Predicting and Forecasting of Vehicle Charging Station Using ECNN with DHFO Algorithm

Abstract

The forecast of the optimal placement of a charging station (CS) according to the real-time consumption of electric vehicles is a subject of urgency in this new era. The demand of a charging station in an area based on the trend of consumption can be predicted by means of interpolation and the extrapolation of historical data using a linear function of prediction model. The prediction of the charging station system was performed with distance relevancy methods. An adaptive optimal learning model was proposed to enhance the prediction performance for charging station management and to represent the pattern of vehicles’ travelling directions. The proposed model uses Distributional Homogeneity Feature Optimization (DHFO) using artificial intelligence (AI) to categorize and forecast the charging station from the database. The prediction performance of this model is improved more than the conventional classification model by filtering the apt features from all the electric vehicular and charging station attributes in the database. The Enhanced Cladistic Neural Network (ECNN) is used to improve the pattern learning model and increase learning accuracy. By comparing statistical parameters with other state-of-the-art methodologies, the suggested model’s overall findings were verified.

1. Introduction

Faster electric vehicle (EV) adoption all over the world aims to achieve the decarbonation of the entire planet, thereby making Earth a safer and more sustainable place to live. Various research has been conducted all over the world to examine the biggest challenges faced by the automobile industry during this transition phase. The range anxiety is the greatest hindrance for all EV users. The uncertainty of running out of battery power during journeys is the predominant barrier for this quantum change. A myriad of conditional factors were reviewed depending on environmental, economic, social, and technical aspects to resolve this situation. The installation of power supplies at apt locations will reduce the fear of EV users of running out of power. This paved the way for studies and experiments to be carried out on the prediction of EV users’ behavior and the optimal placement of charging stations.

Numerous systems are available to forecast the way an electric vehicle will approach a charging station by monitoring the State of the Charge (SOC) of the vehicle [1]. The SOC of the vehicle depends on several environmental and vehicle parameters. It gives an indication of the maximum range that can be covered by a vehicle and helps it to recognize when it is in urgent need of charge. Machine learning algorithms can carry out the forecast process to identify available charging stations using vehicle historical data. The charging stations may not be the most optimal ones as there can be multiple issues while tracking the vehicle’s range, traffic conditions, and maximum charging capacity of the charging station, as discussed in [2]. This causes the charging station prediction and other vehicle routing processes to be misclassified. To get around that, the charging station prediction was followed by a historical data reference for relevancy. Apart from that, the lack of sufficient numbers of charging stations is a primary concern hindering the growth of EV adoption. This has triggered the necessity of installing new charging stations. Hence, the identification of the most suitable location for the same is a challenging task.

The Multi-Objective Optimal Placement model is ideally based on the above perceptions to develop an optimal charging station prediction for EV users. The Multi-Objective Optimal Placement model mainly concentrates on three stages—vehicle allocation, vehicle count selection, and end-user vehicle node placement. In this, the vehicle allocation defines the cluster-based coverage area computation of EVs that is in the proximity of the CS. Based on the vehicle distance to the CS and the power requirement, the parameters are validated and assigned to the best CS. The vehicle count selection takes into consideration the traffic conditions that are to be matched against the maximum charging capacity of the CS. This will select the number of vehicles that can be mapped with the available CS by comparing the available power of the CS along with the requirement of the EV to charge. The end-user vehicle node-based optimal placement refers to the parameters of the electric vehicle that will meet the criteria such as State of the Charge (SOC), based on the power demand, distance to travel, availability of terminals, and capacity of CS. Based on these three stages of the optimization model, the vehicle can identify the best place to charge the vehicle with minimum service time.

The proposed system takes into consideration the three stages of the Multi-Objective Optimal Placement model in its design to predict the most optimal charging station from the available ones for an electric vehicle, and forecasts the most suitable location to install a new charging station, taking only the most necessary vehicle parameter values into account. The study in [2] revealed that historical data for an area also play a prominent role in developing an advanced model of feature prediction, and these are adopted in the proposed classification system. To improve the prediction model, an optimization method was used to forecast the data pattern feature analysis [3] from the vehicle data samples and the weight input properties, which led to the selection of the best attributes. To improve the classification performance and increase prediction accuracy, an adequate amount of data samples were taken during the training process for forecasting the power demand as studied in [4].

As in a linear function, a straight line was drawn to predict values outside the dataset—in the proposed system, the route path of the EV from an origin to destination is the standard baseline taken. The prediction of the available charging station system was then performed with the distance relevancy methods across this line. The power demand of the charging station in an area based on the power consumption rate and the vehicle density can be predicted by means of interpolation and the extrapolation of the historical data using the linear function of prediction model. An adaptive optimal learning model was proposed to enhance the performance of prediction of the most apt charging station from the available ones, and the pattern of vehicle movement in the travelling direction was the primary source of data pattern, along with charging station parameters and power demand. This pattern of EV movement and its intersection across the baseline O-D Trip formed the core computation of the prediction. The proposed model used Distributional Homogeneity Feature Optimization (DHFO) and artificial intelligence (AI) to obtain the optimal feature selection for this prediction. The prediction performance of this model was improved more than the conventional classification model by filtering the apt features from all the electric vehicular and charging station attributes in the database. The Enhanced Cladistic Neural Network (ECNN) was then used to improve the pattern learning model and increase learning accuracy to forecast the co-ordinate positions for the installation of a charging station. The pattern analysis for forecasting the co-ordinates of the new charging station mainly concentrated on the power demand of the existing charging stations, along with their availability. By comparing these statistical parameters with other state-of-the-art methodologies, the suggested model’s overall findings have been verified.

The following are the main highlights of the proposed framework:

• Distributional Homogeneity Feature Optimization (DHFO) technique, which is a novel optimal feature selection system for vehicle parameters required for the prediction of an apt location for a charging station for an EV, was developed;
• The classification performance was enhanced by organizing the best feature properties that were pertinent to the data;
• The weight and vehicle type of feature attributes were used to anticipate and predict the range of charging station site forecast based on the state of power demand;
• An Enhanced Cladistic Neural Network (ECNN) classifier was implemented for the classification approach;
• The performance of the proposed classification model was validated based on the statistical parameters, and the comparison was conducted with existing approaches.

The overall organization of the paper is as follows. Section 2 examines the various optimization techniques for the Optimal Placement Model to determine the optimal charging station direction selection. The working model of the suggested ECNN with the DHFO optimization algorithm for the optimal charging station direction selection is described in Section 3. Section 4 provides an explanation of the simulation graph charts and validation of the outcomes of the suggested model, along with the comparison results. Section 5 concludes with a discussion of the paper’s work and potential improvements.

