The escalating threat of global warming poses a formidable challenge to sustainability, necessitating a transformative shift in the transportation sector. A pivotal solution lies in transitioning from conventional fuel-based vehicles to electric vehicles (EVs) to curtail global warming and unlock significant social and economic benefits. However, this transition is far from straightforward and consists of many challenges, with a major concern being the accurate estimation of the EV population on our roads. Many parameters influence EV adoption, making it crucial to gauge the potential number of EVs on the road. To address this, our study delves into the depths of machine learning (ML), conducting a study to estimate the EV penetration of the Uttarakhand region in India by employing different ML algorithms, including random forest (RF), support vector machine (SVM), decision trees, artificial neural networks (ANN), and K-nearest neighbor (KNN). After the estimation of EV penetration, an approach to determine the energy and power requirements in the grid infrastructure is shown, considering the domestic EV charging scenario. The study shows that the SVM and ANN algorithms can be used for the estimation of EV penetration, achieving a higher R-square score of 0.979 and 0.978 respectively, with less root mean square error (RMSE).
Research Article
Performance Analysis of Machine Learning Algorithms for Estimation of EV Penetration
https://doi.org/10.21203/rs.3.rs-4153186/v1
This work is licensed under a CC BY 4.0 License
Journal Publication
published 29 Oct, 2024
You are reading this latest preprint version
EV Penetration
Machine Learning
Random Forest
Support Vector Machines
Artificial Neural Networks
and Domestic EV Charging
As the population accelerates, the burden on the transportation sector increases to provide more reliable and efficient operations. Numerous commodities and services heavily rely on the transportation sector to meet the growing supply chain and energy generation. Focusing on the energy generation portion, the demand rises as the population density increases, putting pressure on the power generating stations. Therefore, environmentally friendly options are being taken into consideration. Renewable energy resources such as Photovoltaic (PV), Wind, Biomass, and Hydro are being adopted as alternatives for power generation to meet the growing electricity demand. As the power grid undergoes this transition towards environmental alternatives, the time has also become ripe for a revolution in the transportation sector. Transformation from fossil-fueled vehicles to Electric Vehicles (EVs). In the transportation sector, fossil fuel vehicles are the most used for meeting the requirements; fossil fuel vehicles not only increase greenhouse gas (GHG) emissions but also possess high energy inefficiency. On the other hand, EVs are much more environmentally friendly and efficient compared to fossil-fueled vehicles.
The transportation industries have realized the need for this transition and have initiated several steps into this revolutionary transformation. EVs are now replacing fossil fuel-based vehicles in various transportation subsectors like essential delivery areas, public transportation areas, and mobility used in industries for shipping. This integration of EVs in different areas of the transportation sector will provide environmental benefits and enable economic and operational benefits associated with this transformation.
As per the World Health Organization, India, a developing country, has the world's worst air quality. Besides this, India is the world's third-largest fossil fuel and GHG-emitting consumer, mostly from India's transportation sector [1]. To reduce the concerns above, the Indian Government has launched several policies accelerating EV adoption in India. Several policies and campaigns are under action, like the National Electric Mobility Mission Plan (NEMMP) 2020, EV 30@30, and Faster Adoption and Manufacturing of Electric Vehicles (FAME) in India. India is targeting to convert fossil fuel-based vehicle sales to 100% Plug-in EVs (PEVs) by 2030 [2]. As the influx of EVs increases, the expected surge in power demand challenges existing grid infrastructure. Meeting this demand increases power generation and the investment involved in the grid infrastructure, smart grid technologies, and energy storage to provide a more reliable and resilient power supply. The adoption of EVs mostly depends on the accessibility of the charging infrastructure. Therefore, for the adoption of EVs, the government needs to develop a well-distributed network for charging stations. By integrating advanced modelling techniques, the grid operations can estimate the concentration of the EVs in different regions and predict the likely load on the power grid. Load prediction will help the grid operators plan for the upgrades required in the existing grid infrastructure to accommodate the surging demand.
