Discover Applied Sciences

Research

Prediction of electric vehicle charging demand using enhanced gated

recurrent units with RKOA based graph convolutional network
R. Gunasekaran1 · Manjunatha B.2 · Anand S.3 · Piyush Kumar Pareek4 · Sandeep Gupta5 · Anand Shukla6

Received: 7 February 2024 / Accepted: 5 November 2024

© The Author(s) 2024  OPEN

Abstract
Accurate forecasting of traffic patterns plays a crucial role in the effective management and planning of urban trans-
portation infrastructure. In particular, predicting the availability of electric vehicle (EV) charging stations is essential for
alleviating range anxiety among drivers and facilitating the adoption of electric vehicles. This study proposes a novel
deep learning-based predictor model to approximate the demand for charging electric vehicles over the long term. The
methodology integrates the Berkeley wavelet transform (BWT) to decompose input time series data while preserving
its inherent characteristics. The proposed hybrid prediction model combines an enhanced gate recurrent unit with an
optimized convolution kernel within a fusion graph convolutional network (GCN). The Red Kite Optimization Algorithm
(RKOA) is employed to select the convolution kernel of the GCN effectively. Additionally, the construction of the graph lev-
erages both adjacency and adaptive graphs to accurately represent the correlations among nodes in the EV network. The
model extracts multi-level spatial correlations through stacked fusion graph convolutional elements and captures multi-
scale temporal correlations via an improved gated recurrent unit. Furthermore, the incorporation of residual connection
units allows for the fusion of extracted spatiotemporal features with direct data, enhancing predictive performance.
The proposed neural predictor is evaluated using EV charging data from Georgia Tech in Atlanta, USA. The experimental
results demonstrate the effectiveness of the prediction metrics generated by the proposed model compared to existing
methods reported in the literature, showcasing its capability to accurately forecast EV charging demand.

Article highlights
• In this research work, a novel deep learning (DL)-based predictor model is attempted to be developed for charging
electric vehicles.
• To suggests a hybrid prediction model that is built on an upgraded gate recurrent unit and an optimised convolution
kernel of a fusion graph convolutional network (GCN).
• Red Kite Optimisation Algorithm (RKOA) selects the convolution kernel of the GCN optimally. The outcomes demon-
strate the effectiveness of the prediction metrics calculated using the suggested neural predictor for the examined
dataset when compared to earlier methods from published studies.

* Sandeep Gupta, [email protected]; * Anand Shukla, [email protected]; [email protected]

| 1Excel Engineering College, Komarapalayam, Tamil Nadu, India. 2New Horizon College of Engineering, Ring Road, Bellandur Post,
Bengaluru 560103, India. 3Department of EEE, Nitte Meenakshi Institute of Technology Bengaluru, Bengaluru, India. 4Department of AIML
and IPR Cell, Nitte Meenakshi Institute of Technology Bengaluru, Bengaluru, India. 5Department of Electrical Engineering, Graphic Era
(Deemed to Be University), Dehradun, Uttarakhand 248002, India. 6Wollega University, Nekemte, Ethiopia.

Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

KeywordS Traffic prediction · Red kite optimization algorithm · Berkeley wavelet transform · Graph convolutional
network · Electric vehicle charging station

1 Introduction

One of the most important pieces of contemporary urban infrastructure is the intelligent traffic system development.
The two main components of a smart system are typically the public transport infrastructure, such as vehicle stations
and traffic lights, and the traffic highways. The traffic management team faces challenges in the form of road congestion
and car emission concerns due to the significant growth in demand for commuting [1]. Enhancing the effectiveness of
the transport system and mitigating the problems associated with commuting has been investigated through the use
of very accurate and long-term forecasts for public transport facilities and road information, including flow, speed, and
signal systems. An electric-powered vehicle, or EV, is a low-emission, environmentally friendly form of transportation.
The number of EVs being owned is rising quickly as more well-known automakers, like Tesla, release additional EV alter-
natives onto the market. Since electric vehicles (EVs) do not generate CO2, they have a significant potential to prevent
air pollution. Additionally, users find using electric vehicles appealing since charging an EV’s battery is far less expensive
than refuelling [2].

1.1 Importance of EV

In this scenario, the global utility landscape has experienced a significant surge driven by the rapid proliferation of elec-
tric vehicles (EVs). In 2021 alone, the production of new battery-electric cars reached three million globally, constituting
a notable portion of the 6.8 million battery-powered vehicles in operation [3]. The EV market in India exemplifies this
growth trend, with a remarkable 168% increase observed as evidenced by the sale of 329,190 cars in 2021, compared to
122,607 units sold in 2020 [3]. However, the surge in EV usage poses challenges, exacerbated by the expanding global
population and the escalating number of vehicles on the roads [4]. The distribution of electric power converters is integral
to the charging process of EVs, providing the necessary DC power to the battery source. This is essential for the efficient
operation of EVs, with various power ratings and charging options available, as detailed in Table 1. Both direct current
(DC) and alternating current (AC) charging modes are utilized for conductive charging of EVs, ensuring flexibility and
compatibility with diverse charging infrastructures.
When power is used at home and by businesses, the cost of energy is low compared to that of fuel and diesel, and the
cost of battery recharging is economical. Compared to conventional automobiles, electric vehicles require less mainte-
nance since their auto components experience less wear and tear [5]. The cost of maintenance is less complicated than
with combustion engines. To encourage the people to use electric vehicles and enforce the use of green technology, the
government has launched a number of incentives.

1.2 EV challenges

The recent advancements in electric car battery efficiency have led to notable improvements in the driving range of
electric vehicles (EVs), with some models now capable of traveling over 500 km on a single charge. However, despite
these advancements, charging EV batteries remains a time-consuming process, exacerbated by the limited availability of
charging infrastructure compared to traditional petrol stations [6]. As EV ownership continues to rise, several challenges
become more pronounced. Firstly, charging an EV at home using a standard power source typically takes six to eight
hours, while utilizing a high-voltage rapid charging station can still require up to thirty minutes, significantly longer than
the few minutes it takes to refuel a conventional vehicle. Secondly, the scarcity of charging stations necessitates careful

Table 1  EV Charging and Charging mode Power rating (P) Supply
power score
Normal power charging Less than otherwise equal to 7 kW DC and AC
7–22 kW DC and AC
High power charging 22–50 kW only DC supply
50–200 kW Only DC supply

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

planning by drivers to avoid lengthy charging waits, exacerbating the inconvenience of EV ownership [7]. Thirdly, the
charging behavior of EVs exhibits periodicity, characterized by distinct peak and off-peak periods at charging stations,
influenced by factors such as battery endurance and real-time traffic conditions [8].
The primary obstacles hindering widespread EV adoption include concerns regarding charging schedules, insufficient
charging infrastructure, communication technology requirements, and consumer awareness [9]. The use of AI in electric
vehicle charging is another critical concern [10, 11]. The paper showcases a deep learning-based time-series estimate for
the monthly demand for commercial EV charging [10]. A neural network-based approach is suggested for the precise
prediction of EV charging needs [11]. Among these challenges, determining optimal charging times for EVs emerges as
a top priority. Projections suggest a significant surge in EV sales by 2030, with electric vehicles comprising nearly half
of all automobiles on the road [12]. However, this rapid growth in EV usage poses potential strains on the utility system,
resulting in increased energy consumption, unexpected demand peaks, and fluctuations in frequency and voltage.
Furthermore, simultaneous charging of multiple EVs could lead to power outages if not managed efficiently.

