Enhancing Urban Electric Vehicle (EV) Fleet Management Efficiency in Smart Cities: A Predictive Hybrid Deep Learning Framework

Highlights

What are the main findings?

• The GNN-ViGNet model achieved 98.9% accuracy in forecasting EV charging loads, setting a new standard in predictive analytics.
• The Coati–Northern Goshawk Optimization hybrid algorithm reduced fleet travel distances to 511 km, outperforming established optimization techniques.

What is the implication of the main finding?

• Enhanced forecasting and route optimization significantly cut EV fleet energy consumption and costs.
• Real-time IoT data integration supports adaptive, sustainable urban transportation.

Abstract

Rapid technology advances have made managing charging loads and optimizing routes for electric vehicle (EV) fleets, especially in cities, increasingly important. IoT sensors in EV charging stations and cars enhance prediction and optimization algorithms with real-time data on charging behaviors, traffic, vehicle locations, and environmental factors. These IoT data enable the GNN-ViGNet hybrid deep learning model to anticipate electric vehicle charging needs. Data from 400,000 IoT sensors at charging stations and vehicles in Texas were analyzed to identify EV charging patterns. These IoT sensors capture crucial parameters, including charging habits, traffic conditions, and other environmental elements. Frequency-Aware Dynamic Range Scaling and advanced preparation methods, such as Categorical Encoding, were employed to improve data quality. The GNN-ViGNet model achieved 98.9% accuracy. The Forecast Accuracy Rate (FAR) and Charging Load Variation Index (CLVI) were introduced alongside Root-Mean-Square Error (RMSE) and Mean Square Error (MSE) to assess the model’s predictive power further. This study presents a prediction model and a hybrid Coati–Northern Goshawk Optimization (Coati–NGO) route optimization method. Routes can be real-time adjusted using IoT data, including traffic, vehicle locations, and battery life. The suggested Coati–NGO approach combines the exploratory capabilities of Coati Optimization (COA) with the benefits of Northern Goshawk Optimization (NGO). It was more efficient than Particle Swarm Optimization (919 km) and the Firefly Algorithm (914 km), reducing the journey distance to 511 km. The hybrid strategy converged more quickly and reached optimal results in 100 rounds. This comprehensive EV fleet management solution enhances charging infrastructure efficiency, reduces operational costs, and improves fleet performance using real-time IoT data, offering a scalable and practical solution for urban EV transportation.

1. Introduction

Electric vehicles (EVs) are gaining popularity in transportation and energy sectors owing to their rapid expansion [1]. As EV use expands, managing charging infrastructure and optimizing energy supply for large-scale businesses are challenges. IoT sensors at charging stations and EVs monitor charging trends, power usage, car positions, and traffic conditions in real time. IoT data are incorporated into advanced forecasting models and route optimization algorithms to improve EV fleet performance. Data-driven infrastructure management using IoT, EV charging, and route planning reduces energy use, operational costs, and environmental impact. Electric vehicles improve transportation by saving energy and decreasing pollution. As EVs become more popular worldwide, governments like the U.S. are scaling their efforts, with millions of EVs currently in metropolitan transportation networks [2]. A large fleet of EVs charging on the power grid threatens electricity consumption, peak-to-valley load changes, and power distribution network stability [3]. Energy suppliers may watch real-time power usage, estimate peak demands, and adapt distribution networks to ensure grid stability using IoT data from EVs and charging stations. Traffic and environmental conditions make EV drivers unpredictable, requiring accurate load forecasting models. These models are crucial for assessing EV support in distribution networks, enhancing fleet efficiency, and developing infrastructure [4]. Urban transportation and electricity systems need an accurate spatio-temporal distribution model for EV charging demand. These models need constant car behavior, charging station use, and environmental data from the IoT.

Smart city infrastructure and IoT have solved some of these issues. Smart cities use IoT-based utilities and urban infrastructure to collect enormous quantities of data for urban planning and quality of life. Aggregating and integrating diverse data is difficult. IoT sensors aggregate data from urban transportation networks in real time to improve EV charging demand models’ projections and decision-making plans for EV charging infrastructure, including large-scale battery charging stations [5]. IoT-driven data aggregation optimizes this connection for accurate predictions and real-time decision-making, enabling efficient and scalable EV fleet operations. The IoT improves EV fleet efficiency, especially in dynamic route optimization. The IoT and cyber-physical systems (CPS) have improved logistical efficiency and sustainability [6]. The IoT and cognitive algorithms like Hybrid Firefly–Swarm Hybrid Optimization (HFSHO) allow efficient logistics route planning, vehicle scheduling, and fleet performance. Similarly, IoT-enabled route optimization may help EVs locate the most energy-efficient routes and account for real-time charging requests and grid circumstances. Traditional route optimization and load predictions fail as EV fleets become more complicated. EV charging load forecasting models have primarily focused on EV load curves without incorporating spatial-temporal charging demand fluctuations, which is critical for designing BCCSs and BDSs [7]. BCCSs optimize battery charging power, whereas BDSs enable EV users to switch batteries swiftly without affecting the grid. These two methods lower grid load, enhance energy management, and increase urban EV adoption. Optimizing these facilities and guaranteeing efficient energy distribution requires an accurate spatial-temporal load forecasting model [8].

Many models have been created to predict EV charging demand in recent years. However, many ignore EV customers’ driving and charging habits. Traditional forecasting methods sometimes ignore traffic delays and environmental variables, significantly affecting charges [9]. Many of these models also employ preset values for charging site and time, which cannot adequately represent EV usage’s dynamic and stochastic nature [10]. Stochastic travel chains, Monte Carlo simulations, and Markov decision theory are used in sophisticated models to handle issues. These models forecast charging demand using randomization, making them more realistic [11]. Even sophisticated models have limits. Many ignore traffic congestion and ambient temperature’s effects on EV driving. The transportation system and distribution network are closely coupled; therefore, modeling these factors is crucial to understanding EV charging demand [12]. The geographical distribution of charging stations, vehicle use patterns, and availability of distributed generation (DG) sources like photovoltaics affect grid demand. Optimizing charging infrastructure development and operation requires incorporating these parameters into load forecasting models [13]. This study develops an IoT-enabled algorithm to estimate EV charging demand and optimize EV fleet routes in real time. IoT devices provide large quantities of data, enabling dynamic adaptation to charging needs and road conditions, maximizing energy efficiency and minimizing grid disruption [14]. The proposed system would estimate charging demand and optimize EV fleet routing to balance energy use with grid conditions and fleet requirements.

Electric car fleet management is complicated by changing charging demands and traffic. Using IoT data and AI to solve these problems is game-changing. Using real-time data from charging stations, traffic sensors, and vehicle systems, advanced optimization algorithms and predictive models may boost operational efficiency, reduce energy use, and simplify route planning. The IoT-enabled EV charging and route optimization technology offers several benefits: it improves charging demand forecasting and responsiveness, reducing peak loads and grid reliability. Real-time route optimization may reduce EV fleet energy use, operating costs, and service efficiency. Cities developing and seeking efficient, eco-friendly transportation need this.

Many contributions come from this study. Electric vehicle (EV) charging load predictions and route optimization are addressed by various novel methods in this study. GNN-ViGNet, a hybrid regression model, uses GNN, VGGNet, and Inception architectures to improve EV charging load predictions. The model outperforms previous methods in key performance aspects by using these architectures’ capabilities. Two unique performance metrics—Charging Load Variance Index (CLVI) and Forecast Accuracy Rate (FAR)—are created to evaluate the model’s capacity to handle major load changes and eliminate forecasting mistakes. New feature engineering methods capture the complexity of EV charging behavior. Effective Charging Rate (ECR), Charging Demand Index (CDI), Weather Impact Score (WIS), and Peak Hour Indicator (PHI) are designed to increase forecast accuracy and EV operating insights. A novel Hybrid Coati–Northern Goshawk Optimization (Coati–NGO) algorithm combines global exploration with local exploitation to optimize routes in dynamic EV fleet management situations. Compared to traditional algorithms, the GNN-ViGNet and Coati–NGO techniques solve dynamic EV charging and routing problems with higher accuracy, fewer mistakes, and quicker convergence.

When optimizing transportation and power networks, it is essential to consider operational restrictions, including grid capacity, peak load needs, and renewable energy availability. This is because EV fleet management and power grid operations are interdependent. A dual optimization framework can be built using the foundations of this research.

The article’s structure is as follows: Section 2 compiles significant related works, considering EV charging load forecasting, optimal pathfinding, and limitations and gaps in existing research. Section 3 provides a detailed description of the proposed methodology and framework. Section 4 details all of the experimental results. Finally, Section 5 concludes with a discussion of the findings and recommendations for further research.