In this paper, a model is developed to predict the EV density for Uttarakhand state in India using different machine learning algorithms like Random Forest (RF), Support Vector Machine (SVM), Decision Trees (DT), Artificial Neural Network (ANN), and K-nearest neighbor (KNN). The model depends upon several parameters like income, population density, altitude, and grid availability of that subregion in Uttarakhand, India. Based on the total electric vehicles registered in the different sub-regions, the weights of these parameters are calculated using multiple linear regression and are employed for the EV modelling. The model provides the total population of EV users in different sub-regions of Uttarakhand, India. Finally, based on the EV population, the EV load curve is determined considering the domestic charging scenario. The determined EV load curve will eventually be helpful for the grid operators in planning upgrade requirements in the grid infrastructure to meet the EV load demand.
The paper has VI sections, starting with an introduction to EVs and why measuring their penetration is crucial. Section II discusses ongoing research conducted in consideration of EVs and other parameters correlated to them. Section III explores ML techniques utilized for the modelling of EVs and their other aspects. Section IV shows the overall methodology, showcasing the process of data collection, data processing, weightage calculation of different parameters and penetration estimation model. Section V shows the results of the study, describes the efficacy of the ML models used for the study, total EV penetration, and, based on the estimation of the total energy and power requirement. Section VI is the conclusion section highlighting the outcomes of the study conducted and any future work that may be possible in this study.
The global shift towards sustainable transportation calls for attention towards the influence of EVs on the energy and power sector industry. Several researchers are focusing on the various aspects of EVs, examining their charging & discharging behavior, their impact on the grid, and possibility of exploiting them as the emergency backup services. In [3] impact of Electric Vehicle (EV) penetration on the power system is studied using a multi agent simulation, employed with an EV charging and other control algorithms. The study signifies that uncontrolled EV charging can threaten the power system stability. Due to this uncontrolled charging behavior, the overall peak demand increases by around 50%. The study observes that scheduled charging compared to uncontrolled charging provides less peak demand increase and reduces load variability in the power system. A hybrid transfers learning model [4] using Convolution Neural Network-Bidirectional Long Short-Term Memory (CNN-BiLSTM) is proposed for forecasting EV charging profile using insufficient historical data of EV charging. The models include both commercial and domestic charging data. The model exhibits superiority compared to the other traditional models considering case with limited data availability. However, the model’s limitation is its scalability for larger and complex power systems. In [5] prediction of increase in energy consumption due to EV penetration is conducted on Jeju Island, Korea. The prediction is performed by employing statistical analysis and fuel economy data. The study is performed on the context of Jeju’s aim to become carbon free island by 2030. The authors have predicted growth of EV usage and corresponding to it, increment in the energy demand up to 2022. Another study is performed to predict penetration of EVs in Costa Rica’s power system [6]. The model is developed using Bass and Gompertz mathematical approaches, the study is conducted to show the necessity to assess impact of the EV penetration and its integration on the power grid. The research concludes that the prediction of EV penetration is challenging especially in limited historical data scenarios. The study [7] evaluates the impact of EV charging on the transformer loading present at distribution grid. Modelling of the driving habits & charging demand is estimated using survey data. The model performs the effect of EV charging using two scenarios, immediate charging performed at the arrival of the EV at home, and the other one is delayed charging which is initiated 4 hours later. The findings suggest that delayed charging results in a more balanced load on the transformer side, indicating it as a mitigation technique against the overloading due to EV charging. An E-mobility road map scenario analysis [8] is performed to examine the impact of EV penetration on Singapore’s distribution grid. Transformer loading is analyzed under different charging scenarios including single phase, three phase, and DC fast charging. The study demonstrates that single phase and three phase charging might not substantially influence the distribution grid till 2050. A micro simulation-based strategy [9] is employed which models the individual behavior of drivers and their charging choices. The model exhibits spatial and temporal context, covering technical specifications of EVs. The proposed strategical approach helps in assessing the impact of EV penetration in different areas of a city, having varied consumption behaviors at charging points in terms of power & energy demand. The proposed model utilizes Monte Carlo simulation considering various factors including distribution system topology, penetration level in the area, available charging power, vehicle battery capacity, State of Charge (SoC) of battery, EV users’ behavior, and daily energy requirements of EV users. The authors have proposed a methodology flowchart to assess the impact of on power demand due to EV penetration. The study [10] involves assessment of EV penetration at two different levels (20% & 80%) on the power grid, especially focusing on the impact on the feeders located at rural and urban areas. The study shows that urban areas with 80% penetration levels jeopardize power grid with highest percentage increase of feeder loading.
