
PSO-LSTM Winter–spring transition 3.806 29.012 5.386
Spring of model results.
Table 6. Comparison 3.910 29.796 5.458

Model Season MAE MSE RMSE

Based on the performance metrics
Winter 4.335(MAE, MSE, RMSE) for different models
33.171 across the
5.760
GRU seasons in Table
Winter–spring 6, this study presents4.365
transition a detailed analysis from the following three 5.778
33.382 perspectives:
• Spring
Winter Period (1 January–3 February):
4.219 The MAE of 31.046 the PSO-LSTM is 3.896, which is a
5.572
reduction of 6.6% compared to4.170
Winter LSTM (4.170) and 10.1% compared to GRU
29.701 (4.335). Its
5.450
LSTM Winter–spring
RMSEtransition 4.050 than those of both
(5.359) is significantly lower 30.625
LSTM (5.450) and GRU 5.534 (5.760),
Spring
indicating stronger stability in4.019 31.292
high-variance load scenarios. 5.594
• Winter
Winter–Spring Transition Period 3.896
(4 February–4 March):28.717
PSO-LSTM achieves 5.359
the lowest
PSO-LSTM Winter–spring
MAE transition 3.806
(3.806) and RMSE (5.386), demonstrating 29.012 to dynamic change
its adaptability 5.386patterns.
• Spring
Spring Period (5 March–29 April): 3.910 Although the performance
29.796 5.458
differences between
models are narrow, the RMSE of PSO-LSTM (5.458) still outperforms GRU (5.572),
Based on the
indicating performance metrics
its competitiveness (MAE,
in stable MSE, RMSE) for different models across
scenarios.
the seasons
From a comprehensive analysis of the three seasons’analysis
in Table 6, this study presents a detailed data, thefrom
GRUthe following
model three
consistently
perspectives:
exhibits higher and more fluctuating error metrics across all seasons, with poor adaptability
to seasonal changes, making it less effective in accurately capturing the variations in
charging load across different seasons. The original LSTM model performs better than GRU
World Electr. Veh. J. 2025, 16, 150 17 of 19

in terms of prediction accuracy but still falls short compared to PSO-LSTM in most cases,
suggesting that although the LSTM model has some capacity for handling sequential data,
it has limitations when facing complex seasonal charging load patterns. The PSO-LSTM
model, however, maintains a low error level across all three seasons (winter, winter–
spring transition, and spring), reflecting excellent generalization ability and adaptability to
different seasonal data characteristics. Furthermore, it outperforms both GRU and original
LSTM models in key evaluation metrics—MAE, MSE, and RMSE—demonstrating that
the LSTM model optimized by the PSO algorithm offers superior performance and higher
stability in load forecasting.
In conclusion, the LSTM model’s powerful sequence modeling capability gives it an
advantage in time-series forecasting. When combined with PSO algorithm-based parameter
optimization, the model’s performance is further enhanced, improving both prediction
accuracy and computational efficiency. These findings not only validate the effectiveness
of the LSTM model and PSO algorithm in time-series forecasting but also provide new
directions for future research [27]. This involves examining the integration of various opti-
mization algorithms with deep learning models and exploring their potential applications
across a broader range of fields.

4. Conclusions
This study proposes a PSO-LSTM network model to enhance the accuracy and ro-
bustness of electric vehicle charging load forecasting. The LSTM model is trained using
historical load data, and the PSO algorithm dynamically optimizes its key hyperparame-
ters, such as the number of hidden layer neurons, dropout rate, batch size, and number
of training epochs. Experimental results show that the PSO-LSTM model significantly
outperforms traditional methods across different seasonal scenarios. For example, in winter,
its MAE (3.896), MSE (28.717), and RMSE (5.359) are reduced by 10.13%, 13.43%, and 6.96%
compared to the GRU model and by 6.57%, 3.31%, and 1.67% compared to the original
LSTM model, validating the effectiveness of the PSO algorithm in parameter tuning. Addi-
tionally, the model maintains stable prediction performance in the winter–spring transition
period and spring, demonstrating its ability to effectively capture seasonal load variation
patterns and reduce the impact of climate and user behavior differences on forecast re-
sults, providing a highly adaptable solution for power system scheduling and charging
facility planning.
Future research can expand in the following directions: First, by integrating me-
teorological data (e.g., temperature, humidity, wind speed) and traffic information, a
multi-variable input model could be constructed to more comprehensively characterize the
coupling relationship between charging behaviors and external environmental factors [28].
Second, a hybrid model architecture combining BiLSTM and a CNN could be developed,
leveraging the bidirectional temporal modeling capability of BiLSTM and the spatial feature
extraction advantages of CNN to further enhance the model’s ability to analyze complex
spatiotemporal patterns [29].