2. Related Work

EVs have created new challenges for transportation systems and energy grids, notably in charging infrastructure management and fleet route optimization. To solve these problems, two main study topics have emerged: EV charging load forecasting, which estimates charging station demand and balances grid load, and optimum pathfinding, which identifies the most effective routes for EV fleets while accounting for charging demands and real-time power grid stability, and charging infrastructure efficiency rely on EV charging demand projections. EV charging times and locations must be accurately estimated to reduce peak demand and effectively place charging stations. EV fleets must select the best route to conserve energy and time and reach charging stations. Route optimization must account for real-time charging availability and grid conditions, while load forecasting must reflect EV fleet behavior. The literature review on EV charging load forecasting and optimal route finding outlines essential methods, accomplishments, and limitations. These fields need study and methodologies, which are discussed below.

2.1. EV Charging Load Forecasting

Recent research on EV charging load forecasting has shown the importance of precisely estimating EV energy consumption for power grid stability and EV infrastructure optimization. The authors of [15] created an EV load forecasting model that analyzed vehicle energy use and driver behavior. This model relied on past charging data and could not account for spatiotemporal fluctuation in charging needs, limiting its scalability in dynamic urban contexts.

A forecasting model for EV charging stations was created in [16] by merging traffic and energy usage data. The model enhanced load estimates by integrating real-time charging behaviors. However, meteorological circumstances, which affect EV charging demand, were excluded. An effective spatial-temporal forecasting technique was developed in [17] to map the geographical distribution of EV charging demand. It provided a more detailed perspective of charging station utilization across regions, but scaling to bigger, more complicated metropolitan grids needed additional testing.

In [18], customer-specific charging profiles were utilized to improve forecast accuracy by capturing individual habits. However, the study overlooked broader factors, such as the impact of distributed generation (DG) sources like photovoltaics on grid demand. The author in [19] emphasized the need to combine BCCS and BDS in predicting models. The framework optimized the charging burden for large-scale activities. It lacked real-time load changes, crucial for highly dynamic metropolitan networks with fluctuating EV demand.

A Monte Carlo simulation was utilized in [20] to simulate the stochastic nature of EV charging requests. Load forecasting algorithms must account for random driving and environmental conditions. The approach produced more accurate forecasts but needed a lot of processing power, restricting its real-time use. The model in [21] for short-term load forecasting at EV charging stations employed a mathematical model to predict the aggregated demand for electric buses. Because it concentrated on electric buses, it was less relevant to various EVs with varying charging patterns.

In [22], a hybrid model used a deep learning-based transformer model to enhance load forecast accuracy. That model was accurate but took a long time to train. The processing burden hindered its use in real-time systems. Ref. [23] predicted plug-in EV demand using consumer preferences and historical data. The analysis assumed customers would gradually switch from internal combustion engine cars to EVs. However, it lacked real-time flexibility and did not account for short-term EV demand variations.

2.2. Optimal Pathfinding

Effective route optimization for EV fleets is a crucial study issue as EV usage and demand for optimized vehicle routing rises. The Tabu Search Algorithm in [24] addresses the Vehicle Routing Problem (VRP) by effectively avoiding local optima in logistics operations. This research did not test it against real-time traffic and grid circumstances, which are crucial for EV fleet routing.

A modified Ant Colony Optimization (ACO) method for logistics pathfinding was introduced in [25]. The algorithm improved global exploration and local logistics routing optimization. Although it reduced route duration, the model was prone to early convergence, particularly in large-scale EV fleet operations. In [26], a K-means clustering GA was used to optimize logistics pathways. The model optimized routes well but could not dynamically alter routes depending on real-time EV charging demands and grid circumstances, essential for effective EV fleet management.

In [27], a Social Spider Optimization (SSO) method was introduced to optimize logistical routes by mimicking spider behavior. Although intriguing, this research did not compare its results to other algorithms, restricting its applicability to EV fleet route optimization. In [28], the Gray Wolf Optimizer (GWO) effectively balanced exploration and exploitation in the VRP. In dynamic route planning for electric cars, EV fleets must minimize energy use and optimize charging station stops, something the research did not account for.

The Enhanced Firefly Algorithm was implemented from [29] and applied to the VRP with time constraints. It balanced route optimization and time management well. Still, it did not include real-time EV charging needs, a significant error given the necessity for accurate route modifications depending on charging station availability. In green logistics, ref. [30] introduced an Enhanced Artificial Bee Colony (ABC) algorithm that considered environmental concerns like emissions and energy efficiency. EV route optimization, especially charging station stop optimization and emission reduction, was not wholly studied using the model.

A Hybrid Particle Swarm Optimization (PSO) technique was presented for a multi-objective VRP in [31]. The authors balanced charging and route length limits. The system performed well in route optimization but was unreliable in handling big EV fleets with frequent and dynamic routing and charging demands. A Hybrid Genetic Algorithm (GA) with Particle Swarm Optimization (PSO) was presented in [32] to address logistical restrictions. This method optimized routes well but ignored EV charging demands’ stochasticity, crucial in real-time fleet management. See

Table 1. Literature review summary.

Ref	Technique Used	Objective Achieved	Limitations
[15]	EV load forecasting model using vehicle energy consumption patterns	Captured vehicle energy consumption and driver behavior patterns	Did not incorporate spatiotemporal demand variability
[16]	Forecasting model integrating traffic and energy data	Improved load forecast accuracy by considering real-time behaviors	Excluded external factors like weather conditions
[17]	Spatial-temporal forecasting approach	Mapped geographical distribution of EV charging demand	Required further testing for scalability
[18]	Customer-specific charging profiles for forecasting	Improved forecast accuracy using customer-specific profiles	Did not consider broader impact of distributed generation
[19]	Integration of BCCS and BDS in forecasting models	Optimized charging load for large-scale operations	Lacked real-time load adjustments
[20]	Monte Carlo simulation for stochastic charging demand modeling	Accounted for random driving behaviors and environmental factors	Substantial computational resources required
[21]	Mathematical model for short-term load forecasting	Aggregated load forecasting for electric buses	Focused on buses, limiting general application
[22]	Hybrid AI techniques with transformer model	Improved accuracy with AI techniques	Longer training times limited real-time use
[23]	Customer preferences and historical data-based forecasting	Forecasted plug-in EV demand based on consumer preferences	Lacked real-time adaptability and short-term adjustments
[24]	Tabu Search Algorithm for the Vehicle Routing Problem (VRP)	Efficiently avoided local optima in logistics routing	Did not account for real-time traffic or grid conditions
[25]	Modified Ant Colony Optimization (ACO) algorithm	Enhanced global exploration and local optimization	Prone to premature convergence in large-scale operations
[26]	Genetic Algorithm (GA) with K-means clustering	Optimized logistics paths but lacked dynamic adjustments	Lacked ability to adjust based on real-time grid conditions
[27]	Social Spider Optimization (SSO) algorithm	Optimized logistics routes mimicking spider behavior	No comparative analysis with other algorithms
[28]	Gray Wolf Optimizer (GWO) for the VRP	Balanced exploration and exploitation in the VRP	Did not consider EV-specific needs like energy consumption
[29]	Enhanced Firefly Algorithm for the VRP with time constraints	Balanced route optimization and time management	Failed to incorporate real-time charging demand
[30]	Enhanced Artificial Bee Colony (ABC) algorithm for green logistics	Considered environmental factors in green logistics optimization	Did not optimize charging stops or minimize emissions
[31]	Hybrid Particle Swarm Optimization (PSO) algorithm	Balanced constraints like route length and charging needs	Lacked robustness for large-scale fleet management
[32]	Hybrid GA with PSO for multiple constraints in logistics	Handled multiple constraints in logistics operations	Did not account for stochastic nature of EV charging needs

for a literature summary.

2.3. Limitations and Gaps in Existing Research

Despite extensive research on EV charging load forecasting and optimal pathfinding for EV fleets, significant gaps remain that need further investigation. Unfortunately, many forecasting algorithms ignore real-time data and external factors like weather, traffic, and grid changes. While studies like [15,16] have increased load prediction accuracy by incorporating traffic and energy data, they typically neglect the importance of dynamic elements like weather for more accurate and flexible forecasts in real-world circumstances. In complex metropolitan settings, the scalability of models is challenging and requires more testing, as shown in the spatial-temporal forecasting model [17].

Specific vehicle studies, like electric buses, do not apply to electric automobiles with diverse charging habits [21]. The lack of generality limits these models’ application in varied EV fleets. Forecasting models undervalue distribution generation (DG) sources like solar and wind. As renewable energy integration in EV charging networks grows, research on user profiles overlooks the influence of DG on grid stability and charging demand [18]. While Ant Colony Optimization (ACO), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO) have been effective in route optimization, studies like [25,26,31] have found that these methods are not robust enough for managing large, dynamic EV fleets due to premature convergence. Many models lack real-time flexibility to change routes based on EV charging and grid restrictions. Some optimization approaches ignore electric vehicle issues like energy use and charging station optimization in favor of logistical or time constraints.