2. Related Works

This section provides a comprehensive overview of the current techniques for forecasting data and predicting charging stations for electric vehicles The challenges and limitations inferred from these studies, which paved the way to scope out further research, have been summarized in Figure 1. Most of the recent techniques mainly give emphasis to socioeconomic and environmental factors, but vehicle density and state of charge play dominant roles and need to be considered. A study of the influence of the social, economic, and environmental factors on the identification of the charging station location was made in [4,5,6,7]. The work [4] proposed a suitable ranking system based on the above influencing factors but lacks the detailing on the technical aspects such as the electric demand and state of charge of the vehicles travelling a particular path that may affect the forecasting of optimal charging station placement. The Charging Station Allocation Data were approximated using the time series and data samples from the Charging Station Time Chart in [8]. In doing so, the production analysis in the basin-scale vehicle level allocation model makes use of the OSM data samples. This also refers to the variation in charging power and the availability forecast based on dataset history. In [9], a gravity-dependent Huff model was described, taking into consideration Sioux-Fall network for illustration. The study’s main focus is from the perspective of a private investor of in a charging station and leverages an optimal pricing strategy. A dynamic pricing strategy based on the Markov decision process and multi-agent neural network was proposed in [10]. The study suggested that charging during off-peak times could be more cost effective. In [11], graph-based computations were used for the maximum comfort and ease of EV users. The schedule time of route flow was predicted in this work using data modeling from short-term availability forecasts. This was operated based on historical data samples for station power and demand characteristics, as well as availability.

Based on the available data and demand forecast, Ref. [12] developed a station availability retrieval model. This analyzed various classification models to compare the efficacy of classification based on statistical variables and to estimate station availability from satellite data. The behavior pattern of EV users for charging was thoroughly analyzed in [13,26,27,28]. The cost of land invested in by the operators was also taken into consideration for predicting the best charging station location. These are all forecasted based on the power consumption and availability factors for the various station availability circumstances, but lack vehicle parameters and SoC detailing. In a similar vein, Ref. [14] put forth a prediction model for the SOC level in the charging station that considered the uncontrolled power loss and the unequal distribution of the available content load. According to this prediction, after the flooding effect, the status of the station will increase the range of SOC > 6.5 to a maximum value of 7. The construction of the charging station at the proper station is improved by this estimation. With this improvement in mind, Ref. [15] suggested a power logging prediction model for the charging station field to supply route flow-level data. This helps to grow the charging station in the field more by estimating the amount of power supply and other associated features based on the power level. A CNN feature-based path planner for EV was studied in [29]. Based on the traffic flow the waiting time to become charged was predicted. This can be used in further studies to track the most suitable CS when the SOC of a vehicle gets below the minimum threshold.

Several additional techniques have improved the construction range and the other decision plan based on the forecasting data and the station parameter prediction process. Based on this, Ref. [30] suggested using an optimization model to estimate the quality of charging stations as a decision support for corridors. This validates the power availability for the public in highways as special segment of highways are corridors dedicated to EV charging. It classified the CS in corridors as better locations than other CSs. The vehicle level quality was determined by the fuzzy interference system, which also segments each division independently. But this study is suitable only for high density traffic areas. Ref. [17] presented the estimation of geographical variance in the field of land used for charging station installation. During power fills, this forecasts the SOC level of the car. To forecast what kind of car-level charging station can develop there, this also assesses the redox condition of the station. Ref. [18] converted double-layer vehicle levels to ratoon charging station fields to lessen carbon emissions and enhance ecosystem benefits. This will result in a higher annual yield, lower demand, and lower car battery charging costs because of the ratoon vehicle-level (RV) technology. The analytical report of CS construction was used to analyze this together with various data points in the entire article.

Numerous techniques are aimed at estimating the necessity of CSs taking into consideration the OD (Origin Destination) trip and SOC level of the vehicles [17]. Considering this, Ref. [19] suggested a model to allocate optimal CSs for the bus network in Toronto. Depending on the vehicle-level SOC cycle and bus trip route, the fastest charging time with minimum cost was assured. An additional state-of-prediction for determining the allocation of discharge according to station characteristics, vehicle range estimation, and battery SOC conditions was the focus of [20,31]. To make the best choice from the available charging stations, the range of an electric vehicle is forecasted based on a fuzzy classifier. This study provided a report of the maximum distance an EV can travel to reach the most suitable charging station. An intelligent route flow system to increase the CSs based on weight and station availability-level projected from the user preference features was proposed by the author in [32]. Based on the remote sensing data, Ref. [21] suggested a city-level prediction for the deployment of charging stations with a minimum budget. The study gave priority to socioeconomic factors such as customer satisfaction, as well as population density. Using the combination of deep learning with phenological properties, this was estimated and forecasted [33]. An optimal placement of charging station in distribution networks was suggested in [22]. To safeguard the charging station CS, the author of [34] provided research based on several charging station discharge categorization models. The suggested model was a data-based station range prediction based on the station’s availability, the cost value of the station’s power demand, and other pertinent aspects and parameters. A business approach to a charging station was proposed in [23]. This study led to the motivation of finding the most optimal placement of charging station, but the technical aspects of the prediction algorithm for the same was not elaborated. Although the battery swapping method as in [35,36] is a great trend these days, taking into consideration the time, it may not be practical in all adverse scenarios. A time series-based approach was discussed in [37] regarding charging demand, but the optimal placement of CS is not suggested in that study.

All the studies discussed so far may mainly be categorized into the following domains based on the decision parameters affecting EVs and their charging stations: power distribution, socioeconomic–environmental factors, EV vehicle, CS parameters, and geographical–topological factors. Most of the studies summarized in the existing literature are based on a plethora of features which are specific to regions. The datasets used by most of the studies are real-time or recorded transportation data focused on a particular town or country, taking into consideration the attribute features of that region. A thorough study was conducted based on the literature summarized above to identify the constraints in each of the specified domains as listed in Figure 1. The traffic and vehicle parameters play a prominent role as influencing factors in all situations. Based on these features, a new optimal prediction system was proposed in our study. The most significant features and their relationship to computing the demand of CSs are discussed in detail in the next section. The proposed framework enhances the accuracy of predicting the best CS from existing ones, as well as forecasting the optimal location for installing a CS, taking into consideration the electric vehicle parameter features obtained from the dynamic tracking of vehicle movements, as well as its State of Charge.