In [11], the study delves into evaluating the influence of Plug in Electric Vehicle (PEV) charging on the distribution grid of New South Wales, Australia. The study considers different penetration levels of PEV. The study is evaluated using three main tools, tool A focusses on the modelling of the energy demand required and charging availability of PEVs. Tool B develops the charging load profile considering unmanaged charging. Tool C is used for the cost estimation of the upgradation required for meeting estimated energy demands. The research shows that even small penetration levels of PEVs can increase the distribution assets ratings, increasing the overall cost of the equipment. In [12], the author conducted a case study on EVs, emphasizing the impact of charging on Turkey's distribution network. It was noted that Turkey's electrical grid faces challenges in accommodating the electricity needs of EVs due to the absence of communicative charging stations and unclear load capacities. The paper [13] explores the challenges encountered by distribution system operators (DSOs) as EVs become more prevalent. The paper introduces an innovative approach for estimating charging concurrency, incorporating factors like popularity time, waiting time, and visiting time. Utilizing the concurrency factor, a load profile is generated, aiding DSOs and researchers in strategically planning the seamless integration of EVs into the grid. In [14], authors introduce an innovative approach to signal a charge warning and offer path planning strategies to mitigate electricity shortages while driving. The model assesses real-time electricity consumption within the vehicle, issuing timely warnings when energy levels are insufficient. Additionally, the model suggests an optimal driving route, considering variables like queuing time at charging stations. The Dijkstra algorithm is employed to determine the most efficient path, demonstrating its effectiveness in reducing driver travel time. In the context of paper [15], the impact of charging EVs and Plug-in Hybrids (PHEVs) on a low-voltage power system is discussed. The study utilizes unsymmetrical power flow calculations to assess voltage imbalances caused by single-phase charging vehicles in a weak grid scenario, considering load profiles, photovoltaics (PV), and distributed EVs. The voltage imbalance exceeds allowable limits, reaching approximately 2% for 10 minutes and 4% at a single time during EVs and PHEVs charging. A grouping strategy proposed in [16] enables EVs to contribute to peak shaving in day-ahead plan generation, considering both grid and consumer needs. As the number of EVs increases, the strategy proves effective in mitigating the impact of dimensionality disaster. Moreover, a hybrid optimization algorithm presented in [17] manages energy storage in PV-integrated EV charging stations. This algorithm addresses uncoordinated charging behavior and variable solar output, incorporating considerations like band allocation and cost degradation models. The three-part algorithm includes real-time electricity price categorization, calculation of solar output from solar radiation, and optimization to achieve minimal operating costs. From literature survey it can be concluded that many researchers have developed models to predict the EV penetration, to analyze the impact of EVs charging on the distribution grid. Furthermore, fruitful research has been conducted to control the unmanaged charging behavior, ensuring that grid reliability and performance remains untouched. However, it observed, prediction of EV penetration is performed mostly based on insufficient historical data, and on the charging behavior of the user. Whereas the socio-economic parameters like income, population density, existing grid availability, altitude and various other parameters also influence EV penetration in a particular regional study.
Understanding the importance of socio-economic factors including income, population density, grid availability, and altitude becomes necessary especially for effective prediction of EV penetration. Income level directly influences the affordability of EVs, higher income normally results in higher willingness to invest in sustainable transportation options. Population density influences the demand for EV charging infrastructure which directly aligns with grid availability in that region. Consequently, population density and grid availability are two factors correlated to each other. On the other hand, altitude affects the performance of EV. Regions with higher altitude eventually result in degrading of the battery efficiency due to additional burden on the EV motor.