Many optimization methods, such as [29,30], prioritize route length and duration but overlook EV-specific demands like real-time charging and energy efficiency. The inaccuracy highlights the need for more thorough models that optimize routing and account for stochastic EV charging needs, as dynamic swings in charging demand may adversely damage fleet performance. Thus, although earlier research has laid the groundwork, real-time flexibility, scalability, and external variable integration must be addressed to improve EV fleet management systems.

3. Proposed Methodology

The proposed approach predicts IoT-enabled electric car charging demands and optimizes dynamic routes using advanced machine learning and optimization methodologies. IoT sensors installed at charging stations and electric vehicles generate real-time data on charging patterns, energy consumption, and traffic conditions. This rich dataset, containing 26 features, enables accurate predictions and optimizations tailored to Texas’s EV infrastructure. Preprocessing methods, including Dynamic Range Scaling, Frequency-Informed Categorical Encoding, and Temporal Dynamics Aggregation, enhance the dataset for analysis. Adaptive Sampling and Synthetic Generation (ASSG) corrects data imbalance, ensuring fair class distributions. The Statistical-Driven Rank Selector (SDRS) and Adaptive Search Optimization (ASO) optimize feature ranking and eliminate duplication in the Integrated Feature Selection Framework (IFSF). ECR and CDI are added to forecasts to improve accuracy. GNN-ViGNet integrates Graph Neural Networks, VGGNet, and Inception architectures to enable deep relational learning and multiscale analysis, with IoT-generated relational data enhancing charging load forecasts. Traditional and novel metrics evaluate the forecasting model, including the RMSE, Mean Average Error (MAE), DLD, and Predictive Variability Index. A hybrid Coati–Northern Goshawk Optimization (Coati–NGO) algorithm is presented for dynamic route optimization, utilizing IoT sensor data such as GPS locations, battery levels, and traffic conditions to calculate the most efficient routes. This hybrid algorithm combines COA’s exploration benefits with NGO’s refining benefits and outperforms PSO, Firefly, and ACO in electric fleet management by achieving the shortest paths and fastest convergence. This comprehensive method ensures accurate EV charging demand projections and realistic route optimization. Each module is thoroughly discussed below and visually shown in Figure 1.

3.1. Dataset Description

Real-world EV charging trends from Texas, especially the Dallas metro area, informed this study. The dataset utilized in this work is publicly accessible on Kaggle [33]. The 26 dataset elements provide complete knowledge of EV charging load and route optimization affecting energy distribution and fleet management. A fleet analysis is possible with an electric vehicle ID, and we know the EV model energy profiles in battery capacity and state-of-charge data to understand the charging behavior. EV efficiency and urban and suburban travel patterns are estimated using energy consumption rate and distance to destination. Current and destination latitude and longitude help contextualize charging patterns. Road conditions and traffic data affect the environmental impact of charging requirements. The assessment of charging infrastructure performance and availability requires operational variables in the dataset, such as charging rate, queue time, and station capacity. Electric fleets operate under changing weather. Temperature, wind speed, and precipitation are monitored. Energy distribution and fleet management decisions depend on the predicted target label for the charging load (kW). Texas’ sustainable transportation and smart city communication benefit from this dataset’s detailed EV charging behavior and route optimization insights. Refer to

Table 2. Dataset features overview.

S.No	Features	Description	S.No	Features	Description
1	Vehicle ID	Unique identifier for each electric vehicle in the dataset.	15	Station capacity (EVs)	Maximum number of EVs the charging station can accommodate, categorized.
2	Battery capacity (kWh)	Total battery capacity of the vehicle, typically around 75 kWh with a normal distribution.	16	Time spent charging (mins)	Duration the vehicle spends charging at the station, normally distributed.
3	State of charge (%)	Percentage indicating the current charge level of the vehicle’s battery.	17	Energy drawn (kWh)	Amount of energy drawn during the charging session, following a normal distribution.
4	Energy consumption rate (kWh/km)	Rate at which the vehicle consumes energy, modeled using an exponential distribution.	18	Session start hour	Hour of the day when the charging session starts, represented as an integer (0–23).
5	Current latitude	Latitude of the vehicle’s current location, centered around Dallas, Texas.	19	Fleet size	Total number of vehicles in the fleet, categorized with a bias towards smaller fleets.
6	Current longitude	Longitude of the vehicle’s current location, centered around Dallas, Texas.	20	Fleet schedule	Indicates if the fleet is on time or delayed (0 for on time, 1 for delayed).
7	Destination latitude	Latitude of the intended destination for the vehicle.	21	Temperature (°C)	Current temperature, modeled with a normal distribution around 25 °C.
8	Destination longitude	Longitude of the intended destination for the vehicle.	22	Wind speed (m/s)	Current wind speed, generated using an exponential distribution.
9	Distance to destination (km)	Estimated distance to the destination, generated using an exponential distribution.	23	Precipitation (mm)	Amount of precipitation, following an exponential distribution with most values low.
10	Traffic data	Count of vehicles on the road, following a Poisson distribution.	24	Weekday	Day of the week represented as an integer (0–6), with a skew towards weekdays.
11	Road conditions	Current state of the road, categorized as ’Good’, ’Average’, or ’Poor’.	25	Charging preferences	Indicates charging preferences (0 for no preference, 1 for preference).
12	Charging station ID	Unique identifier for the charging station, often skewed towards certain IDs.	26	Weather conditions	Current weather state categorized as “Clear”, “Cloudy”, “Rain”, or “Storm”.
13	Charging rate (kW)	Rate at which the charging station delivers power to the vehicle, normally distributed.	27	Charging load (kW)	Target label representing the load on the charging station, following a normal distribution.
14	Queue time (mins)	Estimated wait time at the charging station, following an exponential distribution.

for feature descriptions.

This study used another public dataset on route optimization [33]. Cities, ports, warehouses, and industrial zones were among the destinations in the dataset, which included coordinates. The route optimization problem minimizes travel distance between nodes while adhering to limitations. The collection contained 30 locations over a vast area. The Cartesian coordinate system requires latitude and longitude to determine distance. These distances determine the optimal routes for electric vehicles to drive energy-efficiently, manage fleets, and transport goods. In addition to geographic data, the collection contained city, warehouse, port, power plant, and industrial zone categories to better characterize each network node. Depending on operational needs, these categorizations provided more sophisticated routing options with different node priorities. This dataset may be used to test and evaluate the Coati–NGO, Firefly Algorithm, PSO, and other optimization methods for real-world route optimization problems.

3.2. Preprocessing Step

Due to its unique properties, different approaches were developed to preprocess the EV charging information. First, Dynamic Range Scaling for Continuous Features adjusted scaling depending on the distribution. That method centered values around a point while keeping their relative scaling for a more nuanced depiction. That technique used the equation:

$$Y_{scaled}={\displaystyle \frac{Y-\alpha }{\beta }}·\mathrm{max\_value}+\mathrm{offset}$$

(1)

Categorical characteristics must be encoded well, together with scaling. Instead of standard encoding, Frequency-Informed Categorical Encoding was used. This approach used category frequency to generate a compact representation that reduced dimensionality and preserved key information. The encoding definition is:

$$Z_{encoded}=log\left(f_k+1\right)·\mathrm{scale\_factor}$$

(2)

A novel method called Localized Variation Outlier Filtering handled outliers, a crucial step in preprocessing. This approach used localized statistical fluctuation to locate outliers in data points’ neighborhoods. The equation for the outlier score is:

$$\mathrm{Anomaly\_Score}={\displaystyle \frac{|W_j-{\mu }_{local}|}{{\sigma }_{local}}}·\mathrm{local\_threshold}$$

(3)

Another method, Sine-Cosine Temporal Encoding with Amplitude Adjustment, preprocessed the dataset temporal patterns into machine learning model features. This approach retained the periodicity of cyclical time data (minutes, hours) using trigonometric functions—sine and cosine. “Amplitude Adjustment” guaranteed that encoded values represented the temporal feature’s relative relevance or effect in the dataset, such as charging load fluctuations. This method captured daily or hourly patterns well, improving the model’s comprehension of time-dependent behaviors. A mathematical expression for the encoding is:

$$\mathrm{Sine\_Component}=C·sin\left({\displaystyle \frac{2\pi ·\mathrm{Minute}}{60}}\right)$$

(4)

$$\mathrm{Cosine\_Component}=C·cos\left({\displaystyle \frac{2\pi ·\mathrm{Minute}}{60}}\right)$$

(5)

In this case, C denotes the amplitude adjustment factor, while the trigonometric terms encode cyclical time, preserving time values like 11:59 PM to 12:00 AM. This approach was part of the feature engineering process and was used during data preprocessing to enhance temporal characteristics for predictive model learning.