The CS characteristics such as power demand, charging piles, as well as frequency of vehicle occurrences and the distance to reach the CS, are also considered. The forecasting model of the CS represents the best location to place a new CS based on the parameters of power demand, frequency of vehicle occurrences, availability of existing CS, and the peak time of vehicle visits. This will predict the best location for placement of CS according to the suggested co-ordinate positions. For this purpose, cutting-edge techniques like the Enhanced Cladistic Neural Network (ECNN) and the Ensemble Modulation Pattern (EMP) system are recommended.

3. Proposed Methodology

This section explains the proposed procedure of the optimal charging station prediction and forecasting model. When an EV travels from an origin to its destination, its initial co-ordinate positions are initialised based on the OSM data samples. As its journey continues, the state of charge (SOC) of the vehicle decreases and the distance that it can travel depends on vehicle parameters such as its battery make—lithium or hybrid—as well as battery power capacity, driving speed, and the traffic density. The existing charging stations near to the EV current position are made available to the EV. The cluster similarity measures form a cluster of CS where EV can place a charging request when the EV battery has a minimum SOC. Vehicles may prefer level1, level2, or DC fast charging, depending on its capability and charger compatibility. Each charging station has its own charging capacity which is divided into the available charging piles. Hence, the demand of a charging station depends on the availability of the number of charging piles, as well as its maximum charging capacity. That is, if a charging station has a maximum of 100 kW and two charging piles, then the charging capacity of that charging station is restricted to a maximum of 50 kW per pile. Hence the vehicle approaching the charging station must ensure this power demand of the CS by checking if the charging piles are available and other EVs are not plugged in at the station.

The demand of CS is computed based on the generated dataset which takes into consideration the number of charging piles, vehicle density, cost, and charging capacity of the charging station.

The power demand ‘$P_D$’ in a charging station can be calculated by referring to the below equation.

$$P_D=\left(V_C\times \beta \times R\right)-C$$

where, $V_C$—Vehicle Count (number of vehicles that are connected in the piles in individual CS).

$\beta$—Power constant for charging a vehicle.

$R$—Random value to represent the power consumption of each EV vehicle that are connected to the piles of a CS. Each vehicle will consume different amounts of power based on the battery state and the range of battery, therefore, a random value will refer to the different amount of power consumption for connected EVs, along with the charging constant ‘$\beta$’.

$C$—Total capacity of CS that is the ability to supply power to the connected EVs in the piles.

This demand value, if negative, indicates that some piles are empty and vehicles can be connected to the CS. On the other hand, if the demand value is positive, it indicates that the charging stations are fully occupied, and then, based on the historical data and pattern of the feature attributes taken into account in the study, a new charging station location is recommended within the coverage area of that EV. Thus, the forecasting of the CS is based on the vehicle demand and the CS availability, as well as recognizing the area where the vehicle cannot reach the CS with its minimum SOC of battery level.

This prediction is computed using the ECNN algorithm which performs a pattern analysis while tracking the dynamic movement of the EV to available CS with respect to time and the SOC parameters. SOC parameters mainly refer to the percentage of SOC battery level and the distance that can be covered by that EV to a CS. This distance varies from time to time as it depends on the vehicle parameters such as the battery make and capacity, driving speed, and traffic. Hence the proposed system predicts the optimal CS for an EV, and if the CS is already occupied it forecasts an optimal location to install a CS in the future.

Figure 2 displays the main architecture diagram for the proposed charging station prediction model. The methodology is divided into three main stages which are labelled in the Figure 2.

The first stage is the architecture initialization where the source and destination of the path to be travelled by the user are initialized, along with all the necessary vehicle parameters.

The second stage comprises the DHFO algorithm which mainly concentrates on the optimal feature selection based on the demand computation and vehicle parameters. It recognizes the geographical co-ordinates of the CS that the vehicle can reach with its minimum SOC of battery level. The cluster similarity measures form a cluster of CS where EV can place a charging request when the EV battery has minimal SOC.

In the third stage, the Enhanced Cladistic Neural Network (ECNN) technique optimizes the clustering and feature selection, and makes the final prediction. The ECNN algorithm preprocesses the testing weight data using the concept of normalization. The preprocessed data blocks are then split into groups to extract data patterns. Every block group is then subjected to the Ensemble Modulation Pattern (EMP). This efficiently extracts geometrical characteristics to raise the general level of classification accuracy. Finally, the classifier is utilized to classify whether the location ID is satisfied with the relevance ratio or not. The relevance ratio defines the ratio between the number of known values and the predicted values of the data samples.

Thus, the entire forecasting of the CS is based on the power demand and availability of the CS, and the recognition of the area where the vehicle cannot reach the CS with its minimum SOC of battery level. The cluster similarity measures form a cluster of CS where EV can place a charging request when the EV battery has minimal SOC. This will predict the best location for the placement of a CS by suggesting geographical co-ordinate positions extracted using a map matched tool. Figure 3 is the flowchart of the proposed system.

The proposed algorithms are explained in depth in the ensuing subsections.

Section 3.1: Distributional Homogeneity Feature Optimization (DHFO);

Section 3.2: Enhanced Cladistic Neural Network (ECNN) Algorithm.

3.1. Distributional Homogeneity Feature Optimization (DHFO)

The Distributional Homogeneity Feature Optimization is a multi-objective optimization algorithm which uses the various parameters in different objective conditions to finalize the most suitable combination of parameters that will enhance the classification accuracy. Here, the optimization algorithm is used to reduce the size of feature database (before training) of a classifier to enhance the prediction performance. For that, the input of the DHFO optimization algorithm is the full data of the training feature set. As described in the algorithm below, the input is the geographical co-ordinates of the vehicle obtained from the OSM Data Samples and Demand Cost ($M_{NID}$) (the entire charging station parameters as in Figure 4) of randomly generated CS. From these inputs, the output of the DHFO algorithm is the optimal selected feature attributes among the overall database. Based on the selected indexes of the training database, the selected attributes are passed as the input for the classifier in both the training and testing processes.

In Algorithm 1, the multi-objective DHFO method’s precise steps and equation model are made available. To construct the root direction, the input was selected as the co-ordinate position of the available charging stations, and the starting direction was formed based on the vehicle’s geographic latitude–longitude co-ordinates setting of 0.5 if the link difference obtained from the datapoints was less than the defined range as in algorithm below. The coverage area for the randomly allocated CS location in terms of latitude and longitude was ±0.5 from the route path co-ordinate position of EV movement. The number of divisions to allocate the random CS was specified as 10 numbers which split the CS location into different positions. These randomly allocated CS contained the additional parameters along with the latitude and longitude, such as total capacity (kW), number of charging piles, cost (Ps per kWh), vehicle count already connected to the CS piles, and power demand in the CS (kW).