1.3 Aim and motivation

The aforementioned limitations serve as the driving force behind this research study’s motivations. Accordingly, in order
to get around these restrictions, a better predictive model must be created and developed that can most accurately
forecast the charging request for electric vehicles using the input limits that have been taken into consideration and
the features that have been derived from them [13]. Because it will alert users and drivers to take extra precautions
when charging is necessary, an accurate prediction perfect for electric car charging demand is crucial. Maintaining the
balance time spent travelling, amount of time needed for charging, and cost will be made easier with the use of a more
appropriate prognostic model for predicting the charging needs of electric cars [14].
The predictor model is intended to estimate the number of electric cars that will require charging at a specific sector
in terms of demand. According to the datasets under consideration, for instance, if 50 cars need to be charged during
hours, the corresponding demand would be higher than it need to be charged [15]. Therefore, stations will be able to
better meet the need once the charging demand has been analysed and anticipated with regard to the sectors and time
zones. If more pass by in a lane, the prediction and increasing demand will aid in the installation of charging stations,
reducing the amount of time that EV users must wait to charge.

1.4 Contribution of the study

To address the challenges outlined above, there is a critical need for an intelligent and coordinated approach to electric
vehicle (EV) charging. This entails the ability to adjust or reschedule charging times based on factors such as demand
fluctuations, current pricing structures, and user preferences. Accurate predictions of EV load demand are therefore
essential to facilitate this process. Forecasting EV load demand not only aids utility companies in strategic planning and
decision-making but also contributes to system stability and error prevention [16].
The primary objective of this study is to develop a predictive model that anticipates EV charging demand, thus ena-
bling better management of resources. By achieving this goal, it becomes possible to strike a balance between expendi-
ture and factors such as travel time, distance, and charging duration, thereby optimizing the overall charging experience.
The following are our research’s primary contributions:

• To create a new recurrent neural network perfect based on deep learning that uses autoencoders and decoders to
handle non-linear data related to charging electric vehicles.
• Using the intrinsic frequency components of the BWT, break down the data to obtain the temporal properties.
• The adaptive graph generated by layer is integrated with the adjacency graph based on geographic location to cre-
ate a fusion graph structure that improves the representation of node interactions. Better exploration of the spatial
correlations among traffic data is achieved by using a network that is based on the fusion graph.
• In order to overcome the shortcoming of recurrent neural networks in collecting long-term dependencies of data,
an improved gated recurrent unit was created. In order to store and transmit the correlation info in flow data more
effectively, this unit employs gating mechanisms and attention scores.
• The suggested model will be tested and validated using Georgia Tech’s electric vehicle charging dataset in Atlanta,
USA.
• To guarantee that the waiting time for EVs is shortened within a 24-h period based on the prediction made

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

The remainder of the study is divided into the subsequent sections: Sect. 2 provides a detailed overview of relevant
efforts that have been done in this field in earlier publications. The construction of the suggested datasets used in this
investigation are elucidated in Sect. 3. Section 4 provides the simulated findings and the discussions regarding the
achieved solution set. Section 5 addresses the paper’s conclusions.

2 Related works

The Multi-Objective Particle Swarm Optimisation (MOPSO) algorithm was proposed by Chakraborty and Pal [16] as a
solution to the problems of power loss and voltage fluctuations. In order to analyse DS and EV load time series and make
predictions regarding dynamic pricing, they leveraged OpenDSS and MATLAB. The test system demonstrated squared
voltage changes of 0.0235 p.u. and a power loss of 164.36 kW regardless of whether or not it was equipped with Electric
Vehicle Charging Infrastructure (EVCI). Both the voltage fluctuation and the power loss were reduced to 0.0182 p.u. and
201.40 kW, respectively, as a result of the implementation of their plan with five EVCIs, each of which contained twenty
EVCSs operating at 50 kWh. The value of the R-squared for the dynamic price prediction was 0.9999, which indicates
that it is highly accurate.
A model that takes into account the heterogeneity of charging behaviour and demand schemes among electric vehicle
(EV) users was presented by Jiang et al. [17]. This model makes it possible to accurately forecast the demand for public
charging based on travel paths. The spatial–temporal accuracy of their model is beneficial to the planning of transport
network electrification, as demonstrated in a case study conducted in Los Angeles County that predicted the require-
ments for public charging stations under a variety of different scenarios.
An agent-based model (ABM) was utilised by Williams et al. [18] in order to investigate the effects that electric vehicle
charging has on the residential power networks in New Zealand. Using a Monte Carlo simulation that lasted for twenty
days and took place in a neighbourhood with seventy-one houses, they determined the rate of "Exceedance" of the 300
kVA transformer limit. Immediately charging led to an increase in exceedance, whereas algorithms for delayed charg-
ing led to a reduction in peak power consumption. According to their findings, New Zealand’s residential networks are
capable of managing a forty percent transition to electric vehicles through the use of altruistic charging, but they face
significant challenges caused by self-serving pricing.
A machine learning algorithm that makes use of KNN, SVR, DT, and RF regressors was proposed by Das et al. [19]
in order to estimate the health of a battery based on variables that can be measured. By utilising the battery dataset
provided by Sandia National Laboratories, their models demonstrated varying degrees of precision and practicability.
To illustrate, the mean absolute error (MAE) at 25 degrees Celsius for SVR, DT, KNN, and RF was 5.593E−02, 2.379E−03,
2.429E−03, and 5.073E−03, respectively. This demonstrates that KNN and DT have a higher degree of accuracy when it
comes to electric vehicle lithium-ion batteries.
Using an online booking system (OBS) and a primary charge control system (PCCS), Zhao et al. [20] developed a hierar-
chical charge scheduling technique with the goal of balancing peak demand and improving grid load. By incorporating
deep reinforcement learning (DRL), their model demonstrated significant improvements in efficacy, ranging from 6.7
to 49.44%, in comparison to alternatives that only utilised a single strategy. Utilising this strategy allows for the shifting
of charging demand from peak to off-peak hours, which optimises the representation of the actual charging process.
Through the utilisation of cumulative distribution functions (CDFs) and multi-feeder investigations, Azzopardi and
Gabdullin [21] presented tools that were designed to evaluate the effects of electric vehicle charging on the stability of
the grid. Their regression analysis revealed that there were strong relationships between feeder processes, which aided
in the management and planning of the network. Their study provides insights for future challenges, despite the fact
that there are no significant technological hurdles at the moment, as no low-voltage feeder on the Maltese grid has more
than one percent electric vehicle charges.
An electric vehicle charging time prediction model was developed by Ullah et al. [22] using data collected from 500
electric vehicles in Japan. Through the utilisation of feed-forward neural networks (FFNN) and metaheuristic techniques,
they were able to improve the accuracy of the parameters of the machine learning parameters. The findings of their pre-
diction showed that models based on Grey Wolf Optimisation (GWO) performed better than others, offering significant
improvements in the forecasting of the amount of time required for electric vehicle charging.
Using both traditional and deep learning time series methods, such as recurrent neural networks (RNNs), long short-
term memory (LSTMs), and transformers, Koohfar et al. [23] proposed estimating charging demand. Their research, which
was among the first to apply these models for predicting the demand for electric vehicle charging, demonstrated that