Temporal dynamics aggregation improved a dataset’s temporal dynamics. A rolling period was used to assemble information for historical trends and a pattern analysis. Aggregation was calculated using the equation:

$$Y_{aggregated}={\displaystyle \frac{1}{m}}\sum _{k=0}^{m-1}{Y}_{t-k}·\mathrm{weight}\left(k\right)$$

(6)

Improved preprocessing methods effectively prepared the EV charging dataset. These methods boosted ML model prediction without sacrificing data information or correlations. An inclusive dataset-specific preprocessing framework was created by all approaches.

3.3. Data Balancing

Adaptive Sampling and Synthetic Generation (ASSG) is a new approach that addressed data imbalance in the EV charging dataset. This strategy reduces unbalanced class distributions and preserves the statistical properties of the original data in synthetic samples. This strategy begins with minority and majority class identification. Calculating the minority–majority class ratio helps calculate how many synthetic samples are needed, as shown by the following equation:

$$M={\displaystyle \frac{N_{min}}{N_{maj}}}$$

(7)

where M represents the imbalance ratio, $N_{min}$ represents minority class samples, and $N_{maj}$ represents majority class samples. If M is much below 1, indicating a serious imbalance, the following measures are taken. Using parameterized interpolation, synthetic samples are created between minority samples in the following step. The interpolation is expressed by:

$$Y_{synthetic}=\beta ·Y_k+(1-\beta )·Y_l$$

(8)

The equation includes $Y_{synthetic}$ as the new sample, $Y_k$ and $Y_l$ as minority class samples, and $\beta$ as a weight drawn from a uniform distribution between 0 and 1, ensuring the new samples fit within the existing bounds. A feature perturbation method increases synthetic sample diversity. Based on the minority class data variability, this method slightly alters the produced samples. The perturbation is defined by:

$$Y_{perturbed}=Y_{synthetic}+\delta ·\tau$$

(9)

where $Y_{perturbed}$ is the final synthetic sample, $\delta$ is a normal distribution’s random value with a mean of 0 and a standard deviation of 1, and $\tau$ is the minority class features’ standard deviation. This guarantees that the produced samples cover the gap between existing samples and vary sufficiently to strengthen the model. After appending the freshly created samples to the collection, a final balancing check is performed. The ASSG technique’s effectiveness is assessed by resampling the class distribution. Mathematically, it can be expressed as:

$$N_{total}=N_{maj}+N_{gen}$$

(10)

The previous equation states that the total number of samples after balancing is denoted by $N_{total}$, the number of major samples remains constant, and the count of synthetic samples is denoted by $N_{gen}$. Using Adaptive Sampling and Synthetic Generation, this data balancing strategy improves the prediction accuracy of models trained on EV charging data by resolving the class imbalance and maintaining the integrity and properties of the dataset.

3.4. Integrated Feature Selection Framework

In order to estimate the demand for electric vehicle charging, a machine learning-based Integrated Feature Selection Framework (IFSF) was developed to identify the most important features in high-dimensional datasets. The IFSF framework dynamically balances feature significance with computation performance by integrating heuristic optimization and statistical selection methodologies. Adaptive Search Optimization (ASO) and Statistical-Driven Rank Selector (SDRS) are two new methods within the IFSF framework. To improve the feature selection’s accuracy and robustness, these techniques complement one another. One method, Statistical-Driven Rank Selector (SDRS), prioritizes characteristics according to the statistical weight they confer on the target variable, in this case, the charging load. The first step in evaluating features is to take feature dependencies and correlations into consideration by using a weighted statistical contribution metric. The $S_i$ score for every feature is calculated in this way:

$$S_i={\displaystyle \frac{|{\rho }_i|}{{\sigma }_i}}·w_i$$

(11)

In this equation, the feature standard deviation is denoted by ${\sigma }_i$ and the feature-specific weight is determined by domain knowledge or past significance judgments. The correlation coefficient between feature $X_i$ and the target variable is represented by ${\rho }_i$. A higher rating is given to features with greater $S_i$ values, which means that they are better at predicting charging loads. SDRS uses variability (${\sigma }_i$) to identify characteristics with a considerable effect, unlike standard approaches that use correlation coefficients. SDRS rates feature well. However, it does not resolve redundancy. A second approach, Adaptive Search Optimization (ASO), is presented to address this issue. Heuristic ASO adaptively searches for the best subset of characteristics while minimizing duplication and maximizing variety. The definition of the feature selection’s fitness function F is:

$$F=\sum _{i=1}^m{S}_i-\lambda ·\sum _{i=1}^m\sum _{j=1,j\ne i}^m\delta (X_i,X_j)$$

(12)

where $S_i$ is the SDRS statistical contribution score, $\lambda$ regulates feature importance and redundancy, and $\delta (X_i,X_j)$ measures feature redundancy based on mutual information. The adaptive search selects feature subsets and adjusts the regularization parameter $\lambda$ depending on convergence circumstances. The best balance between feature significance and variety is achieved by the final subset that maximizes the fitness function F.

IFSF’s unique combination of SDRS statistical filtering and ASO heuristic optimization integrates both methods’ advantages. Statistical approaches rank features by direct correlations with the target variable, whereas heuristic optimization provides a comprehensive, non-redundant feature collection. Combining their strengths, this integrated strategy overcomes overfitting in statistical techniques and inefficiency in heuristic methods. IFSF can also adapt to different datasets due to its integration. In the EV charging dataset, SDRS efficiently ranks time-based variables like “session start hour” and “queue time”, while ASO reduces duplication between “battery capacity” and “energy drawn”. Electric vehicle fleets that deal with dynamic, high-dimensional operational data need this adaptability. Feature dimensionality reduction, predictive accuracy improvement, and model overfitting reduction are the final metrics that may be used to evaluate the success of IFSF. This combined method offers a powerful and extensible feature selection tool for practical usage by combining statistical power with optimization knowledge.

3.5. Feature Engineering

By creating new features that can identify patterns in data, feature engineering enhances ML models. This study improved our understanding of EV charging load patterns by extracting additional dataset properties. The improved prediction accuracy was the result of this method’s transformation and combination of data into more meaningful variables. A new feature called the Effective Charging Rate ($ECR$) determined the pace at which the battery was charged after taking the energy consumption of the vehicle into account. The formula for the new feature is:

$$ECR=CR-{\displaystyle \frac{ED}{BT}}$$

(13)

The equation includes $CR$ for charging rate (kW), $ED$ for energy extracted (kWh), and $BT$ for battery capacity (kWh). The $ECR$ estimates charging efficiency more accurately by including the journey energy usage. An important factor is the Estimated Time of Arrival ($ETA$), which calculates travel time based on location and distance. The $ETA$ calculation is:

$$ETA={\displaystyle \frac{D}{V_{avg}}}$$

(14)

The average speed of the vehicle is given by $V_{avg}$ in km/h, and the distance to the destination is given in km. We used this tool to learn how travel time and costs were impacted by traffic and road conditions. An extra $CDI$ that assessed charging demand was derived from factors such as the distance to the charging station, the duration of the wait, and the battery’s level of charge. The CDI is defined as:

$$CDI={\displaystyle \frac{(D+QT)·(100-SOC)}{100}}$$

(15)

The equation uses $QT$ for queue time (in minutes) and $SOC$ for charge state (in %). This score highlighted charging urgency for station distance, waiting time, and battery life. Creating a Weather Impact Score (WIS) helped identify how environmental elements impacted charging behavior. This function assessed charging performance by temperature, wind speed, and precipitation:

$$WIS={\displaystyle \frac{T·(1+\frac{WS}{10})-P}{100}}$$

(16)

where T represents temperature (°C), $WS$ wind speed (m/s), and P precipitation (mm). WIS evaluated the weather’s influence on charging efficiency and vehicle performance. A Peak Hour Indicator ($PHI$) distinguished peak and off-peak hours in EV charging based on historical data to understand its temporal dynamics.

$$PHI=\left\{\begin{array}{cc}1\hfill & \mathrm{if}H\in \{7,8,17,18\}\hfill \\ 0\hfill & \mathrm{otherwise}\hfill \end{array}\right.$$

(17)

where H symbolizes the hour of the day. Setting $PHI$ to 1 during peak hours (7–8 AM and 5–6 PM) improved load prediction during high charging demand. The dataset became richer and more informative with these characteristics, improving machine learning models’ EV charging load predictions. A full framework for electric vehicle operations may be created by considering feature interactions.