Each charging station contained an individual range of all these listed parameters. These formed the training database for the classifier to define the available CS with/without power demand.

The initial value of SOC was set at 20%. The distance covered by the car in terms of SOC was defined as EV-km-per SOC Unit, being 1.5. The critical value of SOC was 5. This was followed by the feature pattern analysis method to extract the most suitable features for prediction of apt CS using the population strength of the vehicle which was generated in this study along with the power capacity of the vehicle that was connected to the electric vehicle model.

Algorithm 1: DHFO Algorithm

Input: Input Data [co-ordinates and demand Cost ($M_{NID}$)] Output: Predicted CS and demand, $A_{ij}$, $R_{ij}$, $C_{id}$ Step 1: Construct availability matrix and demand cost matrix Let $S_i$ and $S_j$ be size of matrix ($M_{NID}$) and set K = 2; where, $NID$—Node ID For i = 1 to $S_i$            For j = 1 to $S_j$                       Compute the initial position of CS data, $A_{ij}$ by                       $A_{ij}=\left\{\begin{array}{c}0.5 if \left(M_{NID}\right(i,j)<=0.5) \\ 0 Otherwise \end{array}\right.$            End for j; End for i; Step 2: Construct and update responsibility matrix and availability matrix; For $X_k$ = 1 to k            For i = 1 to $S_i$                       For j = 1 to $S_j$                                                        Compute the relevant vector, $R_{ij}$ by                                                        $R_{ij}=\left\{\begin{array}{c}{M}_{NID}(i,j)-A_{ij} If (A_{ij}<=M_{NID}(i,j\left)\right) \\ 0 Otherwise \end{array}\right.$                       End for j            End for i            For i = 1 to $S_i$                       For j = 1 to $S_j$                            Let ${temp\_R}_{ij}=0$;                            For m = 1 to ${\mathrm{S}}_{\mathrm{i}}$                                      ${temp\_R}_{ij}={temp\_R}_{ij}+R_{im}$                            End for m;                            ${temp\_R}_{ij}=\left\{\begin{array}{c}{temp\_R}_{ij}+R_{ij} if ({tem{p}_R}_{ij}\le 0) \\ R_{ij} Otherwise \end{array}\right.$                                                                                  If ($i!=j$), then                                      $A_{ij}=\left\{\begin{array}{c}0 If ({tem{p}_R}_{ij}<0) \\ {temp\_R}_{ij} Otherwise \end{array}\right.$                            Else                                      $A_{ij}=\left\{\begin{array}{c}{temp\_R}_{ij} If ({temp\_R}_{ij}>0) \\ 0 Otherwise \end{array}\right.$                            End if                      End for $S_j$          End for $S_i$ End for $X_k$ Step 3: Compute exponential matrix and the average of relevant vector based on the estimated list index and then update in the list by ${Expm}_{ij}=\left\{\begin{array}{c}1 if\left(A_{ij}+R_{ij}\right)>0\\ 0 Otherwise \end{array}\right.$ $av{g}_{idx}={\displaystyle \frac{1}{x}}\sum _{x=1}^{size\left({Idx}_{ls}\right)}{R}_{i\left({Idx}_x\right)}$ Update $av{g}_{list}🡨av{g}_{idx}$; For y = 1 to $S_j$          Compute the related average of overall vector ${avg}_R$ by          ${avg}_R={\displaystyle \frac{\sum _{j=1}^{S_j}{R}_{ij}}{S_j}}$          Compute the distance list between the average list and the related parameter by          ${dis}_{ls}=\sqrt{{avg}_{List}^2-\left({{avg}_R}^2\right)}$                      Update $C_{id}🡨\mathit{min}\left({dis}_{ls}\right)$ End for y;

The size of the dataset for the N number of attributes in this case is the RN database model. The total dataset size is expressed in terms of M. The key relevancy prediction between feature characteristics is represented by the matrix K_i and K_j. Next, ${\mathrm{wd}}_x, {\mathrm{wd}}_y, \mathrm{a}\mathrm{n}\mathrm{d} {\mathrm{wd}}_{xy}$ are the particle parameters of $K_i \mathrm{a}\mathrm{n}\mathrm{d} K_j$. As indicated by the following equations, the binary representation of these particles is similarly calculated as $M_{xy},N_{xy} \mathrm{a}\mathrm{n}\mathrm{d} N_{yx}$.

The demand cost matrix represents the Euclidean distance between each demand value arranged in the form of a square matrix. In this, Equations (1) and (2) represent the availability matrix and the responsibility matrix, respectively, for the index of ‘i’ and ‘j’. This availability matrix is evaluated with the demand cost and with the responsibility matrix to refresh the matrix based on the update of vehicle availability.

$$A_{ij}=\left\{\begin{array}{c}0.5, if \left(M_{NID}\right(i,j)<=0.5) \\ 0, Otherwise \end{array}\right.$$

(1)

$$R_{ij}=\left\{\begin{array}{c}{M}_{NID}(i,j)-A_{ij} If (A_{ij}<=M_{NID}(i,j\left)\right) \\ 0 Otherwise \end{array}\right.$$

(2)

Next, the ratio of ${\displaystyle \frac{1}{\log \left({\mathrm{Derv}}_{xy}\right)}}$ and ${\mathrm{dif}}_v$ is used to compute the distance ${\mathrm{dist}}_{xy}$. As a result, the sum of $T_{\mathrm{dist}} \mathrm{a}\mathrm{n}\mathrm{d} {\mathrm{dist}}_{xy}$ is updated in the $T_{\mathrm{dist}}.$

$${dis}_{ls}=\sqrt{{avg}_{List}^2-\left({{avg}_R}^2\right)}$$

(3)

$$av{g}_{idx}={\displaystyle \frac{1}{x}}{\sum }_{x=1}^{size\left({Idx}_{ls}\right)}{R}_{i\left({Idx}_x\right)}$$

(4)

The values of $T_{\mathrm{dist}}$ are then used to update the value of $O_{\mathrm{dist}}$, and lastly, the updated ${\displaystyle \frac{O_{\mathrm{dist}}}{M}}$ is used to estimate $M_{\mathrm{NID}}\left(i,j\right).$

The charging station architecture optimized the station’s orientation and set it up for quicker data transfer based on the evaluation of these criteria. The output of DHFO was the optimal feature database (training database), and the selected attributes of the overall training database were used as the input for the ECNN classifier.