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

Fig. 1  Workflow of the pro-

posed model

the Transformer model performed exceptionally well in both short-term and long-term forecasts, with superior RMSE
and MAE performance across the board.
Through the utilisation of a multi-feature Long Short-Term Memory (LSTM) model that incorporated historical mete-
orological data, Aduama et al. [24] improved the ability to predict the demand for electric vehicle charging. The fact
that their method yielded a prediction error of 3.29 percent demonstrated significant improvements in comparison to
conventional LSTM models. According to this strategy, electric vehicle load station management is optimised to meet
energy demands.
In order to forecast the deterioration of battery capacity in electric vehicles, Zhao et al. [25] developed machine learn-
ing methods. Through the utilisation of a dataset consisting of 420 cells and 9 battery packs, they utilised a two-stage
noise reduction approach and stacking ensemble learning, resulting in an MPE of 0.28% and an RMSPE of 0.55% during
the process. Using their model, they were able to accurately estimate the remaining usable lifespan (RUL), which dem-
onstrates that there is significant potential for predicting battery performance.

3 Research gap

Each prediction technique has advantages and disadvantages that should be considered in relation to the comprehen-
sive literature analysis conducted on the various prediction models used for predicting the changing needs of plug-in
vehicles. Millions of electric cars are now on the road worldwide, meaning that the chosen charging system must con-
stantly be improved. The different shortcomings found in the current prediction models taking into account the review,
are listed as follows:

Fig. 2  Histogram of charging

period and distribution plot

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

Table 2  Example datasets of Charging time Cost incurred Energy (kwh) Ghg savings (kg) Gasoline sav-
charging (hh:mm:ss) (usd) ings (gallons)

01:54:49 1.78 5.153 2.164 0.647

06:20:20 6.17 19.28 8.098 2.42
04:47:36 4.73 12.334 5.18 1.548
03:12:16 2.73 9.975 4.19 1.252
01:58:14 1.68 6.075 2.551 0.762
02:23:58 2.04 7.758 3.258 0.974
01:37:25 1.39 9.55 4.011 1.199
03:24:24 2.9 11.275 4.735 1.415
01:01:50 0.88 6.061 2.546 0.761
01:23:29 1.19 4.019 1.688 0.504
02:29:45 2.12 9.085 3.816 1.14
01:11:50 1.02 6.249 2.625 0.784
00:58:15 0.83 4.352 1.828 0.546
01:11:24 1.02 4.341 1.823 0.545
03:19:30 4.12 7.857 3.3 0.986
03:10:19 2.7 12.815 5.382 1.608
03:27:05 2.94 13.826 5.807 1.735
04:37:06 4.25 12.408 5.211 1.557
04:48:07 5.56 18.272 7.674 2.293
03:57:31 4.52 9.594 4.03 1.204
04:43:32 4.5 12.967 5.446 1.627
02:06:29 1.96 4.789 2.011 0.601
06:11:21 8.96 19.524 8.2 2.45
04:45:17 4.05 15.359 6.451 1.927
05:07:46 4.36 12.336 5.181 1.548
04:18:45 3.67 14.575 6.122 1.829
00:01:34 0 0.081 0.034 0.01

• Existence of metrics of dissimilarity within the algorithms

• Insufficient mapping of the necessary feature parameters
• For some systems to make the prediction, additional information about the movement and tracking of electric vehicles
is needed.
• Different charging elapse times require longer prediction times. There is seldom data sharing between charging sta-
tion providers and stations.
• Non-linearity with time-dependent data exists; certain machine learning algorithms have delayed and premature
convergence.
• Existence of both global and local optima while being below the saturation limit
• The limited reliability and high mobility of electric cars
• The performance of current prediction models is affected by environmental conditions and occupant behaviour.
• Analysing the long-term energy consumption prediction is difficult.
• Depending on the state-of-charge level
• The impact of driving ability on preserving electric vehicle charging capacity.
• Non-linear data and a patchy charging infrastructure.
• Linear networks’ inadequate descriptive power in complicated situations
• The greatest rate of public charging demand; some predictive representations’ short-term focus; challenges with
handling temporal characteristics in algorithms

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

4 Proposed methodology

In this section, the brief explanation of the proposed model with its mathematical expression is given in Fig. 1, where
each block is defined in detailed manner.

4.1 EV charging datasets

The data used to validate and assess the performance of the proposed predictor model pertains to the electric vehicles
(EVs) operating on the Georgia Tech campus in Atlanta, USA. Approximately 150 vehicles were observed navigating the
campus and utilizing the charging infrastructure available at the station [26, 27]. These vehicles have an average driving
distance of thirty-one kilometers.
To provide insights into the charging behavior of these EVs, a probability distribution curve and a histogram plot
illustrating the charging duration are presented in Fig. 2.
Additionally, Table 2 displays the sample datasets comprising key variables relevant to EV charging, including energy
savings (in kilograms), petrol savings (in gallons), charge time (in hours, minutes, and seconds), and the variable under
consideration for forecasting the charging demand. The proposed predictive model for electric vehicle charging is devel-
oped and evaluated using these datasets, offering insights into the effectiveness of the model in predicting charging
demand in real-world scenarios.

4.2 Wavelet transform

The Berkeley wavelet transformation (BWT) is used in this study to break down the input signals. Wavelet technology is
centred on breaking down a function, operator, data, or information into multiple frequency components so that each
can be examined separately. Since they can be characterised by Eq. 1 and serve as the starting point for other wavelets,
fundamental wavelets, also known as mother wavelets (t), are the source of all other wavelets.
� �
1 t−𝜏
Ψs,𝜏 = √ Ψ (1)
s s

where s and 𝜏 are the translation and scale factors, correspondingly.

The BWT is a two-dimensional, three-dimensional wavelet that is applicable to signal processing. The threshold
marker and selection are required for the seed point in the BWT transformation. The BWT is fully orthonormal and
serves as a useful tool for representing signal change. Wavelet transformations create a full, orthonormal basis in 2-D
form by translating and scaling the entire set with a single wavelet constant term. The 𝛽x𝜑 transformation wavelet of
the mother is piece by piece. The mother wavelets 𝛽x𝜑 are generated by mother wavelet translation and scaling in the
mother wavelet at different pixel positions on the two-dimensional plane and are depicted on the Eq. 2.
1 𝜑 s
𝛽x𝜑 (𝜏, s) = 𝛽 (3 (x − I), 3s (y − j) (2)
s2 x
The mother wavelet transformation is called the Berkeles mother wavelet, where μ and s are translation and scale
parameters of wavelet transformation, and 𝛽x𝜑 is the function.
The Berkeley wavelet transform is employed to decompose the input time series data into multiple frequency
components. This decomposition helps in isolating the underlying patterns in the data at different scales, which is
crucial for capturing both short-term and long-term dependencies in the EV charging demand. The wavelet transform
preserves the inherent characteristics of the data by maintaining the time-domain information, which is essential
for accurate temporal analysis.
Integration with Enhanced Gated Recurrent Units (GRUs):
After decomposing the input data using BWT, the resulting wavelet coefficients are fed into the enhanced gated
recurrent units. These GRUs are designed to capture the temporal dependencies in the decomposed data effectively.
The enhanced GRUs utilize an optimized architecture to handle the multi-scale temporal features extracted from
the wavelet coefficients, ensuring that both short-term fluctuations and long-term trends are accurately modeled.
Interaction with the Graph Convolutional Network (GCN):