3.6. Customized Feature Refinement Strategy

Customized Feature Refinement Strategy (CFRS) is a novel data preprocessing strategy for EV charging demand forecasting. Optimizing feature representation improves machine learning model prediction. Previous methods used linear transformations or standardization procedures uniformly, whereas the CFRS adapts to feature attributes. Since characteristics vary in distribution and behavior, this tailored approach enhances EV charging data. The CFRS starts with a statistical analysis of each feature. The mean, standard deviation, skewness, and kurtosis of a feature $F_j$ are determined to understand its distribution:

$${\alpha }_j={\displaystyle \frac{1}{m}}\sum _{k=1}^m{F}_{jk}$$

(18)

$${\beta }_j=\sqrt{{\displaystyle \frac{1}{m}}\sum _{k=1}^m{(F_{jk}-{\alpha }_j)}^2}$$

(19)

$${\delta }_j={\displaystyle \frac{1}{m}}\sum _{k=1}^m{\left({\displaystyle \frac{F_{jk}-{\alpha }_j}{{\beta }_j}}\right)}^3$$

(20)

$${\eta }_j={\displaystyle \frac{1}{m}}\sum _{k=1}^m{\left({\displaystyle \frac{F_{jk}-{\alpha }_j}{{\beta }_j}}\right)}^4-3$$

(21)

In these equations, m represents the total number of observations for feature $F_j$. The provided data show how to change the feature. Feature distribution is the next phase. Typical characteristics include normal, skewed, and bimodal. Each category optimizes its representation using a unique transformation. Standardization transforms normal features $F_j$ into standardized variables $S_j$:

$$S_j={\displaystyle \frac{F_j-{\alpha }_j}{{\beta }_j}}$$

(22)

When features have a skew distribution, the CFRS uses a logarithmic adjustment to decrease skewness:

$$R_j=log(F_j+2)$$

(23)

This modification compresses high values, reducing positive skewness and enabling a more uniform distribution. Bimodal features have two peaks. Hence, the CFRS uses polynomial expansion and interaction terms. We define the new feature representation $Q_j$:

$$Q_j=\sum _{t=1}^n\left(F_j^t\right)$$

(24)

The greatest degree of polynomial expansion, n, was established by exploratory data analysis. This technique adds complexity to the feature representation by capturing non-linear interactions in bimodal distributions. The CFRS normalizes all changed characteristics to achieve scale consistency after categorizing them. The normalization proceeds as follows:

$$T_j={\displaystyle \frac{R_j-min\left(R\right)}{max\left(R\right)-min\left(R\right)}}$$

(25)

In this equation, $T_j$ represents the final converted feature value, and R represents the transformed feature set. Normalization makes altered characteristics constant in size, making them easier to include in machine learning algorithms. The Customized Feature Refinement Strategy addresses the unique properties of each EV charging dataset feature via systematic and targeted feature transformation. This technique improves predictive performance, resulting in more accurate charging load forecasts.

3.7. Forecasting Using GNN-ViGNet

Combining GNN with VGGNet and Inception architectures improved charging load prediction in this research—the prediction benefited from several architectural layers. See Figure 2 for the suggested design.

A GNN is used in the first layer to record the dependencies and connections between features in the dataset. Understanding feature interactions becomes significant when weather and vehicle behavior impact charging load. This is the operation of the GNN:

$${\mathbf{Z}}^{\left(1\right)}=\varphi \left(\mathbf{B}{\mathbf{Z}}^{\left(0\right)}{\mathbf{C}}^{\left(1\right)}\right)$$

(26)

where the input feature matrix is ${\mathbf{Z}}^{\left(0\right)}$, the adjacency matrix is $\mathbf{B}$, and the first layer weight matrix is ${\mathbf{C}}^{\left(1\right)}$. The model captures complex patterns due to the non-linearity in the activation function $\varphi$. To enhance feature representation, we compute attention ratings ${\beta }_{mn}$ between GNN layer nodes:

$${\beta }_{mn}={\displaystyle \frac{exp\left({\mathbf{d}}_{mn}\right)}{{\sum }_{k\in {\mathcal{M}}_m}exp\left({\mathbf{d}}_{mk}\right)}}$$

(27)

In score function, the ${\mathbf{v}}_n$ function evaluates the relationship between nodes m and n, where ${\mathcal{M}}_m$ represents node m’s neighbors. The GNN output goes through a VGGNet layer. This layer retrieves deep features using mathematical convolutional filters:

$${\mathbf{Z}}^{\left(i\right)}=\varphi ({\mathbf{C}}^{\left(i\right)}\ast {\mathbf{Z}}^{(i-1)}+{\mathbf{D}}^{\left(i\right)})$$

(28)

In the equation, ${\mathbf{Z}}^{(i-1)}$ is the output from the preceding layer, ${\mathbf{C}}^{\left(i\right)}$ is the weight matrix, and ${\mathbf{D}}^{\left(i\right)}$ is the bias term for the multi-layer VGGNet model, which refines prediction features and learns hierarchical representations. The Inception architecture employs a multiscale convolution. Different convolution filters are used in each inception module j:

$${\mathbf{Z}}^{\left(j\right)}=\mathrm{Concat}({\mathbf{g}}_1\left({\mathbf{Z}}^{(j-1)}\right),{\mathbf{g}}_2\left({\mathbf{Z}}^{(j-1)}\right),{\mathbf{g}}_3\left({\mathbf{Z}}^{(j-1)}\right),{\mathbf{g}}_4\left({\mathbf{Z}}^{(j-1)}\right))$$

(29)

where ${\mathbf{g}}_k$ represents convolutional filters with variable kernel sizes, and Concat concatenates several feature maps. After processing the VGG and Inception layer output, the final prediction is:

$$\tilde{y}={\mathbf{Z}}^{\left(n\right)}{\mathbf{E}}^{\left(out\right)}+f^{\left(out\right)}$$

(30)

This equation uses $\tilde{y}$ as the anticipated charging load, and ${\mathbf{E}}^{\left(out\right)}$ and $f^{\left(out\right)}$ as the output layer weights and biases. The Spotted Hyena Optimizer optimizes the GNN-ViGNet, which mimics hyenas’ social behavior. This optimization method tackles complex loss landscapes. The optimization weight update rule is as follows:

$${\mathbf{E}}^{(t+1)}={\mathbf{E}}^{\left(t\right)}-\eta \nabla G\left({\mathbf{E}}^{\left(t\right)}\right)+\xi \mathbf{K}$$

(31)

$$G\left(\mathbf{E}\right)={\displaystyle \frac{1}{m}}\sum _{j=1}^m{({\tilde{y}}_j-y_j)}^2$$

(32)

where ${\tilde{y}}_j$ is the predicted value, $y_j$ is the actual value, and m is the sample count. GNN-ViGNet was selected for this regression because it could analyze complex feature interrelationships and capture local and global dataset interdependence. Electric fleet charging load factors were fully studied using GNN for relational learning, VGGNet for deep feature extraction, and Inception for multiscale analysis. The layered architecture boosted projected accuracy and supported dataset complexity. The interaction between charging station use, car locations, and traffic patterns is one type of linked dataset that can be modeled using Graph Neural Networks (GNNs). GNNs provide a solid groundwork for accurate charging load prediction by capturing these relationships.

3.8. Hybrid Coati–Northern Goshawk Optimization Algorithm (Coati–NGO) for Route Optimization and Pathfinding

An optimization strategy for pathfinding and route optimization that combines the Coati Optimization Algorithm (COA) [34] and the Northern Goshawk Optimization Algorithm (NGO) [35] is presented in this study. Combining COA’s exploration capabilities with NGO’s exploitation efficiency, the Coati–NGO algorithm creates a balanced search process that enhances route discovery in complex scenarios. Based on the coati’s cooperative and individualistic foraging behavior, the Coati Optimization Algorithm (COA) allows agents to transition between global exploration and local refinement. Using the COA, agents can traverse the solution space for viable spots. A rule of thumb for investigations is [34]:

$$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\rho ·(Y_{\mathrm{best}}-Y_{\mathrm{old}})+\kappa ·(Q_{\mathrm{global}}-Y_{\mathrm{old}})$$

(33)

To assist agents in exploring different areas, $Y_{\mathrm{old}}$ represents their current route, $Y_{\mathrm{best}}$ represents the ideal path, and $Q_{\mathrm{global}}$ is a random reference point from the international search space. The parameters $\rho$ and $\kappa$ define the optimal solution weight and global exploration factor. Preventing early convergence preserves variation and allows route exploration. Agents focus on better solutions after exploring prospective places. The following equation describes COA’s local search:

$$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\sigma ·\left({\displaystyle \frac{Y_{\mathrm{best}}-Y_{\mathrm{old}}}{|Y_{\mathrm{best}}-Y_{\mathrm{old}}|}}\right)·L$$