3.2. Enhanced Cladistic Neural Network (ECNN) Algorithm

The optimal feature database (training database), which is the output of the above-described DHFO, was the extracted attributes of the overall training database that was used as the input for the ECNN classifier. It was computed like the evolutionary relationship concept in the construction of a cladogram. Cladograms are often used to create a linear representation of hypotheses. They describe the evolutionary relationships between varied features, showing how they are organized depending on co-operative characteristics. The ECNN-based classification algorithm is used to predict the CS position based on the EV and CS features that are optimized from the DHFO algorithm. This will estimate the parameters of the vehicle, and allow the CS module to train itself on a suitable combination of indexes. From that, if a vehicle requests any CS prediction, this will find the best co-ordinate position of CS that is related to the satisfactory level of the vehicle for it to charge its battery. This classification model can also forecast and recommend the location to install a CS later, based on the demand estimation and the history of trained data from the available features. This will update the training database for improving the performance in accuracy and sensitivity levels. The ECNN algorithm takes into consideration the vehicle capacity that can be integrated into the architecture depending on the minimal link coverage distance that is the time it takes to reach the available CS, which may vary from time to time. This algorithm’s primary contribution is its ability to forecast the optimal CS, based on the parameters influencing this link coverage. The link coverage varies from time to time as it depends on the vehicle parameters such as its model, the battery make and capacity, driving speed, and traffic.

The model training begins with data preprocessing. The main phases of preprocessing undergone in the proposed framework are (i) data normalization, (ii) feature optimization using DHFO, and (iii) train test split.

3.2.1. Data Normalization

The normalization process aims to scale the numerical values in a dataset to a common range without affecting the range of values. The normalization process applied in the proposed model is min–max normalization applied as in equation,

$$X^{\prime }={\displaystyle \frac{X-Xmin}{Xmax-Xmin}}$$

3.2.2. Feature Optimization

Feature optimization is the process to minimize the number of features to decrease the complexity of computation, as well as to improve the efficiency of the training model. The feature optimization technique used in the proposed model is DHFO which has been described in Section 3.1. Every optimization iteration received the instance’s data matrix as input. Here it is used to reduce the size of the dataset before training the classifier to enhance the prediction performance. As mentioned in Algorithm 1, the input is Co-ordinates and Demand Cost (MNID) (the full table parameters are in Figure 4) of randomly generated CS. Thus, the training feature set of the classifier model comprises the collected OSM data co-ordinate positions, the selected attributes from the parameters of CS and EV such as SOC, Km can travel by vehicle, Cost of CS, and Demand in (KW) from the randomly generated CS. The optimal feature database obtained from DHFO and the selected attributes are passed as the input to the ECNN classifier in both the training and testing processes. The data matrix was updated every time the ideal charging station direction was chosen. The training model of the ECNN was used to classify the classes of live moving vehicle parameters.

3.2.3. Train Test Split

The final phase in preprocessing is to split the dataset into train and test sets. The process of splitting the dataset is a typical phase in learning algorithms to address overfitting and for evaluating the performance of the trained model properly on unseen data. In the proposed work, the dataset was split in the ratio of 70:30, where 70% was used for training and remaining 30% was retained for model testing.

3.2.4. Classification Model Description

In this process, the ECNN is a multi-level classifier used for the prediction of the best location for a charging station based on the feature attributes that are collected from the parameters of the CS and EV. The classification result from the ECNN is the prediction of the nearest CS location in the route path based on the parameters of SOC, Km can travel by vehicle, Cost of CS, and Demand in (KW). If there was no CS predicted in the coverage area of the current moving vehicle, then this recommended the location of CS based on the demand level and the distance to the CS in the coverage area. The labels that are considered for the classification are:

(a)

CS station available with sufficient power to charge the EV—Class 1;

(b)

CS station available with power demand (KW)—Class 2;

(c)

CS station not available—Class 3;

(i)

Recommend new CS station;

(ii)

Recommend the longitude and latitude of new CS station.

These classes are defined by the training set for the classifier extracted from the parameters of the CS and EV. These parameters are arranged in the form of feature attributes to construct the training model for the classifier. This training model will be a one-time process while installing the system on the manufacturing side. While driving the EV, the parameters from the live running vehicle will be arranged as the testing feature vector. At this testing stage, the parameters from each time point are used as the input to the classifier for testing. The classifier predicts the class according to the best match with the training database (feature set) and gives out the class label as 1, 2, or 3. If it predicted class 3, then the classifier will again validate the parameters to recommend the best location for a new CS, and this produces the longitude and latitude for the new CS.

The ECNN model’s algorithm steps are given in Algorithm 2:

Algorithm 2: ECNN Algorithm

Input: Input vehicle pattern set $T_D\left(s\right)$ Output: Classified result $V\left(k\right)$ The initial pattern for the prediction is arranged as, $T_D\left(s\right)=\left\{T_{D1}\left(s\right), T_{D2}\left(s\right), \dots , T_{Dm}\left(s\right)\right\}$ In the input layer of the neural network, the data sequence can be formed as the matrix in the equation below. $X_D\left(s\right)=\left[\begin{array}{c}{T}_{D1}\left(s\right)\\ T_{D2}\left(s\right)\\ \begin{array}{c}\dots \\ T_{Dm}\left(s\right)\end{array}\end{array}\right]$ From the matrix arrangement, the block correlation feature can be estimated by $F\left(X_D\left(s\right).X_D^{\ast }\left(s\right)\right)$. This can be represented as $F\left(X_D\left(s\right).X_D^{\ast }\left(s\right)\right)=X_D^{\ast }.e^{T-T_m}$ Where, ‘$T$’ and ‘$T_m$’ represent the attribute values from matrix $X_D\left(s\right)$. Estimate the kernel model of classifier $K_m={\displaystyle \frac{1}{2^{q-1}}}{\left({\displaystyle \frac{\sqrt{2q}}{l}}\right)}^q{k}_q\left({\displaystyle \frac{\sqrt{2q}}{l}}r\right) \forall q=1, 2, \dots , N$ //’$r$’ represents the range of feature distance, ‘$l$’ represents the length of the feature vector. Estimate the relevancy using the kernel function with feature points. $t_n=F^T{\omega }_n$ Where, ‘${\omega }_n$’ weight value of attributes. $u_n=F^T{\omega }_n$ Extract the training features and form the network by $T_r=\left\{t_1, t_2,\dots , t_n\right\}$ $X_b=\overline{X_b}+\sum _{i=1}^N{t}_i\left(d\right)p^i$ Estimate the matching score for the correlated blocks by $\widehat{T_s}={\left({\left(X_b^{\overline{d}}-\overline{X_b}\right)}^T\left(P^T\right)\right)}^T$ Where, the relevance factor $X_b^{\overline{d}}\in R^{\left(T-T_p\right)M}$ can be written as $R^{\left(T-T_p\right)M}={\widehat{T_s}}^T{Q}^T+\overline{t{t}_a}$ Where, ‘P’ and ‘$Q^T$’—predicted component. The predicted label can be represented by $V\left(k\right)={\displaystyle \frac{d_{ij}}{R_j-R_i}}$ Where, $d_{ij}$—distance matrix for ‘$i$’ and ‘$j$’ of the relevance matrix ‘R.