Machine Learning (ML) is a cutting-edge field in computer innovation that falls under the category of artificial intelligence [18]. The competencies of these algorithms are associated with applications for managing the EV infrastructure and grid operations. By processing the real-time inputs and historical data, these algorithms can accurately predict the EV load curve, enabling the DSOs to forecast demand and optimize the charging infrastructure placement and capacity, ensuring grid stability. In [19], a human-machine reinforcement learning framework is proposed for energy management of EVs. By employing deep deterministic policy gradient (DDPG) and deep Q learning (DQN) techniques, the structure optimizes the energy consumption by maintaining the voltage magnitude at local level. The proposed algorithm overall decreases the decision-making process and human intervention with ML, results in significant reduction in losses during learning, realize emergency control, and identify superior control policy. A deep learning framework is proposed in [20] for EV load management within the smart grid. The structure makes probabilistic short-term forecasting for EV charging demand; it employs a partial convex neural network (PCNN) to predict the distribution of day-ahead charging demand for each EV charging station and mollify issues like quantile crossing. The framework also employs LSTM and convex learning layers to capture the probabilistic distribution of stochastic scenarios. A novel application of non-intrusive load monitoring (NILM) [21] to identify distributed energy resources (DERs), specifically EVs and rooftop solar panels, in a low-voltage (LV) distribution grid. The proposed NILM method utilizes three machine learning approaches: KNN, RF, and multilayer perceptron. The approach is employed to analyze the aggregated measurements obtained from the LV distribution side transformers. The study involves the evaluation of these approaches in different scenarios of DER integration, evaluating the F1 score for EV and PV identification. The methodology overall showcases the feasibility of real-time identification of DERs from low-frequency electric measurements present at the LV distribution networks. In the study [22], enhancement in electric vehicle load curve (EVLC) forecasting using nine different methodologies, such as statistical, machine learning, and deep learning, is performed. Evaluation of these methods is done based on four public, real-world EV datasets. The study concludes that machine learning algorithms such as XGBoost and Multilayer Perceptron (MLP) outperform deep learning models in prediction EVLC. The study also shows the potential of simpler machine learning algorithms, particularly in scenarios with limited data availability.
The above studies show that ML provides a significant potential for the application of EV load management. The algorithms can be employed to predict the EV load curve, optimize charging strategies, and enhance EV fleets' participation in the energy market. The ML models used for the different applications can be broadly classified into supervised and unsupervised learning. In this paper, for the prediction of the EV population, supervised learning algorithms like RF [23], SVM [24], Decision tree [25], ANN, and K-nearest neighbor [26]. The general specifications of these supervised machined learning algorithms are shown in Table 1.
| Algorithm | Principle | Pros | Cons | Parameters Required for Modeling | Formula/Key Concept |
|---|---|---|---|---|---|
| RF [27] | Ensemble learning using decision trees | High accuracy and robustness | Complexity can lead to slow training and prediction | Number of trees, tree depth, minimum samples per leaf, maximum features considered for splitting | Decision trees with random feature and data subset selection |
| SVM [28] | Separates data with a hyperplane in high-dimensional space | Effective in high-dimensional spaces | Sensitivity to noise and overfitting | Kernel type, regularization parameter (C), kernel-specific parameters | Maximizing margin with constrained optimization |
| ANN [24] | Mimics the structure and function of the human brain | Ability to model complex relationships | Prone to overfitting, requires large amounts of data | Number of layers, number of neurons per layer, activation functions | Weighted sum of inputs with activation function |
| Decision Trees [29] | Hierarchical partitioning of data based on attributes | Interpretable, easy to understand | Prone to overfitting, sensitive to small variations in data | Criteria for attribute selection, tree depth, minimum samples per leaf | Gini impurity or entropy for attribute selection |
| KNN [26] | Classifies based on majority class or averages values of k-nearest neighbors | Simple, no training phase, works well with small datasets | Computationally expensive during prediction with large datasets | Number of neighbors (k), distance metric (e.g., Euclidean distance) | Distance metric (e.g., Euclidean distance) |
3.1 Random Forest
RF is an ensemble learning technique that builds many decision trees during training and predicts the class that appears most frequently (classification) or the individual trees' average prediction (regression). The introduction of randomization into the model, achieved by randomly selecting subsets of features and data points, mitigates overfitting and improves generalization. The final prediction is determined by calculating a weighted average or majority vote of the various trees, which makes the RF method highly resilient and adaptable.