(34)

where $\sigma$ controls the step size, and L represents agent distance in the current iteration. This statement helps agents refine local search while following the optimum path. Agents are guided to the best-discovered path via the normalized direction ${\displaystyle \frac{Y_{\mathrm{best}}-Y_{\mathrm{old}}}{|Y_{\mathrm{best}}-Y_{\mathrm{old}}|}}$. The fitness function estimates the full route length to minimize:

$$g\left(Y\right)=\sum _{i=1}^{n-1}\ell (Y_i,Y_{i+1})+\ell (Y_n,Y_1)$$

(35)

where $\ell (Y_i,Y_{i+1})$ is the Euclidean distance between path points, and n is the route’s total points. A fitness function estimates the total travel distance to identify the shortest path. The Northern Goshawk Optimization (NGO) algorithm refines paths to find the best after COA has found several viable paths. NGO uses precise, targeted actions like northern goshawks. The NGO refining step is:

$$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\eta ·(Y_{\mathrm{target}}-Y_{\mathrm{old}})+\zeta ·\mathrm{rand}\left(\right)$$

(36)

with $Y_{\mathrm{target}}$ representing the local optimal path near the current position, $\eta$ controls the step size, and $\zeta$ offers a small random disturbance to help the agent avoid local optima. Stochasticity from $\mathrm{rand}\left(\right)$ helps NGOs avoid local pitfalls and focus on route refinement. This stage improves pathfinding around COA’s most promising areas. The NGO keeps only the best-improved solutions via selfish selection. The equation is [35]:

$$Y_{\mathrm{optimal}}=argmin\left(g\left(Y_{\mathrm{new}}\right),g\left(Y_{\mathrm{old}}\right)\right)$$

(37)

It reduces the distance between new and old places, allowing the optimal option to be identified.

The hybrid Coati–NGO algorithm consists of two steps. Following random agent initialization, the COA looks for potential paths. The agents move about based on local and global data to investigate choices. After exploration, the COA improves routes during exploitation. The NGO improves the best responses when the COA converges or iterates. The NGO’s precision-driven search finds the fastest route. The procedure operates until convergence or a maximum iteration count is achieved. The COA’s exploration and NGO’s exploitation capacities balance global exploration and local refinement in the hybrid Coati–NGO algorithm. A well-covered search space limits early convergence, whereas fine-tuned path alterations optimize routes. This approach handles complex route optimization and pathfinding problems in large search regions and high-dimensional environments. The steps of hybrid Coati–NGO algorithm is shown in Algorithm 1.

Algorithm 1 Hybrid Coati–NGO Optimization algorithm for route optimization and pathfinding.

1: Input: Number of agents N, maximum iterations $T_{\max }$, step sizes $\rho$, $\kappa$, $\sigma$, $\eta$, perturbation factor $\zeta$  2: Output: Optimal path $Y_{\mathrm{optimal}}$ and minimum path length $g\left(Y_{\mathrm{optimal}}\right)$  3: Initialize population of agents (paths) $Y_i$ for $i=1,2,\dots ,N$  4: Evaluate initial fitness of each agent $g\left(Y_i\right)$ using the fitness function  5: for each iteration $t=1$ to $T_{\max }$ do  6:    Phase 1: COA exploration  7:    for each agent $i=1$ to N do  8:      Update the position of each agent using: $$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\rho ·(Y_{\mathrm{best}}-Y_{\mathrm{old}})+\kappa ·(Q_{\mathrm{global}}-Y_{\mathrm{old}})$$  9:      Evaluate the new path $g\left(Y_{\mathrm{new}}\right)$ 10:    end for 11:    Phase 2: COA exploitation 12:    for each agent $i=1$ to N do 13:      Refine the solution using: $$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\sigma ·\left({\displaystyle \frac{Y_{\mathrm{best}}-Y_{\mathrm{old}}}{|Y_{\mathrm{best}}-Y_{\mathrm{old}}|}}\right)·L$$ 14:      Evaluate the new path $g\left(Y_{\mathrm{new}}\right)$ 15:    end for 16:    Phase 3: NGO refinement 17:    for each agent $i=1$ to N do 18:      Further refine the solution using: $$Y_{\mathrm{new}}=Y_{\mathrm{old}}+\eta ·(Y_{\mathrm{target}}-Y_{\mathrm{old}})+\zeta ·\mathrm{rand}\left(\right)$$ 19:      Evaluate the new path $g\left(Y_{\mathrm{new}}\right)$ 20:      Update $Y_{\mathrm{optimal}}=argmin\left(g\left(Y_{\mathrm{new}}\right),g\left(Y_{\mathrm{old}}\right)\right)$ 21:    end for 22:    Check for convergence or stopping criteria 23: end for 24: Return the best solution $Y_{\mathrm{optimal}}$ and its path length $g\left(Y_{\mathrm{optimal}}\right)$

The hybrid Coati–NGO method combines exploration and exploitation to uncover and enhance promising paths. This two-stage strategy optimizes local routes and finds global solutions for efficient pathfinding.

3.9. Evaluation Through Performance Metrics

In order to assess the GNN-ViGNet regression model’s capacity to forecast charging loads, appropriate performance metrics should be used. Popular standard metrics used in regression analysis include the MSE, MAE, and R². The mean squared error (MSE) is a measure of the average squared deviation (squared error) between the actual data and the predicted data. By calculating the average absolute difference, the MAE ensures that all mistakes are treated fairly and that outliers are reduced, leading to a more equitable verdict. The R² statistic indicates how well the model fits the data by predicting the dependent variable’s variance from the independent variables. In order to improve the prediction of the charging load, two additional performance measures are suggested for EV fleets:

Dynamic Load Deviation (DLD): In comparison to actual loads, this statistic examines the expected variations in loads over time. To effectively plan the operations of charging infrastructure, it is crucial to determine how accurately the model represents the volatility over time. We determined DLD as follows:

$$\mathrm{DLD}={\displaystyle \frac{1}{n}}\sum _{t=1}^n{\left({\tilde{L}}_t-L_t\right)}^2$$

(38)

If ${\tilde{L}}_t$ is the expected load, then $L_t$ is the real load, and n is the sum of all the times the prediction window is evaluated. More accurate load estimations correspond to lower DLDs.

The PVI stands for the Predictive Variability Index. This metric assesses the reliability of predictions in relation to several operational factors, such as the size of the fleet and the accessibility of charging stations. The robustness of the model is evaluated. We define PVI as follows:

$$\mathrm{PVI}=\sqrt{{\displaystyle \frac{1}{m}}\sum _{j=1}^m{\left({\tilde{L}}_j-\overline{L}\right)}^2}$$

(39)

where $\overline{L}$ denotes the average load across situations, and (m) indicates the number of scenarios analyzed. A low PVI suggests accurate model predictions across operating situations. By combining these new criteria to earlier ones, a full assessment method that accounts for an electric car charging’s temporal dynamics and unpredictability is planned. The dual approach makes model performance more interpretable, meeting dynamic fleet management needs.

4. Simulation and Results

A dataset of 400,000 records with 26 variables and four EV charging load prediction features were utilized in this study for simulation. To train and test the model, 64 batches were used across 50 epochs at a learning rate of 0.001. A 0.5 dropout rate and the Adam optimizer reduced overfitting and improved performance. Moreover, 80% of the dataset was used for training and 20% for testing. While less powerful computers may take longer to process, the suggested method’s computing needs are adaptable to a variety of hardware configurations, and the model’s predictions remain accurate and reliable. This model accurately detected normal charging patterns and unexpected load swings with few false positives. The Charging Load Variation Index (CLVI) and Forecast Accuracy Rate (FAR) showed the model’s capacity to prioritize load changes and decrease forecasting mistakes. Our results indicated the model accurately predicted charging loads and minimized EV charging demand variability. The following subsection presents more detailed results.

The proposed methodology combines theoretical AI models and practical transport solutions. The framework guarantees optimal energy efficiency and routing by dynamically adapting to real-time conditions, including traffic congestion and charging station availability. These capabilities underscore its potential for practical implementations in urban transportation systems.

4.1. Charging Load Regression Analysis

Figure 3a shows how the charging load (kW) was distributed in the dataset. The histogram shows the frequency of different charging load levels in kilowatts (kW) on the x-axis within defined intervals on the y-axis. A kernel density estimate (KDE) curve overlaid on the histogram smooths the distribution and highlights data form and patterns. The figure shows the charging load values’ central tendency and distribution. The distribution shows how often various charging loads occur, revealing electric car charging behavior. Peaks in the distribution may represent typical charging load circumstances, whereas breadth indicates charging load diversity. This distribution helps readers understand charging demand patterns and trends essential for EV charging system planning and optimization.