4. Results and Discussion

The outcomes of the suggested model for classifying charging stations according to availability and weight characteristics were verified and contrasted with the approaches already in use. The first Section 4.1 was a briefing on the experimental setup and the dataset. The tracking of the map-matched data samples with the output of the experiment was discussed in Section 4.2. The performance indicators used for the study were discussed in the second Section 4.3, followed by the quantitative evaluation based on the metrices used to calculate the performance of the suggested classification model in Section 4.4. The computational complexity of the system was analyzed in the following subsection E. An investigation based on the road map matched with the OSM samples and experimental results obtained was elaborated, followed by the performance comparison with other existing approaches in the last Section 4.5.

4.1. Experimental Setup

Python 3.19 was used for the experimental study and testing. To perform the weight flow analysis, the position and co-ordinate data were chosen based on the coverage area of the electric vehicle.

The data samples for the experimental setup were generated from the OpenStreetMap (OSM). OSM is a map embed API that provides map-matched data for the Origin Destination (O-D) Road Trip as planned for the EV user (https://www.openstreetmap.org/#map=5/21.494/66.665, accessed on 26 January 2024). The geographic data are collected from the traffic co-ordinate positions in the OSM database. Similarly, an archival database of several map-harmonized tours are taken into consideration. According to the direction on the map, the CS co-ordinate positions and the specification of CS are simulated and verified. This includes weight data from the vehicle database, with some data created as placeholders for missing values to forecast and predict the data. In this work, there were five different road trip datasets generated for the locations from Bangalore to Mysore(D1), Bangalore to Pune(D2), Bangalore to Tumkur(D3), Bangalore to Cochin(D4), and Bangalore to Hyderabad(D5) in random directions on the map. These trips interconnecting different Sates in Southern India were of variable length, as shown in Figure 5a. Figure 5b shows the map direction D1, from the source of Bangalore to the destination of Mysore. Each part of this trip was further fragmented into several road segments. Random position changes and the creation of a situation where data are missing are then generated and simulated in various modules. The Open-Source Routing Machine (OSRM) routing service provided the available charging stations on the routes chosen for the study.

Figure 4 shows the initialized demand parameters of CS at each co-ordinate position which comprises co-ordinate positions of the CS in terms of latitude and longitude, the maximum output power that a CS provides in kW, number of piles available in the CS, cost per kWh (varying from 5 to 10), and number of vehicles already getting charged. These parameter values are used to compute the power demand of CS. The CS with and without demand are further examined using the DHFO and ECNN algorithm to predict the most optimal charging station for the EV users.

4.2. Track Analysis with Map

The map-matched route trips taken from the OpenStreetMap Database were utilized to perform the proposed study. The data of available charging stations on the trips taken were extracted using the Open-Source Routing Machine (OSRM). All the available charging stations near to the EV were grouped together based on clustering. The demand value of each CS was computed as described in the DHFO algorithm, taking into consideration the vehicle SOC and traffic density. The demand value in the positive range indicates that a greater number of vehicles can be charged there. The negative value of demand suggests unavailability of suitable charging stations, and hence there is an urge to place one there.

Figure 6 shows the simulation update of vehicle movement in all the O-D road trip datasets that were generated for the locations from Bangalore to Mysore (D1), Bangalore to Pune (D2), Bangalore to Tumkur (D3), Bangalore to Cochin (D4), and Bangalore to Hyderabad (D5) in random directions on the map. The dynamic prediction of optimal CS placement is represented by the legend marks. In this, the black color represents the vehicle and the orange color represents the movement tracking of the EV. The green dots represent the co-ordinate positions of all the charging stations adjacent to the route of the electric vehicle, traced using the clustering algorithm. The red dot in the figure indicates the co-ordinate positions of the CS which are of high demand in the path for the specified source to destination. The purple dot indicates the co-ordinate positions of the CS based on the optimization algorithm. It represents the forecasted and recommended location to place the CS. It was obtained based on the demand estimation with vehicle distance. The prediction was done dynamically, based on the classification algorithm taking into consideration the selected feature parameters from the DHFO algorithm and computing the demand MNID (i, j), as described in Algorithms 1 and 2 above. As the EV travels towards to charging station, the SOC reduces. A minimum threshold of 5% was set for the SOC to ensure that the EV would be charged before it was totally drained. The EV, on reaching this minimum threshold of SOC, will identify the predicted CS and become charged. Accordingly, the SOC will be updated and the vehicle travels towards its destination. Each of the data samples of the O-D trip taken (D1–D7 described in the following section) were further divided into more than 50 datapoints, overall. Thus, we considered approximately 350 datapoints. The ensuing subsections address the performance outcomes.

4.3. Performance Indicators

By calculating the statistical probability between the number of misclassified results and the properly classified data samples, one can determine the parameters utilized in the performance metric-based analytical procedure. The confusion matrix layout is used to generate all of these, which are then assessed by comparing the categorized results to the dataset’s ground truth. The true positives and false positives of the predictions made are used to compute the precision, recall, accuracy, and F1 score. Each of these measures is significant, according to the varying situations, to get a broader understanding of the errors and performance of the model. The accuracy is the baseline measure used in the classifier model to get an indication of how efficient the system is. It is computed as in Equation (5).

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}, \mathrm{A}\mathrm{c}\mathrm{c}={\displaystyle \frac{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{c}\mathrm{o}\mathrm{r}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{t} \mathrm{l}\mathrm{a}\mathrm{b}\mathrm{e}\mathrm{l}\mathrm{s}}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{N}\mathrm{o}. \mathrm{o}\mathrm{f} \mathrm{S}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}}$$

(5)

To ensure that the system performs well, even when the datasets are not properly balanced, we utilized the precision value as it gives the ratio of true positives across all the positive values. It is beneficial in scenarios when the estimate of false positives is high. The mathematical equation of precision is mentioned in Equation (6) stated below.