3.2 Support Vector Machine
SVM is a type of supervised ML algorithm that categorizes data into distinct classes by identifying the hyperplane that optimally separates the data points of one class from those of another. The primary objective of SVM is to optimize the margin between different classes by decreasing the classification error. The algorithm accomplishes this by transforming the input data into a space with many dimensions and identifying the hyperplane that optimizes the distance between the closest data points from various classes.
3.3 Neural Networks
Neural Networks are a class of algorithms inspired by the structure and functioning of the human brain. They consist of interconnected layers of nodes (neurons) with each connection having an associated weight. Neural networks learn by adjusting these weights during training, allowing them to capture complex patterns and relationships in the data. Deep neural networks, or deep learning, involve multiple layers, enabling the model to learn hierarchical representations.
3.4 Decision Trees
Neural Networks are a class of algorithms inspired by the structure and functioning of the human brain. They consist of interconnected layers of nodes (neurons) with each connection having an associated weight. Neural networks learn by adjusting these weights during training, allowing them to capture complex patterns and relationships in the data. Deep neural networks, or deep learning, involve multiple layers, enabling the model to learn hierarchical representations.
3.5 K-Nearest Neighbor
The K-Nearest Neighbors algorithm is a straightforward and intuitive method for classification and regression tasks. The classification of a data point is determined by either identifying the majority class or calculating the average values of its k nearest neighbors in the feature space. The selection of 'k' defines the number of neighbors considered throughout the decision-making procedure.
| Districts | Population Density (Population/km2) | Grid Availability (Capacity in MVA) | Altitude (Meters) | Income (INR) | Total Electric Vehicle Registered |
|---|---|---|---|---|---|
| Dehradun | 637.3591321 | 1,045 | 640 | 235770 | 11372 |
| Haridwar | 929.190678 | 1,097 | 315 | 362688 | 12096 |
| Tehri Garhwal | 197.1334432 | 113 | 1750 | 103345 | 15 |
| Uttarkashi | 47.76696607 | 87 | 1,158 | 107281 | 3 |
| Pauri Garhwal | 152.4347992 | 215 | 1,789 | 108640 | 8 |
| Chamoli | 56.57061021 | 168 | 1,676 | 127330 | 0 |
| Rudraprayag | 141.6587702 | 36 | 895 | 93160 | 1 |
| Bageshwar | 134.2306322 | 39 | 1,006 | 98755 | 9 |
| Pithoragarh | 79.09576869 | 137 | 1,514 | 18678 | 13 |
| Almora | 230.0436445 | 145 | 1,651 | 100844 | 18 |
| Nainital | 260.4897671 | 293 | 1,938 | 190627 | 5267 |
| Champawat | 170.5503964 | 59 | 1,642 | 116136 | 874 |
| Udham Singh Nagar | 657.7462173 | 347 | 550 | 215689 | 8277 |
| Mean | 284.1746789 | 291 | 1271.077 | 144534.0769 | - |
| Standard Deviation | 276.0159061 | 358.5253256 | 540.3762 | 87111.7341 | - |
From the literature survey, it can be observed that estimation of penetration level of EVs is an area that has not been explored widely. To address this, the section shows the overall methodology of the study to estimate EV penetration, including data collection, weightages calculation, data preparation, mathematical model, EV penetration calculations. The steps involved are explained furthermore in the section and are visually depicted in Fig. 1, providing a clear overview of the approach adopted for this study.