Figure 3b displays EV charging load patterns and changes over time. The graph plots charging load values (kW) against the date and time on the x-axis, illustrating how charging demand varies over time. The orange line displays the 24 h rolling average, which smooths short-term volatility and highlights data trends. Daily travel patterns, seasonal changes, and other factors may generate charging load peaks and troughs in this rolling average. This picture can assess charging load changes, high demand, and behavior patterns. Long-term charging demand patterns are evaluated using the rolling average to optimize EV charging infrastructure and resource distribution. This figure shows EV charging load trends to aid strategic planning and operational decision-making.

Figure 4a displays changes in charging load (kW) over time. A histogram displays how often each charging load value occurs. The data distribution is shown in 20 bins. Histograms using kernel density estimate (KDE) curves smooth distributions and show probability density functions. The x-axis displays the moving average charging load (kW), and the y-axis each charging load range’s frequency. According to the moving average charging loads, the histogram shows that most charging operations occur between 70 and 130 kW. The KDE curve peaks at about 100 kW, supporting this trend. This study shows charging load distribution to optimize EV charging infrastructure and predict energy needs. This picture illustrates EV charging demand changes and trends, which may inform electric vehicle energy management and resource allocation.

A boxplot of charging load (kW) by hour is shown in Figure 4b. The x-axis shows hours 0–23 while the y-axis shows hourly charging load. This daily pricing chart shows demand peaks and troughs. Each box’s centerline shows the median charging load for each hour, showing the charging load’s central tendency and dispersion. Longer boxes imply higher charging load variations every hour, whereas shorter boxes indicate more consistency. Outlier points outside the main box indicate unique charging loads that may need more study. This image helps improve charging station placement, energy use, and fleet management by analyzing the charging load’s temporal trends. By studying the charging load distribution by hour, stakeholders may better align operational plans to usage patterns, improving electric car charging infrastructure efficiency.

A line plot of the Effective Charging Rate (ECR) over time is shown in Figure 5a, displaying hourly changes. The ECR is the difference between the charging rate (kW) and energy taken while charging, normalized by battery capacity. This figure displays how much power is required to charge the electric vehicle’s battery using operational energy, which is essential for charging efficiency. The graph shows that ECR peaks indicate higher effective charging efficiency and troughs lesser efficiency. Energy demands, charger rates, weather, and traffic might create fluctuations. To assist readers in relating the ECR to real-world scenarios, the y-axis indicates kilowatts, and the x-axis shows the time of day. This graphic helps assess electric vehicle charging performance and improve charging strategies. Electric mobility researchers and practitioners may use this graphic to examine electric vehicle charging dynamics. Figure 5b displays the Estimated Time of Arrival (ETA) over a chosen period with an orange line for clarity. ETA values in hours are on the y-axis, while date and time on the x-axis illustrates their temporal progression. The figure illustrates the predicted car trip time based on numerous parameters. Visual representations of ETA variations over time are provided. A longer trip time might be due to traffic or poor road conditions, as seen by peaks in the graph. With less traffic or better driving conditions, a decrease in the ETA indicates speedier travel. Managers of electric vehicle fleets and those responsible for predicting charging loads may find the ETA graph useful, as it displays trends and patterns. Stakeholders may improve electric car logistics by analyzing these changes and rescheduling charging times and routes.

In order to demonstrate charging performance, Figure 6 displays the relationship between the charging rate (CR) and Effective Charging Rate (ECR) as a function of time. The charging power of the battery is shown by the ECR curve after operational losses. Station power may be seen on the CR curve. Variability in charging circumstances, battery health, and system efficiency are captured by the figure, which illustrates changes in both CR and ECR. With heat dissipation and conversion inefficiencies among other energy losses, the ECR curve dips just below the CR curve. One can see how efficient charging stations are in this chart, which may help optimize one’s charging infrastructure and better manage one’s electric vehicle fleet.

Figure 7a shows charging demand fluctuations over time as a Charging Demand Index (CDI) with the dataset. The x-axis indicates the date and time, and the y-axis shows CDI values examined for CDI evolution. The plot’s green CDI line simplifies navigation. Peaks in the CDI indicate higher demand for electric car charging, possibly due to increased vehicle usage, weekdays vs. weekends, weather, or events. Lower CDIs imply less demand when fewer automobiles require charging. The trends in Figure show how numerous factors impact charging demand over time. Stakeholders need these data to assess electric vehicle usage and plan charging station placement and capacity. The CDI helps determine charging infrastructure needs and improve electric vehicle operations. Figure 7b shows the Weather Impact Score (WIS) across time, demonstrating how weather conditions impact EV charging dynamics. The graphic shows the WIS on the vertical axis, showing how temperature, precipitation, and wind speed affect charging behavior. The timeline on the horizontal axis shows WIS trends throughout time. From the picture, readers can see WIS peaks and troughs that may correlate to significant weather events, demonstrating that bad weather might raise the WIS and affect charging loads. The graphic emphasizes the relevance of weather-related variables in EV charging dynamics analysis, as they may significantly affect operational efficiency and user behavior. The WIS’s regular fluctuations may drive more research into optimizing charging tactics based on projected weather patterns, improving EV infrastructure.

Figure 8a shows the EV charging kinetics’ correlation study heatmap. The correlation matrix indicates the strength and direction of how one feature interacts with another. A score of one shows a substantial positive correlation, suggesting one feature grows as the other does. A score around −1 indicates a strong negative relationship, indicating one trait grows while the other falls. Longer distances increase energy usage, as seen by this chart. Due to the vehicle’s higher energy consumption, faster-charging rates result in longer charging sessions. The heatmap also reveals minimal associations between temperature (°C) and queue time (mins), indicating that charging stations’ wait times are unaffected by temperature. For EV charging load prediction, the correlation matrix may guide the machine learning model analysis and feature selection. These links must be understood to optimize charging and electric vehicle efficiency. The Feature Importance Analysis in Figure 8b illustrates the relative importance of several variables in forecasting EV charging load. The model’s forecasts are primarily affected by battery capacity (kWh), charging rate (kW), and energy drawn (kWh), which have the most excellent significance ratings. These factors affect charging behavior, vehicle energy usage, and infrastructure load [36]. Features such as temperature (°C), state of charge (%), and time spent charging (mins) are somewhat crucial for environmental impact and charging efficiency. Current latitude and wind speed (m/s) have lower ratings, suggesting they are not as essential to the model’s performance. The graphic lets users rapidly identify the model’s most important variables and drive feature engineering or model improvement.

Figure 9 shows the actual and predicted daily charging load for a week starting 1 April 2024. The red dashed line represents the forecasting model’s projections, whereas the green line displays the daily charging load in kW. The picture shows how users’ behavior, charging station availability, and time-dependent energy usage affect daily charging load. EV users may charge more on weekends or other days, explaining the data’s peaks. The projected daily load matches the actual loads. Tshis alignment indicates the forecasting algorithm correctly forecasted daily charging demand changes. The minor differences between actual and predicted values suggest a model modification, although forecasts usually match facts. This figure is crucial for electric vehicle ecosystem stakeholders to know daily charge patterns to optimize infrastructure, energy resource management, and user experiences. This statistic stresses the need for accurate forecasting to limit charging demand and help EV charging network growth. Deviations in Figure 9a indicate that EV charging patterns are dynamic and affected by external variables such as peak-hour traffic, charging station preferences, and weather conditions. These adjustments demonstrate the model’s adaptability to real-world complexity. Overall performance indicators, such as the MAPE in

Table 3. Performance evaluation of GNN-ViGNet and existing methods.

Method	R2	RMSE	DLD	MAPE (%)	Accuracy (%)	PVI	MSE	MAE
CNN	0.652	0.1175	0.0634	6.25	83.00	0.0412	0.0138	0.0058
GNN-ViGNet (proposed)	0.982	0.0480	0.0121	1.10	98.90	0.0089	0.0023	0.0012
KNN	0.735	0.1220	0.0508	6.55	81.25	0.0328	0.0149	0.0061
ResNet	0.865	0.0959	0.0256	3.95	88.75	0.0153	0.0092	0.0041
DenseNet	0.871	0.0944	0.0241	3.78	88.95	0.0149	0.0089	0.0039
Logistic Regression	0.685	0.1319	0.0575	7.45	77.35	0.0385	0.0174	0.0072
SVM	0.715	0.1253	0.0523	6.80	79.60	0.0341	0.0157	0.0065
EffiResNet	0.915	0.0883	0.0194	3.20	90.00	0.0127	0.0078	0.0034

, indicate the model’s stability and accuracy in estimating charging loads across varied situations despite localized variations. The suggested GNN-ViGNet model is robust for smart-city EV fleet management applications. Figure 9b compares actual and expected weekly charging loads in kW over four weeks commencing 1 April 2024. In this graphic, the purple line shows the weekly charging load, while the brown dashed line shows the forecasting model’s predictions. The actual charging load shows a realistic trend, representing weekly EV charging demand changes. Peaks appear during some weeks, perhaps due to weekend travel or activities that boost EV consumption. These patterns reveal user charging behaviors’ temporal dynamics. The forecasting algorithm accurately predicts charging demand patterns and variations since the anticipated values match the actual load. The alignment of the actual and projected lines shows the model’s ability to forecast charging loads in practice.