Author Contributions: Study conception and design: L.Z.; data collection: X.Y. and X.H.; analysis and
interpretation of results: X.H. and L.Z.; draft manuscript preparation: L.Z. and X.Y.; draft manuscript
editing and reviewing: X.H. and L.Z. All authors have read and agreed to the published version of
the manuscript.

Funding: This research was funded by [the National Natural Science Foundation of China] grant
number [61773243]; and [the Science and Technology Project of Chongqing Municipal Education
Commission] grant number [KJZD-K202101702].
World Electr. Veh. J. 2025, 16, 150 18 of 19

Data Availability Statement: The raw data supporting the conclusions of this article will be made
available by the authors upon reasonable request.

Conflicts of Interest: The authors declare that they have no conflicts of interest.

References
1. Lee, Z.J. Large-Scale Adaptive Electric Vehicle Charging. In Proceedings of the 2018 IEEE Global Conference on Signal and
Information Processing (GlobalSIP), Anaheim, CA, USA, 26–28 November 2018; pp. 863–864. [CrossRef]
2. Wei, J.Z.; Ma, Z.P. Monte-Carlo-algorithm-based Load Prediction of Electric Vehicles Large-scale Charging. Electr. Eng. 2024,
3, 49–53. [CrossRef]
3. Shi, W.Q.; Wu, K.Y.; Wang, D.X. Eclectic Power System Short-Term Load Forecasting Model Based on Time Series Analysis and
Kalman Filter Algorithm. Control Theory Appl. 2018, 37, 9–12+23.
4. Zhang, X.; Li, L. Seasonal Electric Vehicle Charging Load Prediction Based on Random Forest. Softw. Eng. 2024, 27, 11–14+37.
5. Song, M.S.; Li, Z.W.; Song, S. Research on the Optimization Strategy of Electric Vehicle OrderlyCharge and Discharge in Intelligent
Community. Tech. Autom. Appl. 2022, 41, 17–22+27. [CrossRef]
6. Geng, P.; Yang, H.J.; Shi, Z.X. Electric Vehicle Forecasting Charging Demand Based on Spatiotemporal Graph Convolutional
Networks. J. Transp. Eng. 2024, 24, 37–45.
7. Pei, Z.; Zhang, Z.; Chen, J. KAN-CNN: A Novel Framework for Electric Vehicle Load Forecasting with Enhanced Engineering
Applicability and Simplified Neural Network Tuning. Electronics 2025, 14, 414. [CrossRef]
8. Liu, Y.X.; Gao, H. Load Prediction Method of Charging Station Based on SSA-VMD-BiLSTM Model. Guangdong Electr. Power 2024,
37, 53–61.
9. Lin, X.; Zhang, H.; Ma, Y.L. Electric vehicle charging load prediction based on improved LSTM neural network. Mod. Electron.
Tech. 2024, 47, 97–101.
10. Huang, Y.X.; Xiao, S.W. Forecasting of electric vehicle charging load in highway service areas considering meteorological factors.
Appl. Energy 2025, 383, 125337.
11. Yin, W.; Ji, J. Research on EV charging load forecasting and orderly charging scheduling based on model fusion. Energy 2024,
290, 130126. [CrossRef]
12. Ma, S.; Ning, J.; Mao, N.; Liu, J.; Shi, R. Research on Machine Learning-Based Method for Predicting Industrial Park Electric
Vehicle Charging Load. Sustainability 2024, 16, 7258. [CrossRef]