4.1 Data Collection
In this study, EV penetration is estimated for Uttarakhand state in India. The state of Uttarakhand is found in northern India. The state is well-known for its majestic locations from the British era and its gorgeous scenery. There are thirteen districts in all, and each has a unique significance. The state is home to a diverse population of immigrants from across India and other communities. Uttarakhand is situated approximately 1250 meters above sea level on average. Various businesses, including tourism, agriculture, and manufacturing, are the main drivers of the state economy.
Different parameters are considered when developing the EV penetration model for Uttarakhand. This includes the population density of each district, the average annual income per person, the district's total grid availability, the district, the altitude of the district, and the total EVs registered in the district from the 2013–2022 cycle. The data used for the study is collected from several reliable resources, including government reports and official websites [30]–[33]. Here, the total EVs registered in the districts has been considered the target variable whereas other parameters are considered the features. Each parameter's weightage is calculated using Sequential Least Square Programming (SLSQP) algorithm. The overall process of weightage calculation is discussed further in the paper. The data collected for the calculation of the weightage is shown in Table 2.
4.2 Weightages calculation using SLSQP for Predictive Modelling
The section outlines the approach utilized for calculating the weightage using the SLSQP algorithm. SLSQP is an optimization algorithm used to solve non-linear optimization problems. This paper uses the SLSQP algorithm to calculate the weights of the features to calculate the EV penetration.
A linear regression model is employed to predict the EV adopted (y) as the function of several independent variables (xi). The linear regression model equation is shown in Eq. 1.
| \(\varvec{y}={\varvec{w}}_{1}{\varvec{x}}_{1}+{\varvec{w}}_{2}{\varvec{x}}_{2}+{\varvec{w}}_{3}{\varvec{x}}_{3}+{\varvec{w}}_{4}{\varvec{x}}_{4}+\varvec{b}\) | (1) |
Here, wi is the weight associated with each independent parameter, b is the biased term which is equal to 1. The independent parameters include population density, income, altitude, and grid availability denoted by x1, x2, x3 & x4.
Weight calculation using the SLSQP algorithms requires setting the objective function based on which the algorithm starts its optimization process. This study considers Mean Square Error (MSE) as the objective function. The calculation of MSE is shown in Eq. 2.
Where yi is the actual number of EV adoption while ypred represents the predicted number of EV adoption using the linear regression model, and n is the total number of samples present in the dataset. The objective function is to find the optimal weights of the features till the minimum value of objective function is reached. The optimal weights calculated through SLSQP is shown in Table 3.
4.3 Data Preparation
As discussed earlier, various parameters are considered when predicting EV penetration in a certain area. The coefficient of each parameter is determined, considering total EV registration as the target variable and other variables as features.
Now, for model development, a standardized and large data set is required. Each feature's mean and standard deviation are calculated and utilized to generate around 10,000 samples for each variable, as shown in Table 2. These datasets are standardized using the Min-Max technique. This normalization technique brings all the features to a common scale, ensuring equality during analysis. The outliers, i.e., the features with negative values, are detected and removed from the dataset, ensuring the integrity of the data generated for modelling and analysis.
| Weightage | Value | Weightage | Value |
|---|---|---|---|
| W1 | 1.67 | W3 | -0.0832 |
| W2 | 0.0074 | W4 | 8.097 |
4.4 EV Penetration Estimation
EV penetration depends upon factors such as charging behavior, energy prices, range anxiety, government policies and other socioeconomic parameters. These factors show the overall adoption rate of EVs in a particular region. This paper focuses on socioeconomic parameters like population density, grid availability, income, and altitude. The EV penetration is calculated using a mathematical model extracted from the socioeconomic parameters shown in Eq. 3.
Here the coefficients w1, w2, w3 & w4 are the weightages associated with respective socio-economic parameters. The calculation of weights is already discussed in the above section. ϵ is the noise added to the model to check the robustness of the algorithms used for the prediction of EV penetration.