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

The decomposed wavelet coefficients are also processed by the graph convolutional network, which leverages both
adjacency and adaptive graphs to represent the spatial correlations among nodes in the EV network.
The GCN extracts multi-level spatial correlations through stacked fusion graph convolutional elements. The integration
of wavelet-transformed data enhances the GCN’s ability to capture intricate spatial relationships by providing a detailed
multi-scale representation of the input time series.
Impact on Model Performance and Accuracy:
The incorporation of BWT enhances the model’s ability to capture complex patterns in the EV charging demand data,
leading to improved predictive performance. By decomposing the data into different frequency components, the model
can better isolate and analyze variations at multiple scales.
The residual connection units in the model facilitate the fusion of extracted spatiotemporal features with the original
data, further enhancing the accuracy of the predictions.
Experimental results demonstrate that the integration of BWT contributes to a significant improvement in the model’s
prediction metrics compared to existing methods. This highlights the effectiveness of wavelet transformation in preserv-
ing data characteristics and enabling robust spatiotemporal feature extraction.

4.3 Classification

In addition to the input and output layers, the FGCN EGRU model comes with a stacked graph convolution module, a
residual connection unit, an enhanced gated recurrent unit, and a fusion graph creation module. In order to represent
the complicated spatial–temporal properties, the input layer linearly translates the dimensional space. By merging the
adaptive and adjacency graphs, the fusion graph creation module creates a fusion graph. In order to extract multilevel
spatial characteristics, a stacked graph convolutional module is used with the fusion graph and the enhanced data. Uti-
lising the supplemented data and spatial characteristics extracted from stacked graph convolutional, the information.
The spatial–temporal features are combined with the features that were directly derived from the initial input data by
the residual connection unit. The last step in the prediction process is the output layer, which uses the characteristics
fused by the unit to generate forecasts.

4.3.1 Fusion graph convolution

With careful network graph structure construction, networks are able to better capture the extensive spatial correlation
info seen in EV data. A insufficient to adequately depict the interactions between car nodes in an EV network due to
the fact that these relationships are affected by numerous factors. While some methods incorporate adaptive graphs to
uncover hidden links between nodes, these approaches frequently overlook the adaptive graph’s potential as a source
of direct spatial information and neglect to properly combine it with the predetermined network structure. The research
suggests a fusion graph convolutional network (FGCN) to refine the depiction of node interactions and improve the
extraction of spatial correlations among EV data. To start with, it improves the representation of node interactions by
building a fusion graph structure. The fusion graph’s structure and the optimisation that was suggested are then used
to conduct graph convolution actions in order to improve the extraction of spatial correlations.
The adjacency graph G1 = (V , E s , As ) is built based on the EV prediction dataset. The contiguity matrix Aadj of the
adjacency graph G1 is uttered as formula (3).
{ ( ) (( ) )
1, vi , vj ∈ V 𝛿 vi , vj ∈ E s
Aadj = (3)
0, otherwise
( )
The adaptive graph G2 = V , E adp , Aadp is artificially created by the adaptive network learning layer using data. An
adaptive contains two matrices of learnable parameters. P1 ∈ RN×d and P2 ∈ Rd×N that are produced, with N representing
the total sum of nodes in the EV network and d signifying the dimension of matrix. After that, P1 and P2 are increased to
infer the spatial correlation among matrix Acor of adaptive graph G2 is made as formula (4).
( ( ))
Acor = softmax ReLU P1 .P2 (4)

To solve for (4), we normalise the matrix using the softmax

( activation
) function and remove the negative weight among
nodes using the ReLU function. The fusion graph G3 = V , E f , Af is gotten by combination graph and the contiguous-
ness graph. As exposed in formulation (5), matrix Afus of designed by totaling the adjacency Aadj besides matrices Acor.

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

Afus = Aadj + Acor (5)

Adaptive graph matrices are updated continuously throughout training according to the loss function, allowing for the
mining and quantification of dynamic correlation between nodes from data streams. From the static EV network topol-
ogy, the fusion graph may extract the relationship by mixing graph. Accordingly, for any particular network topology,
the fusion graph provides a more accurate depiction of the geographic correlation of EV status across time. By employ-
ing a spectral network [28] to correlation information from EV data, the approach is able to fully utilise the fused graph
structure’s rich representation of node interactions. The Laplacian matrix can be decomposed into L = U ∧ U T , Λ U made
up of eigenvectors of L, and Λ matrix that contains the eigenvalues of L. The convolution process is exposed in Eq. (6)

gc(L, x) = g𝜃 (L)x = Ug𝜃 (Λ)U T x (6)

In this context, gc stands for the kernel, x is the input data, and g_θ (Λ) is the convolution kernel generated from the
spectral domain by Laplacian matrix feature decomposition. The training convergence of the perfect is impacted by
the high computational cost of calculating the convolution domain using Laplacian matrix decomposition, which is
particularly problematic in EV flow prediction tasks due to the very large graph structure of realistic networks. Hence,
the RKOA was employed in the study to resolve the order K.

4.3.1.1 The basics of the red kite optimization procedure Inspired by the social lives of red kites, Gahruei et al. [29] pro-
posed the red kite optimisation algorithm (RKOA), a unique metaheuristic technique. Red kites typically set up their
nests in forested regions and near bodies of water, where they may catch prey. In their communal life, they move at
random, are influenced by one another’s places when flying, and hunt at fast speeds. During moments such as birth,
migration, the discovery of a suitable bait, or water sources, they produce a sound that is referred to as the sound of
oneness. The sound of danger can also describe noises made during dangerous situations, such as an enemy attack, an
earthquake, a storm, or the death of another animal. A red kite’s foraging behaviour can be modelled by defining each
bird according to its current location, evaluation function value, amount of point displacement, danger sound (facing
the individual component), unity sound (facing the social component), new function. Metaheuristic algorithms rely on
efficient navigation of the issue search space to avoid local optima and get good outcomes. In subsequent rounds, it
uses the optimal solution after a slow transition from the exploration to exploitation phase. The three primary steps of
RKOA are outlined below.:
The first stage the birds’ starting positions: At this point, we may randomly initialise the red kites’ positions according
to Eq. (7) as,
Posi,j (t) = lb + rand × (ub − lb), i = 1, 2, … , n and j = 1, 2, … , d (7)

where Posi,j (t) is the position of the ith red kite at repetition t, where lb and ub are borders, n is the size of the populace,
d is the issue dimension, besides rand is sum between 0 and 1.
The second stage leadership position selection: Eq. (8) is used to determine the leader:

�����������⃗ = Pos
Best(t) ����������⃗
i (t)iffi (t) < fbest (t) (8)

where Best(t) signifies site of iteration t, Posi (t) denotes the site of ith red kite in repetition t, fi (t) is charge of the bird
assessment purpose in repetition t, besides fbest (t) is the charge of the assessment purpose of the finest bird in repetition t.
The third stage Bird migration: Red kites, according to Eq. (9), are thought to transition from the exploration to the
exploitation stage of their life cycle in a progressive fashion as the coefficient (D) decreases.
( ( ) )−10
t t
D= exp − (9)
t_max t_max

where t_max is the extreme iteration. Equations (10 and 11) allow the birds to update their positions:

����������������������⃗
pos new ����������⃗
(t + 1) = Pos ����������������⃗ (10)
i i (t) + Pmi (t + 1)

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

( )
P����������������
mi (t + ⃗
1) = D(t) × ���������
P mi
⃗
(t) + ��������
SC(t) ⃗ ⊙ ��������������
Pos rws
⃗
(t) − ����������⃗
Pos i (t)
( ) (11)
�����������⃗ − Pos
��������⃗ ⊙ Best(t)
+ uc(t) ����������⃗
i (t)

where Posrws (t) is the bird position iteration t, posinew (t + 1) where UC is a random vector of individual components and
SC is the new location of the bird. After making position updates, it is necessary to verify the limits of the search space;
Eq. (12) can be used for this purpose.,
( ( ) )
����������������������⃗
pos new ����������������������⃗
(t + 1) = max min pos new
(t + 1) + ub , lb (12)
i i

Any changes made to the evaluation function will result in the new temporary position being substituted. For this reason,
����������������������⃗
Posi (t + 1) is equal to pos new
(t + 1). Each bird’s voice of unity and danger is represented by SC and UC, which are separate
i
components, correspondingly. They are produced according to the following connection.
{
���������������
SC(t + 1)⃗ = r��⃗1
if rand ≤ 0.5
����������������
UC(t + 1)⃗ = r��⃗2

{
���������������
SC(t + 1)⃗ = r��⃗3
����������������
Otherwise (13)
UC(t + 1)⃗ = r��⃗1

where r��⃗1 is a accidental course in [1, 2], r��⃗2 is a chance course in [1, 3], and r��⃗3 is a accidental course in [0, 1]. Based on each
bird’s present position, the RKOA uses a roulette greatest solution identified so far to randomly choose a neighbor’s
position. During the initial rounds of exploration and search, the value of D(t) is almost one. The red kite discovers new
areas by comparing its location with that of a neighbour, chosen at random, in the component-based movement. Addi-
tionally, the algorithm is guided to the global optimum by the social component. The coefficient D(t) gradually lowers
to establish balance among the phases as the algorithm proceeds from the early the acquired decent solutions in the
last rounds, and this coefficient approaches to zero. With minimal governing factors.
The convolution process after the RKOA is exposed in Eq. (14)
K −1 K −1
∑ ( ) T ∑ ( )
gc(L, x) = U ̃ U x=
𝜃k T k Λ 𝜃k Tk L̃ x (14)
k=0 k=0

By skipping feature decomposition and performing convolution directly on the Laplacian matrix, the computational cost
is reduced in the fusion graph convolutional neural network. Further, fusion graph-based graph convolutional networks are
better able to use node-to-node interactions, leading to improved spatial correlation data capture.

Fig. 3  Structure of enhanced GRU​

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

4.3.2 Stacked fusion graph convolution

There is a great variety of geographical relationships among road network data, and the EV network has a complicated rela-
tionship with the roads that connect to it. When congestion happens in one part of the network, it impacts the other parts
nearby as well as the network conditions farther away. This proves that EV data has both local and distant spatial dependen-
cies. Improving the accuracy of EV flow forecast relies on a more precise way to represent the network data’s multi-level spatial
dependence connection. Because our fusion network is limited to K-order neighbour info, it struggles to mine network data
for long-range dependencies and becomes less effective at extracting local features as K rises.
A convolution module with stacked fusion graphs is built to capture the dependence. In order to extract more advanced
spatial correlation information, this module combines three convolutional layers of a network and updates and passes on
node attributes layer by layer. One possible representation of the stacked graph convolution unit is formula (15).
( ( ( )))
Xmgc = gc L, ̃ gc L,
̃ gc L,
̃ x (15)

4.3.3 Enhanced gated recurrent unit

Both the spatial dimensions of EV network data are highly correlated with one another. Consider how the condition
of a vehicle at one instant can impact traffic conditions at several future times in addition to the next neighbour-
ing moment [30, 31]. Accurate flow prediction requires more than just networks to capture the correlation among
flow data. Consequently, the research presents a new method for extracting spatiotemporal correlations: recurrent
neural networks integrated convolution. But classic recurrent neural networks have problems with disappearing
or expanding gradients, early relevant information tends to be forgotten. Because of this, the network struggles to
collect traffic flow data that exhibits long-term temporal correlations.
Our goal in developing the improved gated recurrent unit (EGRU) was to improve the data collecting process for
traffic flow correlation information. The gating apparatus effectively controls the flow of information, reducing the
impact of gradient explosion. Also, in order to keep the early correlation information from getting lost throughout
the time-step propagation, an improved attention method is employed to focus and update it. The enhanced unit’s
construction is portrayed in Fig. 3.
The improved gated recurrent unit takes in data from Ht−1, Xmgc , XT ′ , where, Ht−1 represents the state that is con-
cealed at time t-1 and contains the info of the spatio-temporal characteristics that were retrieved at time t−1. Xmgc
is convolution. XT ′ represents the state that is concealed at time t−1 and contains the info of the spatio-temporal
characteristics that were retrieved at time Xmgc , enabling the subsequent state to retain a greater amount of pertinent
data from the preceding state. Combining X_mgc and which results in the formation of, relies heavily on the atten-
tion scores. Xst , using which the update gate z is later computed. The amount that the prior state affects the present
state is controlled by this update gate.
The exact computation procedure for manipulative the care scores to merge Xmgc and Ht−1 into Xst, which contains
rich spatiotemporal features, is exposed as shadow:
Splice the output Xmgc of the loaded graph u with national Ht−1 to find M as Eq. (16).
M = Xmgc ||Ht−1 (16)

The attention scores of Xmgc and Ht−1 are intended by means of the softmax purpose as shadows:
P = softmax(M)
(17)
Ps , Pt = split(P)

where, split() is the operation, Ps and Pt are the care scores media of Xmgc and Ht−1. Combining Xmgc and Ht−1 using the
consideration scores get Xst as formula (18).
Xst = Ps .Xmgc + Pt .Ht−1 (18)

The Xst uses the EV network data to combine geographical and temporal characteristics. When it’s used to calcu-
late the update characteristics from the old state may be transferred to the new one, which means less important

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

Table 3  Performance metrics Datasets Presentation metrics

with proposed predictor
perfect MAE MSE RMSE Apre

Georgia Tech EV accusing station, USA 0.1083 4.24413 × 10−10 2.0523 × 10−5 0.9712

Table 4  Assessed MSE charge Training Testing

over Epochs
Epochs Mean square error Mean square error

10 4.1230 5.6234
50 1.0173 3.8187
100 0.1942 0.5987
150 3.0654 × ­10−3 7.4234 × ­10−3
200 7.1763 × ­10−5 9.8964 × ­10−5
250 2.0998 × ­10−7 5.1674 × ­10−7
251 4.17716 × ­10−10 5.86534 × ­10−10

data gets lost when the gate is updated. The procedure of scheming the gate unit using Xmgc and Xst , and informing
national with XT ′ , is as shadows.
( )
r = 𝜎 Wr .Xmgc + Ur .Ht−1
( )
z = 𝜎 Wz .Xst + Uz .Ht−1
( ) (19)
h̃ = tanh W .X � + r.U H
h T h (t−1)

Ht = z ∗ H(t−1) + (1 − z) ∗ h̃

The expression is defined as follows: h stands for the candidate hidden unit, H_t for the hidden state at the present
instant, σ for the function, and W and U for understandable weight parameters. Equipped with the initial units, the aug-
mented gated recurrent unit acquires spatio-temporal properties. It uses a new method to calculate the gate mechanism,
which successfully sidesteps gradient problems and lessens correlation information loss during period step propagation.
With the elimination of bias terms in the update algorithm, the network further reduces the number of parameters.