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}, \mathrm{P}=\left(1-\mathrm{F}\mathrm{P}\mathrm{R}\right)={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(6)

Recall is taken into consideration when the estimate of false negatives is high and the equation to compute it is specified in Equation (7) stated below.

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}, \mathrm{R}=\left(1-\mathrm{F}\mathrm{N}\mathrm{R}\right)$$

(7)

The F1 score is also considered to find an equilibrium between the precision and recall. The F1 score is a single metric that is computed as the harmonic mean of precision and recall, as in Equation (8).

$$\mathrm{F}1 \mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}, \mathrm{F}\_\mathrm{S}={\displaystyle \frac{2\mathrm{P}\mathrm{R}}{\mathrm{P}+\mathrm{R}}}$$

(8)

Area under Receiver Operating Characteristic (ROC) curve is another important metric used to evaluate the performance of the model. The ROC curve is plotted as the curve between recall or sensitivity defined in Equation (9) and the false positive rate (FPR) which is (1 − specificity which is mentioned in Equation (10)) or the proportion of incorrect classification made by the model, as in Equation (11).

$$\mathrm{S}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{i}\mathrm{t}\mathrm{y}, \mathrm{T}\mathrm{P}\mathrm{R}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e} \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\left(\mathrm{T}\mathrm{P}\right)}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{N}\mathrm{o}. \mathrm{o}\mathrm{f} \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}}$$

(9)

$$\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}, \mathrm{T}\mathrm{N}\mathrm{R}={\displaystyle \frac{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e} \mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\left(\mathrm{T}\mathrm{N}\right)}{\mathrm{T}\mathrm{o}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{N}\mathrm{o}. \mathrm{o}\mathrm{f} \mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e} \mathrm{s}\mathrm{a}\mathrm{m}\mathrm{p}\mathrm{l}\mathrm{e}\mathrm{s}}}$$

(10)

$$1-\mathrm{S}\mathrm{p}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{f}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{t}\mathrm{y}=\mathrm{F}\mathrm{P}\mathrm{R}={\displaystyle \frac{\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e} \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\left(\mathrm{F}\mathrm{N}\right)}{\mathrm{T}\mathrm{r}\mathrm{u}\mathrm{e} \mathrm{N}\mathrm{e}\mathrm{g}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\left(\mathrm{T}\mathrm{N}\right)+\mathrm{F}\mathrm{a}\mathrm{l}\mathrm{s}\mathrm{e} \mathrm{P}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{t}\mathrm{i}\mathrm{v}\mathrm{e}\left(\mathrm{F}\mathrm{P}\right)}}$$

(11)

4.4. Quantitative Evaluation Based on Performance Measures

The complete outline of the performance of the system was obtained using the above-defined performance metrices. The five defined road O-D trips in Figure 5, as mentioned in the experimental setup, were taken into consideration. The performance measures of precision, recall, F1 score, and accuracy were computed across each of these segments.

The plot showing the tradeoff in these performance measures is visually represented in Figure 7. It was observed that the model obtained a higher recall and accuracy measure across all five defined road trips. The higher recall shows that the system was successful in predicting the maximum number of optimal CSs available for the EV user for a particular road trip. The higher accuracy value across all the road segments ensured that the system’s overall performance was high. Hence it can be inferred that the model will ensure that the EV user will be intimated with all possible optimal CS on the path it is travelling on. The ROC curve outlining the performance of the proposed model on the different datasets is plotted in Figure 8. The plot is, again, an indication that the use of optimal feature selection and demand computation of the proposed system results in elevated results (>0.9) in the proposed application. The area under the curves is labelled in the graph.

4.5. Computational Complexity

The proposed study mainly comprises two main algorithms—DHFO and ECNN. The computational complexity of the system depends on the working of the feature optimization and neural network algorithms, as discussed in [38]. The DHFO algorithm takes into consideration the input co-ordinates of the EV user, as well as the charging station. The availability and the demand matrix are updated for each of the nodes taken into consideration. The construction of the relevancy matrix is nested for loop, which is iterated to the maximum of N, where N is the total number of nodes. It is followed by the updating part of the above-constructed matrix based on the relevant vector as well as the vehicle node co-ordinate positions. This modification again includes an iteration within the above-nested for loop, thereby the final computational complexity is O(n³).

The complexity of the algorithm depends on the training and testing phases. During the training phase, the complexity of the one-time step of the forward pass of each layer involves matrix manipulations resulting in a complexity of O (n_i × n_i+1), where i refers to the layers. The updating of the weights in the backward pass also results in O (n_i × n_i+1) computational time. The estimation of the matching score for the correlated blocks is O (n² R), where R represents the total attributes. Hence, total computational complexity of the ECNN algorithm, taking into consideration k epochs and T time steps, is $\mathrm{O}(\mathrm{k}.\mathrm{T}\sum _{\mathrm{i}=1}^{\mathrm{n}-1}({\mathrm{n}}_{\mathrm{i}}\ast {\mathrm{n}}_{\mathrm{i}+1}))+\mathrm{O}({\mathrm{n}}^2.\mathrm{R})$. For the testing phase, the computational complexity is much less linear in nature as it is directly proportional to the input size only.

4.6. Comparison with Existing Approaches

Several studies have been made to resolve the range anxiety problem of EV users. The optimal prediction of charging station is the primary concern of all these studies. The study in [8] examined deep learning methods for optimal prediction. In the experimental study, the OSM road network from Denmark was used, along with weather and digital elevation information, to make the predictions. The speed profile was investigated based on different trip segments. The statistics of the investigated models were tabulated and, being a regression model, the Root Mean Square Error (RMSE) and Mean Absolute Error (MSE) were tabulated. The results showed that the system performed exceptionally well with a RMSE of approximately 0.8 for all kinds of speed profile when compared to other deep learning models.

In our study we proposed an enhanced cladistic neural network classifier algorithm with Distributed Homogenous Feature optimization, thereby chalking out the most significant features for prediction of the CSS. Hence, for analysis, we used the translation function and converted the above regression results setting into a suitable threshold based on the domain area where we could conclude that most of the DNN-based models showed an outstanding performance. Using the threshold of 2.0, the classification performance metrics computed for each regression method indicated that most methods performed similarly, with high precision above 0.9, recall above 0.8, and F1-scores above 0.85, approximately. Figure 9 depicts the route from Denmark to Hannover with the prediction of optimal CS using the proposed system, which is labelled D6.