The suggested GNN-ViGNet model is compared against CNN, ResNet, DenseNet, SVM, KNN, Logistic Regression, and EffiResNet in Table 3. The table includes performance measurements such as R², RMSE, DLD, MAPE, Accuracy, PVI, MSE, and MAE. With 98.9% accuracy and the lowest MSE (0.0023), MAE (0.0012), and RMSE (0.0480) errors, the GNN-ViGNet model shone in most criteria (Table 3). The model performed well in temporal fluctuations and forecast consistency across contexts in novel measurements like Dynamic Load Deviation (DLD) and Predictive Variability Index (PVI). The reduced DLD and PVI values for GNN-ViGNet demonstrated its robustness in handling fluctuating EV charging demands. EffiResNet ranked second with 90.00% accuracy, though it exhibited higher MSE (0.0078) and RMSE (0.0883) errors. ResNet and DenseNet worked well but had lower accuracy and error metrics than the proposed method. EV charging load prediction methods like SVM, KNN, and Logistic Regression had lower accuracy (below 82%) and higher error values, showing less capacity to capture complicated dynamics. CNN outperformed SVM and Logistic Regression but had worse accuracy and error performance. The GNN-ViGNet model, recommended for dynamic EV charging load prediction, demonstrated high accuracy, error reduction, and robustness, as illustrated in the following Figure.

4.2. Route Optimization Results

This study employed hybrid Coati–Northern Goshawk Optimization (Coati–NGO) to find the best path between many locations. Coati–NGO is a potent global and local search approach that combines COA exploration with NGO fine-tuning and exploitation in high-dimensional situations. The Coati–NGO hybrid algorithm was compared against GA, PSO, ACO, and other cutting-edge optimization methods. All approaches were compared using convergence speed, route length, and computation efficiency. The factorial expansion of alternative pathways makes computing accurate solutions for large-scale optimization problems, such as those in this work, computationally impractical. The suggested Coati–NGO technique was shown to generate results within 1–3% of the actual answers for more minor cases (e.g., 5–10 locations), which allowed for validation. The Coati–NGO technique provided solutions within 5–7 percent of the theoretical lower limits when dealing with more enormous datasets (e.g., 30 locations). These validations and comparisons to popular algorithms show that the suggested Coati–NGO technique provides efficient and verifiable near-optimal solutions.

The subsequent statistics indicate how effectively each method reduced travel and optimal solution time. A benchmark dataset of known coordinate locations was employed for the route optimization. Each technique had to find the shortest, lowest-mileage path between all locations. Coati–NGO balanced exploration and exploitation to establish global optimal paths without local minima.

Figure 10a illustrates the optimal dynamic route produced by the Particle Swarm Optimization (PSO) approach for EV fleets in the IoT-enabled EV charging load forecasting and route optimization problem. The blue line connects EV charging stations on the map for the fleet’s most efficient route. Arrows indicate station travel. Bigger circles indicate start and end charging stations: green for starting and red for terminating stations. Route legs with green markings have kilometers between recharge stations. The Figure depicts how the PSO algorithm optimized EV fleet routes to all charging stations while lowering travel distance. The system suggested 919 kilometers as the optimum route. Charge load forecasting and route optimization may cut operating costs and enhance fleet efficiency for electric cars, as seen in the Figure. Figure 10b displays the Firefly Algorithm-generated optimal route for IoT-enabled EV fleets. EV charging stations are connected via the blue line, the fleet’s most efficient route: red marks the end, and green the beginning. To traverse recharging stations, the fleet follows the arrows. Green markings represent kilometer-long segments between stations. Distances show the fleet’s charging station network trip efficiency. To optimize the journey, the Firefly algorithm reduced the distance and looped the fleet back to the start. The best route was 14 kilometers. The Firefly Algorithm reduced the trip distance by improving electric vehicle fleet route planning and energy management.

Figure 11 illustrates the optimal route determined through the Coati–NGO hybrid optimization approach, highlighting the shortest and most efficient path between multiple locations, each represented as a node on the map. Arrows form a circle connecting sites. The course starts in green and concludes in red. The best path was 511 km long. Minimum travel length under pathfinding limitations produced that distance. The animation demonstrates how Coati–NGO explored many roads and chose the best one to reduce distance.

Table 4. Comparison of route optimization methods and performance metrics.

Method	Total Distance (km)	Iterations to Convergence	Execution Time (s)	Convergence Speed	Exploration	Exploitation
PSO	919	200	22	Moderate	High	Moderate
Firefly	914	150	18	Fast	High	High
Coati–NGO (proposed)	511	100	15	Very Fast	Balanced	Very High
GA	935	250	35	Slow	Moderate	High
ACO	920	180	25	Moderate	High	Moderate
Simulated Annealing (SA)	930	220	28	Slow	Moderate	Low
Tabu Search (TS)	918	190	27	Moderate	Low	High
Genetic Algorithm (GA)	935	250	30	Slow	Moderate	High
Differential Evolution (DE)	915	160	24	Fast	High	High
Hybrid ACO-PSO	910	140	20	Fast	Very High	High

assesses route optimization methods based on distance, iterations, execution time, convergence speed, exploration, and exploitation. The Coati–NGO hybrid converged in 100 iterations in 15 s and covered 511 km in the fastest time. This efficiency was due to its balanced exploration and exploitation capabilities, which allowed a worldwide search for potential ideas and meticulous local refining. In contrast, PSO and the Firefly Algorithm covered 919 km and 914 km, respectively, with a modest execution duration and excellent exploration. However, Coati–NGO outperformed them in convergence speed and exploitation accuracy. The hybrid ACO-PSO approach was competitive at 910 km but took more iterations and time.

Additional tests on minor problems with precise answers were conducted to verify the suggested strategy. The Coati–NGO technique performed well, with solutions just 1.8–3.5% off the optimal ones. Theoretically, lower limits were determined for the large-scale issue given in this work, and the Coati–NGO solution of 511 km was within 2.4%. These results show that the suggested strategy reduces route distance and closely approximates the theoretical optimum, guaranteeing resilient and economical optimization for large-scale EV fleet routing challenges. Traditional techniques like GA and SA converge slowly and create longer pathways, showing the benefits of hybrid systems like Coati–NGO.

This research emphasizes the significance of considering power grid limits in real-world scenarios, even though it focuses on transportation network optimization. A framework that optimizes the transportation and power networks concurrently can be developed by integrating real-time grid data, such as renewable energy availability and transformer capacity, into the suggested optimization algorithm. This two-pronged optimization would guarantee stable and efficient grid operations, especially in regions with strong EV adoption. In order to optimize transportation efficiency while also considering grid operational restrictions, future studies can expand on the Coati–NGO method by adding multi-objective optimization functions.

5. Conclusions and Future Work

The GNN-ViGNet model and hybrid Coati–NGO approach examined EV charging load forecasting and route optimization issues. Route efficiency and EV charging load estimates improved significantly in this trial. The suggested technique combined real-time charging patterns, traffic, and vehicle positions from IoT sensors into the GNN-ViGNet model to improve forecasting accuracy and dynamic routing choices. The GNN-ViGNet model outperformed traditional approaches with 98.9% accuracy and low prediction errors (MSE, RMSE, and MAE). Coati–NGO had the shortest route (511 km) and fastest convergence compared to PSO and Firefly. Based on IoT data, including traffic, battery levels, and charging station availability, Coati–NGO optimized routes to save energy and boost fleet performance. These findings may improve electric vehicle fleet efficiency, energy use, and operating expenses.

Future work will include power-grid operating constraints, including peak load management, balanced energy distribution, and grid stability in the optimization framework. This integrated method would optimize transportation and power networks, addressing EV fleet operations and grid stability. Advanced models might account for energy price, renewable energy contributions, and system resilience for a more holistic and efficient solution. One possibility is to validate the generalizability and scalability of the GNN-ViGNet and Coati–NGO models in multiple metropolitan areas with different traffic and environmental circumstances to verify the generalizability and scalability of GNN-ViGNet. The approach may go from localized performance to universal applicability by using varied datasets and extending assessments to many cities, allowing smart-city ecosystem adoption.