4.3.4 Residual connection unit

Prediction accuracy will suffer if the augmented gated recurrent unit’s spatio-temporal properties are relied upon in
isolation from the direct usable info included in the initial input data. In addition, the model’s performance degrades
and issues like gradient dispersion arise, leading to low prediction accuracy, when the number of activation functions
added increases due to the model’s many network stacks.
The research presents an outstanding joining unit to merge the original input data acquired spatio- characteristics,
which enhances feature utilisation and addresses gradient vanishing [32]. The unique input data X_T are the two input
branches that make up the remaining connection unit. Direct features are extracted from the initial input data by use of
a linear layer in the residual connection’s unit. After that, the spatio-temporal features are combined with these direct
features. The last steps involve applying the function and LayerNorm to the combined features. Future traffic flow data
can be more accurately predicted by using these fused features. Another way to circumvent the disappearing gradient
problem during backpropagation is to use residual connections, which allow gradients to connections.
Formula (20) can be used to define the residual connection unit.
( ( ( ) ( )))
Xrc = LayerNorm relu Linear XT + rele Ht (20)

The study uses the layer to change the output from the residual linking unit X_rc into the last forecast result. Xp . The
cunning process is exposed in Eq. (21).

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

Table 5  Metric performance Methods MAE ME MSE RMSE

of projected methods
ELM 0.1533 0.0866 0.0122 0.1105
XGBoost 0.1455 0.1374 0.0289 0.1700
MLP 0.3371 0.2854 0.0038 0.0619
AE 0.1896 0.1478 0.0207 0.1437
DBN 0.2990 0.2545 0.0647 0.2545
RNN 0.0801 0.0669 0.0014 0.0368
CNN 0.0801 0.0669 0.0014 0.0368
Graph CNN 0.0828 0.0709 0.0050 0.0709
GRU​ 0.1250 0.1341 0.0279 0.1670
Enhanced GRU​ 0.0109 0.0087 0.0002 0.0156
Proposed model 0.3228 0.2369 0.0301 0.1735

Table 6  Comparative analysis Model Training/calibration data Test/validation data

of numerous models
R2 MAE (N) RMSE (N) R2 MAE (N) RMSE (N)

Proposed 0.986 15.176 20.458 0.984 15.792 21.520

GRU​ 0.976 13.738 19.844 0.988 14.127 18.834
GCNN 0.938 35.263 42.501 0.932 36.506 44.321
CNN 0.958 25.804 35.278 0.957 25.685 34.412
RNN 0.946 31.899 39.586 0.946 31.920 39.212
DBN 0.896 44.375 54.967 0.886 48.624 58.081

Xp = Xrc Wo + b𝜎 (21)

where Wo and bo are weight matrices.

5 Results and discussions

At USA, we put the newly created predictive model to the test to see how well it can forecast the energy ingesting for
electric vehicle charging stations. An Inteldualcorei5CPU with 8 GB of physical memory is used to conduct the simulation
in the MATLAB R2021a environment.

5.1 Performance metrics

The proposed predictor perfect is verified for its advantage based on the presentation metrics as given—mean abso-
lute error (MAE), mean square error (MSE), root mean square error (RMSE) and accuracy of prediction ( Apre) and they are
evaluated with the subsequent equations,
N
1 ∑| |
MAE = |Y − Yactual_q | (22)
N q=1 | predicted_q |

N
1 ∑( )2
MSE = Ypredicted_q − Yactual_q (23)
N q=1

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

√
√ N
√1 ∑( )2
RMSE = √ Y − Yactual_q (24)
N q=1 predicted_q

N
{ ( )( )
1∑ 1, if Ypredictedq+1 − Yactual_q Ypredictedq+1 − Yactual_q > 0
Apre = Y ,Y = (25)
N q=1 q q 0, Otherwise

In Eq. (25), ‘N’ characterizes the total sum of data tasters, ‘Yactual ’ signifies the unique EV data and ‘Ypredicted ’ is the foretold
output gotten using the future predictor perfect.