13. Ge, Q.; Guo, C.; Jiang, H. Industrial power load forecasting method based on reinforcement learning and PSO-LSSVM. IEEE
Trans. Cybern. 2020, 52, 1112–1124. [CrossRef] [PubMed]
14. Jin, Y.; Guo, H.; Wang, J. A hybrid system based on LSTM for short-term power load forecasting. Energies 2020, 13, 6241. [CrossRef]
15. Saoud, A.; Recioui, A. Load Energy Forecasting based on a Hybrid PSO LSTM-AE Model. Alger. J. Environ. Sci. Technol. 2023,
9, 2938–2946.
16. Liu, X.; Ma, Z.; Guo, H. Short-term power load forecasting based on DE-IHHO optimized BiLSTM. IEEE Access 2024, 12, 145341–145349.
[CrossRef]
17. Lai, Y.; Wang, Q.; Chen, G. Short-term Power Load Prediction Method based on VMD and EDE-BiLSTM. IEEE Access 2024,
13, 10481–10488. [CrossRef]
18. Kennedy, J.; Eberhart, R. Particle swarm optimization. In Proceedings of the ICNN’95-International Conference on Neural
Networks, Perth, WA, Australia, 27 November—1 December 1995; Volume 4, pp. 1942–1948.
19. Liu, Z.; Chen, X.; Liang, X.; Huang, S.; Zhao, Y. Research on Sustainable Form Design of NEV Vehicle Based on Particle Swarm
Algorithm Optimized Support Vector Regression. Sustainability 2024, 16, 7812. [CrossRef]
20. Dai, X.; Sheng, K.; Shu, F. Ship power load forecasting based on PSO-SVM. Math. Biosci. Eng. 2022, 19, 4547–4567. [CrossRef]
21. Geng, G.; He, Y.; Zhang, J. Short-term power load forecasting based on PSO-optimized VMD-TCN-attention mechanism. Energies
2023, 16, 4616. [CrossRef]
22. Fan, W.; Hu, Z.; Veerasamy, V. PSO-based model predictive control for load frequency regulation with wind turbines. Energies
2022, 15, 8219. [CrossRef]
23. Jain, M.; Saihjpal, V.; Singh, N. An overview of variants and advancements of PSO algorithm. Appl. Sci. 2022, 12, 8392. [CrossRef]
24. Kim, H.J.; Kim, M.K. Spatial-Temporal Graph Convolutional-Based Recurrent Network for Electric Vehicle Charging Stations
Demand Forecasting in Energy Market. IEEE Trans. Smart Grid 2024, 15, 3979–3993. [CrossRef]
25. Güven, A.F. Integrating electric vehicles into hybrid microgrids: A stochastic approach to future-ready renewable energy solutions
and management. Energy 2024, 303, 131968. [CrossRef]
26. Ding, L.; Ke, S.; Zhang, F. Forecasting of electric-vehicle charging load considering travel demand and guidance strategy. Electr.
Power Constr. 2024, 45, 10–26.
World Electr. Veh. J. 2025, 16, 150 19 of 19

27. Zhang, Q.; Lu, J.; Kuang, W.; Wu, L.; Wang, Z. Short-Term Charging Load Prediction of Electric Vehicles with Dynamic Traffic
Information Based on a Support Vector Machine. World Electr. Veh. J. 2024, 15, 189. [CrossRef]
28. Qian, Y.; Kong, Y.; Huang, C. Review of Power Load Forecasting. Sichuan Electr. Power Technol. 2023, 46, 37–43+58. [CrossRef]
29. Zhang, X.W.; Liang, J.; Wang, Y.G.; Han, J. Overview of Research on Spatiotemporal Distribution Prediction of Electric Vehicle
Charging. Electr. Power Constr. 2023, 44, 161–173.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.