Figure 10 represents the geographical picture of the route trip from the road network in Wuhan which is a case study in [39]. The route chosen was from Wuhan to Wanggang, and labelled D7. In the study in [39], different baseline standards were chosen to analyze the influence of urban area in the prediction. The RMSE and MAE for the different baselines were tabulated. These hypothetical regression results, when converted using the trade-off threshold 2.0, clearly revealed that the effectiveness was high, with an accuracy and precision of approximately 0.9, and a precision and F1 score of almost over 0.85.

The performance measures obtained for these two road trips, D6 and D7, using our system, were also of a high standard in terms of accuracy, precision, and recall. In Table 1, the results are evaluated from the tracking of vehicle movement with respect to time response based on the SOC parameters. This result is simulated for routes D6 and D7. With our proposed approach, the optimal charging stations predicted are highlighted as purple dots in Figure 9 and Figure 10.

The proposed technique could classify the results in a generalized approach by effectively extracting the data patterns of the input image and taking into consideration the most optimized parameters from the OSM sample data. The approaches which have been taken for the analysis in the current study also include additional parameters from the urban region, as well as environmental factors such as weather perspective. This proposed study is not specific, and is independent of whether the samples were taken from urban or rural areas, as well of other environmental factors, but still provides a promising and satisfactory prediction of suitable CS.

In this process, the ECNN classifier worked as the multi-level classifier to predict the best location based on the feature attributes that were collected from the parameters of the CS and EV. While driving the EV, the parameters from the live running vehicle were arranged as the testing feature vector. At this testing stage, the parameters from each time point were used as the input to the classifier for testing. The classifier predicted the class according to the best match with the training database (feature set) and gave out the class label as 1:CS station available with sufficient power to charge the EV (green-colored dots), CS station available with power demand (KW)—Class 2 (red-colored dots), or no CS—Class 3 (purple-colored dots as in Figure 9 and Figure 10). If it predicted class 3, then the classifier would, again, validate the parameters to recommend the best location for a new CS that gives out the longitude and latitude for the new CS.

In this work, the data patterns were efficiently extracted using DHFO, and the moving vehicle was accurately detected by using the ECNN technique. From these CS location predictions and the tracking performance evaluation, the parameters representing the performance of the proposed model were satisfactory when compared with other state-of-art methods. The precision, recall, and accuracy defined the sensitivity of the proposed algorithm in the prediction of CS and the forecasting of the CS location to recommend the position where we can optimally place the CS in the path of direction.

Table 1 and Figure 11 show the accuracy and tracking performance rates of the existing approaches: Data-Driven Context Aware Learning DDCAL [39], Probabilistic Deep learning for EV [8], Machine learning Analysis (MLA) based on Linear Regression and Support Vector Machine [2], Fuzzy-based multi-criteria decision-making NWHFMCGDM [40], and Proposed Techniques. All these existing approaches have been studied using the data samples D6 and D7. These results are evaluated based on the successful classification with less time difference. To track the evaluation, the performance measures’ accuracy and tracking performance rates of these approaches were calculated. From the evaluation, it was found that the proposed system provides an efficient result, when compared to the other techniques, and it is stable across all the data routes. To compare the efficiency of feature extraction and classification performance across the systems, the tracking performance rate was used. The tracking performance depends on the computational time taken with respect to the data used, as well as the hyperparameters which are model-specific tunable settings. In this work, the data patterns were efficiently extracted using DHFO, and the moving vehicle could be accurately detected by using the ECNN technique.

From the CS location predictions and the tracking performance evaluation, the parameters representing the performance of the proposed model were better than the other state-of-art methods.

Table 1. Comparative analysis with respect to D6 and D7.

Data Sample Taken	Methods	Accuracy	Tracking Performance %
D6	DDCAL [39]	0.90	80
	PDLEV [8]	0.90	94.45
	MLA [2]	0.90	90.01
	NWHFMGDM [41]	0.89	99.00
	Proposed	0.91	99.89
D7	DDCAL [39]	0.89	89
	PDLEV [8]	0.88	93.29
	MLA [2]	0.89	91.2
	NWHFMGDM [41]	0.88	99.1
	Proposed	0.91	99.95

Table 1. Comparative analysis with respect to D6 and D7.

Data Sample Taken	Methods	Accuracy	Tracking Performance %
D6	DDCAL [39]	0.90	80
	PDLEV [8]	0.90	94.45
	MLA [2]	0.90	90.01
	NWHFMGDM [41]	0.89	99.00
	Proposed	0.91	99.89
D7	DDCAL [39]	0.89	89
	PDLEV [8]	0.88	93.29
	MLA [2]	0.89	91.2
	NWHFMGDM [41]	0.88	99.1
	Proposed	0.91	99.95

5. Conclusions

An innovative charging station prediction model was put forth in this era of the Internet of vehicles [42], utilizing an Enhanced Cladistic Neural Network (ECNN) for charging station prediction and an adaptive Distributional Homogeneity Feature Optimization (DHFO) method for optimal feature selection based on pattern analysis. The results were evaluated from the tracking of vehicle movement, in which the EV could reach the CS with an optimal time difference, and it ensured a high score in all standard metrics across all the OSM data samples used, showing the stability of the system’s performance. From the evaluation, it was found that the proposed system provided a satisfactory result, when compared to the other existing techniques, by taking into consideration only the most significant features obtained on the pattern of vehicle for prediction of the most optimal existing CS, as well forecasting the geographical co-ordinates apt for the installation of the CS. The additional socioeconomic and geographical factors, as described in other existing approaches, influenced the system, but their impact was considerably small on vehicle and CS parameters.

Limitations and Future Prospects

The study mainly concentrated on the charging station prediction, as well as forecasting of location to install a new CS, assuming the higher adoption rate of EV. The study was predominantly based on dynamic vehicle tracking and was not validated for the standard EV datasets, as most of the available datasets are region-specific and may drastically change with time, depending on the adoption rate of EV in those regions. Although the geographical co-ordinates of a future CS have been forecasted, the power distribution and government policies, as well as guidelines, must also be studied prior to the final installation of the charging stations. Based on feature optimization model advancement, this kind of pattern-based prediction model may be tested and implemented in the future for large data in car patterns. Further studies could also be conducted, based on a time series plot with different vehicle models.