The section deals with the results of the study performed to estimate the EV penetration and the EV load curve, considering the domestic charging scenario in the Uttarakhand region of India. The assessment of the model used for EV penetration prediction is done on several metrics like the coefficient of determination (R-square Value) & Root Mean Square Error (RMSE). The R-square score plot and score of different ML techniques are shown in Fig. 2 to Fig. 6 & Table 4. It is evident from the results that the SVM model achieves the highest R-square value of 0.9791, closely followed by the ANN model with a value of 0.9784. Among the models assessed, the SVM model achieved the lowest RMSE value of 0.1503, indicating the highest predictive accuracy among the other evaluated models. However, ANN is closely followed by SVM, having an RMSE of around 0.1662. The two models have shown superiority in predicting compared to other models like RF, decision trees, and KNN.
The decision tree algorithm displays a wider range of prediction errors, spanning between range 1 and − 0.75. Higher compared to SVM and ANN. The prediction error of the KNN algorithm is the highest, lying between 1 & -1.5. It can be concluded that for the prediction of EV penetration, prioritizing the RMSE as the main objective value, SVM and ANN present viable options for future consideration.
After training and testing of ML models for EV penetration estimation, it was found that the highest EV penetration at a particular region reached 9, with total penetration across the whole Uttarakhand region coming to approximately 48,929. Following the determination of EV penetration, the EV load curve is estimated, where charging duration emerges as the pivotal parameter for assessing the duration for which EVs are connected for charging. The charging duration in the context of the charging scenario within the Uttarakhand region, a type 1 charger with a 3.5KW rating, is considered for charging. During the simulation, it is assumed that the EV users consumed constant power at a rate of 3.5KW. The charging duration of each is considered from [34], the mean and standard deviation of charging duration are approximately 170 minutes and 90 minutes, respectively. The distribution of charging duration is shown in Fig. 8. It is seen that most of the EVs are charged for a duration between 100 to 300 minutes duration.
For simulation, February month, having 28 days, is considered for the study. Each day of February month is categorized into peak and non-peak charging hours. These are defined based on predefined timing ranges, i.e., 5 am to 9 am and 6 pm to 10 pm are considered as the peak hours, meaning that maximum EVs will be connected at this period for charging. The rest of the remaining period is considered as non-peak hours. As domestic charging is considered for this study, the peak hour period is measured based on the duration for which the EV user is at home. Taking into account that during the peak hours, the maximum number of EVs connected for charging is 50% of the total EV penetration of the region, as estimated. During non-peak hours, the total number of EVs connected for charging ranges between 10 and 30%. The charging duration of each EV is allotted using the distribution curve of charging time shown above in Fig. 8. The total number of EVs connected for charging during the February month is shown in Fig. 9. The highest no. of EVs connected for charging is around 24,442 on 12th February at 18:30 pm.
| Model | R-square | RMSE |
|---|---|---|
| RF | 0.97438 | 0.166 |
| SVM | 0.97906 | 0.150 |
| ANN | 0.9784 | 0.15 |
| Decision Tree | 0.94799 | 0.236 |
| KNN | 0.95804 | 0.212 |
The simulation sample time is 15 minutes, and energy and power consumption are calculated at each sample time slot. This is achieved by multiplying the average power consumption of EVs by their charging time duration. The overall energy consumption in a day for February month is shown in Fig. 10. It can be observed that during the 12th day of February month, when the highest no. of EVs is connected for charging, the energy requirement on that day is the highest compared to all the other days in the February month. A cumulative energy consumption bar graph is presented in Fig. 10, showing a holistic perspective on energy consumption.
The power load curve shown in Fig. 11 reflects the domestic charging of EVs for February, taking into account the above-mentioned statements. The average power load curve for February month is shown in Fig. 12, which is derived by calculating the energy consumption within each hour; the average energy consumption is calculated by dividing the total energy consumption in each hour by the total number of days in the month. Finally, the average energy consumption is converted into the average power consumption by dividing it by the number of hours in each time slot. From Fig. 12, it can be observed that the power demand increases in the domestic charging scenario especially during the time duration when the users are at home.