5.2 Validation analysis of proposed perfect

The performance metrics with the proposed predictor are described as being perfect in Table 3. During the investigation
of vehicle charging station at Georgia Tech in the United States, the MAE rate was found to be 0.1083, the mean squared
error (MSE) range was found to be 4.24413 × ­10−10 and the root mean squared error (RMSE) was found to be 2.0523 × ­10−5.
Lastly, the corresponding A_pre zx value was found to be 0.9712.
In Table 4 signifies that the Assessed MSE charge over Epochs. In the analysis of training section, the Mean Square
Error at 10 epochs as 4.1230 and testing section, the MSE at 10 epochs as 5.6234 correspondingly. Then the 50 epochs
MSE as 1.0173 and testing section, the Mean Square Error at 50 epochs as 3.8187 correspondingly. Then the 100 epochs
MSE as 0.1942 and testing section, the Mean Square Error at 100 epochs as 0.5987 correspondingly. Then the 150 epochs
MSE as 3.0654 × ­10−3 150 and testing section, the Mean Square Error at 150 epochs as 7.4234 × ­10−3 correspondingly.
Then the 200 epochs MSE as 7.1763 × ­10−5 200 and testing section, the Mean Square Error at 200epochs as 9.8964 × ­10−5
correspondingly. Then the 250 epochs MSE as 2.0998 × ­10−7 and testing section, the Mean Square Error at 10 epochs as
5.1674 × ­10−7 correspondingly. Then the 251 epochs MSE as 4.17716 × ­10−10 and testing section, the Mean Square Error
at 251 epochs as 5.86534 × ­10−10 correspondingly. In the comparisons analysis, we compared different existing algo-
rithms, the proposed model attained better results than other existing methods. The use of enhanced GRUs allows for
better capture of multi-scale temporal dependencies compared to traditional RNNs or GRUs. This improvement facilitates
accurate modeling of both short-term fluctuations and long-term trends in EV charging demand. he model is evaluated
using real EV charging data from Georgia Tech, demonstrating its practical effectiveness and ability to produce accurate
forecasts in real-world scenarios.
In Table 5 characterised Metric performance of projected methods. In the investigation of ELM 0.1533 model attained
the MAE rate as 0.0866 model attained the ME rate as 0.0122 model attained the MSE rate as 0.1105 congruently. Then the
XGBoost model accomplished the MAE rate as 0.14550. and ME rate as 1374 and MSE value of 0.0289 and RMSE value as
0.1700 correspondingly. Then the MLP perfect accomplished the MAE rate as 0.3371 and ME degree as 0.2854 and MSE
rate as 0.0038 and RMSE value as 0.0619 congruently. Then the AE model attained the MAE rate as 0.1896 and ME rate
as 0.1478 and MSE rate as 0.0207 and RMSE value as 0.1437 congruently. Then the DBN model attained the MAE rate as
0.2990 model attained the MAE rate as 0.2545 and MSE rate as 0.0647 and RMSE value as 0.2545 correspondingly. Then
the RNN model attained the MAE rate as 0.0801 model attained the MAE rate as 0.0669 and MSE rate as 0.0014 and RMSE
value as 0.0368 correspondingly. Then the CNN model attained the MAE rate as 0.0801 and ME rate as 0.0669 and ME
rate as 0.0014 and RMSE value as 0.0368 congruently. Then the Graph CNN perfect attained the MAE rate as 0.0828 and
ME rate as 0.0709 and MSE rate as 0.0050 and RMSE value as 0.0709 correspondingly. Then the GRU 0.1250 0.1341 and
ME rate as 0.0279 and RMSE value as 0.1670 congruently. Then the Enhanced GRU 0.0109 0.0087 and ME rate as 0.0002
and RMSE value as 0.0156 congruently. Then the Proposed model 0.3228 model accomplished the MAE rate as 0.2369
model accomplished the MAE rate as 0.0301 and RMSE charge as 0.1735 correspondingly. In the comparisons analysis,
we compared different existing algorithms, the proposed model attained better results than other existing methods.
Various models’ comparative analyses are characterised in Table 6. The proposed perfect accomplished an R2 worth
of 0.986, an MAE range of 15.176, and an RMSE rate of 20.458 in the training/calibration data analysis. The test/validation
model produced an R2 value of 0.984, an MAE range of 15.792, and an RMSE rate of 21.520 in the training/calibration
data analysis. Following that, the GRU model achieved an R2 value of 0.976, which is equal to 13.738, and the Training/
Calibration Data Test/Validation model achieved an ­R2 value of 9.844.The R2 value was 14.127 and the MAE range was
18.834 with a 0.988.Following this, the GCNN model achieved an R ­ 2 value of 0.938, the Training/Calibration Data Test/

Vol:.(1234567890)
 Discover Applied Sciences

Table 7  Performance of proposed model on different environments

Scenario Evaluation metric Results (Example Values) Comments
(2024) 6:605

Normal demand periods RMSE, MAE RMSE: 0.14, MAE: 0.12 Low errors during predictable, regular demand periods
Peak demand periods RMSE, MAE RMSE: 0.23, MAE: 0.27 Higher errors due to demand spikes, but still manageable
Urban areas Accuracy, MAE Accuracy: 93%, MAE: 0.07 High accuracy in densely populated areas with more EVs
Rural areas Accuracy, MAE Accuracy: 89%, MAE: 0.13 Slightly lower accuracy due to unpredictable EV usage in rural settings
With weather/traffic data MAE Without: 0.14, With: 0.13 Significant improvement in prediction when incorporating external data
Unexpected events (outages) RMSE RMSE: 0.37 Higher error during rare events, highlighting areas for improvement
Small charging networks MAE MAE: 0.19 Good performance with a limited number of stations in small networks
Large charging networks MAE MAE: 0.13 Slight increase in error due to the complexity of large spatial networks
Short-term forecasts (next day) RMSE RMSE: 0.07 Very low error for immediate future demand predictions
Long-term forecasts (next month) RMSE RMSE: 0.25 Higher error as long-term demand is harder to predict accurately
Regular charging patterns Accuracy, MAE Accuracy: 97%, MAE: 0.07 High accuracy in cities where charging patterns are predictable
Irregular charging patterns Accuracy, MAE Accuracy: 92%, MAE: 0.23 Lower accuracy in areas with unpredictable or fluctuating charging patterns
Graph noise (node removal) Accuracy, MAE Accuracy: 89%, MAE: 0.17 Slight drop in accuracy when charging stations or relationships are removed
Rkoa tuning RMSE (Before/After Tuning) Before: 0.14, After: 0.15 Performance improves significantly after hyperparameter tuning with RKOA
| https://doi.org/10.1007/s42452-024-06326-x

Vol.:(0123456789)
Research
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

Validation model an R2 value of 35.263, an MAE range of 0.932, and a 42.501 overall. In the comparisons analysis, we
compared different existing algorithms, the proposed model attained better results than other existing methods. In
addition, the performance of the proposed model is tested with some simulated environments in terms of accuracy and
MAE that is shown in Table 7.

5.3 Discussion

Based on the experimental findings shown above, the suggested model performs well in predicting the stations in the
long run, with no degradation in prediction accuracy over time. Considering time intervals more than 30 min is seen as
long- for addressing concerns related to transportation facility usage forecast and traffic limit forecast. As a challenge
to prediction algorithms, the longest interval in this paper is 90 min. Charging stations could provide stable service to
electric vehicle drivers relative to drivers of gasoline-powered vehicles, since there are fewer EVs and drivers of electric
vehicles typically charge their vehicles for longer periods of time than they refuel. This is a key reason why another
existing algorithm also shows little loss over time. On the other hand, by utilising both spatial and temporal data, the
suggested model is able to optimally learn the pattern of charging behaviour influenced by the distribution of urban
charging stations and habits of recurrent users at different stations.

6 Conclusions

This study presents a novel model designed to forecast charging demand by leveraging optimized fusion graph convo-
lution units, specifically tailored for EV datasets. The model utilizes a fusion network and an upgraded gated recurrent
prediction mechanism to extract both spatial and temporal features, which are then fused using residual connection
units. By employing the Red Kite Optimization Algorithm (RKOA), the model effectively reduces computational complex-
ity to identify the optimal K value for the fusion network. Through extensive simulations conducted on EV datasets, the
proposed predictor consistently outperformed existing prediction models, exhibiting superior performance with an
impressive evaluation efficiency of 98%. This surpasses the capabilities of previous state-of-the-art methods, affirming
its practical utility. With enhanced prediction accuracy and minimized error margins, the Fusion Graph Convolutional
Neural Network (FGCNN) and Enhanced Gated Recurrent Unit (EGRU) forecaster demonstrate effectiveness in meeting
the demands of electric vehicle infrastructure.
To further enhance the utilization of charging stations and align predictions more closely with real-world scenarios,
future research could explore the integration of this refined model with other intelligent operational frameworks. Moreo-
ver, operators stand to benefit from employing this model in conjunction with algorithms aimed at optimizing charging
station placement and sizing, thereby enhancing overall operational efficiency.

Acknowledgements Not applicable.

Author contributions Dr. R. Gunasekaran: Supervision, Visualization, Data curation, Formal analysis, Validation. Dr Manjunatha B.: Data curation,
Supervision, Formal analysis, Writing of the original draft. Anand S: Visualization, Data curation, Formal analysis. Dr. Piyush Kumar Pareek: Data
curation, Formal analysis, Validation. Dr. Sandeep Gupta: Conceptualization, Visualization, Data curation, Investigation, Formal analysis, and
validation. Dr. Anand Shukla: Conceptualization, Visualization, Data curation, Supervision, Formal analysis, Validation.

Funding No funding was received.

Data Availability Data availability: The datasets used and/or analyzed during the current study available from the corresponding author on
reasonable request.

Declarations
Competing interests The authors declare no competing interests.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which
permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to
the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You
do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party
material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If

Vol:.(1234567890)
Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x Research

material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds
the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://​creat​iveco​
mmons.​org/​licen​ses/​by-​nc-​nd/4.​0/.

References
1. Luo R, Zhang Y, Zhou Y, Chen H, Yang L, Yang J, Su R, Deep learning approach for long-term prediction of electric vehicle (ev) charging station
availability. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC). 2021;3334–3339.
2. Shahriar S, Al-Ali AR, Osman AH, Dhou S, Nijim M. Prediction of EV charging behavior using machine learning. Ieee Access. 2021;9:111576–86.
3. Bai X, Wang Z, Zou L, Liu H, Sun Q, Alsaadi FE. Electric vehicle charging station planning with dynamic prediction of elastic charging demand:
a hybrid particle swarm optimization algorithm. Compl Intell Syst. 2022. https://​doi.​org/​10.​1007/​s40747-​021-​00575-8.
4. Ma TY, Faye S. Multistep electric vehicle charging station occupancy prediction using hybrid LSTM neural networks. Energy. 2022;244:123217.
5. Hecht C, Figgener J, Sauer DU. Predicting electric vehicle charging station availability using ensemble machine learning. Energies.
2021;14(23):7834.
6. He C, Zhu J, Lan J, Li S, Wu W, Zhu H. Optimal planning of electric vehicle battery centralized charging station based on EV load forecasting.
IEEE Trans Ind Appl. 2022;58(5):6557–75.
7. Ma TY, Faye S, Multistep electric vehicle charging station occupancy prediction using mixed LSTM neural networks. 2021. arXiv preprint arXiv:​
2106.​04986.
8. Andrenacci N, Valentini MP. A literature review on the charging behaviour of private electric vehicles. Appl Sci. 2023;13(23):12877.
9. Anand S, Kulkarni V, Gupta S, Parthasarathy V, Battery management system (BMS) for electric vehicle applications. In: 2022 fourth international
conference on emerging research in electronics, computer science and technology (ICERECT), pp. 1–6. IEEE, 26–27 December 2022.
10. Gupta S, Tripathi RK, FACTS modelling and control: application of CSC based STATCOM in transmission line. In: 2012 IEEE students conference
on engineering and systems. 2012;1–5.
11. Yi Z, Liu XC, Wei R, Chen X, Dai J. Electric vehicle charging demand forecasting using deep learning model. J Intell Transp Syst.
2022;26(6):690–703.
12. Kalakanti AK, Rao S. Charging station planning for electric vehicles. Systems. 2022;10(1):6.
13. Liu J, Lin G, Rehtanz C, Huang S, Zhou Y, Li Y. Data-driven intelligent EV charging operating with limited chargers considering the charging
demand forecasting. Int J Electr Power Energy Syst. 2022;141:108218.
14. Shanmuganathan J, Victoire AA, Balraj G, Victoire A. Deep learning LSTM recurrent neural network model for prediction of electric vehicle
charging demand. Sustainability. 2022;14(16):10207.
15. Zhou D, Guo Z, Xie Y, Hu Y, Jiang D, Feng Y, Liu D. Using bayesian deep learning for electric vehicle charging station load forecasting. Energies.
2022;15(17):6195.
16. Chakraborty P, Pal M. Planning of fast charging infrastructure for electric vehicles in a distribution system and prediction of dynamic price. Int
J Electr Power Energy Syst. 2024;155:109502.
17. Jiang Q, Zhang N, He BY, Lee C, Ma J. Large-scale public charging demand prediction with a scenario-and activity-based approach. Transp Res
Part A Policy Pract. 2024;179:103935.
18. Williams B, Bishop D, Hooper G, Chase JG. Driving change: Electric vehicle charging behavior and peak loading. Renew Sustain Energy Rev.
2024;189:113953.
19. Das K, Kumar R, Krishna A. Analyzing electric vehicle battery health performance using supervised machine learning. Renew Sustain Energy
Rev. 2024;189:113967.
20. Zhao Z, Lee CK, Ren J. A two-level charging scheduling method for public electric vehicle charging stations considering heterogeneous
demand and nonlinear charging profile. Appl Energy. 2024;355:122278.
21. Azzopardi B, Gabdullin Y. Impacts of electric vehicles charging in low-voltage distribution networks: a case study in Malta. Energies.
2024;17(2):289.
22. Ullah I, Liu K, Yamamoto T, Shafiullah M, Jamal A. Grey wolf optimizer-based machine learning algorithm to predict electric vehicle charging
duration time. Transp Lett. 2023;15(8):889–906.
23. Koohfar S, Woldemariam W, Kumar A. Prediction of electric vehicles charging demand: a transformer-based deep learning approach. Sustain-
ability. 2023;15(3):2105.
24. Aduama P, Zhang Z, Al-Sumaiti AS. Multi-feature data fusion-based load forecasting of electric vehicle charging stations using a deep learning
model. Energies. 2023;16(3):1309.
25. Zhao J, Ling H, Liu J, Wang J, Burke AF, Lian Y. Machine learning for predicting battery capacity for electric vehicles. ETransportation.
2023;15:100214.
26. Campus Electric Vehicle Charging Stations Behavior. Available online: www.​kaggle.​com/​datas​ets/​clayt​onmil​ler/​campu​selec​tric-​vehic​le-​charg​
ing-​stati​ons-​behav​ior/​metad​ata. Accessed on 19 Aug 2023.
27. Gupta S, Poonia S, et al. Design and implementation of the electric bicycle with efficient controller. In Intelligent Computing Techniques for
Smart Energy Systems, Lecture Notes in Electrical Engineering, Springer. 2022;862:541–52.
28. Kipf TN, Welling M, Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:​1609.​02907 2016.
29. Gahruei JR, Beheshti Z. The electricity consumption prediction using hybrid red kite optimization algorithm with multi-layer perceptron neural
network. J Intell Proced Electr Technol. 2022;15:1–22.
30. Farah S, Humaira N, Aneela Z, Steffen E. Short-term multi-hour ahead country-wide wind power prediction for Germany using gated recurrent
unit deep learning. Renew Sustain Energy Rev. 2022;167:112700.
31. Bhardwaj R, Gupta S. Evolutionary progress of the electric car market with future directions. Latest Trends Renew Energy Technol Lecture
Notes Electr Eng Springer. 2021;760:315–21.

Vol.:(0123456789)
Research Discover Applied Sciences (2024) 6:605 | https://doi.org/10.1007/s42452-024-06326-x

32. Andrienko N, Gennady A, Peter G, Impact of data and task characteristics on design of spatio-temporal data visualization tools. In: Exploring
geovisualization. Elsevier. 2005;201–222.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Vol:.(1234567890)