Lastly, the model used for the prediction of EV penetration can be utilized by the DSO to study the upgradation required in their existing infrastructure by estimating the energy and power demand that arises due to the rising penetration of EVs.
The study delves into the estimation of EV penetration and the characterization of EV load curves in domestic charging scenarios within the Uttarakhand region of India. Different machine learning algorithms, notably Support Vector Machine (SVM) and Artificial Neural Network (ANN), have shown high efficacy in accurate estimation for EV penetration estimation. The superior performance of the SVM and ANN models is evidenced by high R-square values and low root mean square error (RMSE) scores. Moreover, analysis of the EV load curve provides valuable insights into the dynamics of energy consumption patterns, revealing peaks of the charging period and distribution of charging duration among EV users. By introducing peak and non-peak hours and assessing the energy demand fluctuations throughout February, the study could further optimize the charging infrastructure and enhance the grid resiliency against the sudden increase in load demand due to EV charging. The EV load curve is estimated considering the domestic charging scenario, where the overall duration of the day is classified into peak and non-peak hours. The charging time duration is the key parameter to determine the energy and power requirement. In the future, more refining could be done in predictive modelling, introducing dynamic peak hours considering the energy market scenario and the different EV charging associated parameters, including state of charge and battery size. To study the impact of EV integration, the EV load curve can be integrated into different test systems to understand the dynamics of the power distribution network.
Competing Interest Declaration:
All authors declares that here is no Competing Interest for this manuscript
Data Availability Declaration:
All authors declare that the relevant data is available in the manuscript itself. There is no any other data.
Author Contributions StatementAbhay Chhetri: Conceptualization, Methodology, Investigation, Writing- Original Manuscript preparationDevender Kumar Saini: Conceptualization, Analysis, Investigation, SupervisionMonika Yadav: Conceptualization, Reviewing, SupervisionNitai Pal: Conceptualization, Reviewing, Supervision
- R. Dua, S. Hardman, Y. Bhatt, and D. Suneja, “Enablers and disablers to plug-in electric vehicle adoption in India: Insights from a survey of experts,” Energy Reports, vol. 7, pp. 3171–3188, 2021, doi: 10.1016/j.egyr.2021.05.025.
- P. Plötz, J. Axsen, S. A. Funke, and T. Gnann, “Designing car bans for sustainable transportation,” Nat. Sustain., vol. 2, no. 7, pp. 534–536, 2019, doi: 10.1038/s41893-019-0328-9.
- Z. Wang and R. Paranjape, “An evaluation of electric vehicle penetration under demand response in a multi-agent based simulation,” Proc. - 2014 Electr. Power Energy Conf. EPEC 2014, pp. 220–225, 2014, doi: 10.1109/EPEC.2014.14.
- P. Banda, M. A. Bhuiyan, K. N. Hasan, and K. Zhang, “Assessment of hybrid transfer learning method for forecasting EV profile and system voltage using limited EV charging data,” Sustain. Energy, Grids Networks, vol. 36, no. August, p. 101191, 2023, doi: 10.1016/j.segan.2023.101191.
- S. W. Kim and G. M. Lee, “Estimating increase of electric energy according to penetration of electric vehicles at the Jeju Island in Korea,” 2016 IEEE Transp. Electrif. Conf. Expo, Asia-Pacific, ITEC Asia-Pacific 2016, pp. 947–949, 2016, doi: 10.1109/ITEC-AP.2016.7513086.
- G. A. Gomez-Ramirez and R. Solis-Ortega, “Electric Vehicle Penetration Modelling for Costa Rica Power System,” 2021 IEEE Chil. Conf. Electr. Electron. Eng. Inf. Commun. Technol. CHILECON 2021, 2021, doi: 10.1109/CHILECON54041.2021.9703070.
- M. Coban and S. S. Tezcan, “Analysis of Impact of Electric Vehicles on Distribution Grid Using Survey Data,” 3rd Int. Symp. Multidiscip. Stud. Innov. Technol. ISMSIT 2019 - Proc., pp. 1–4, 2019, doi: 10.1109/ISMSIT.2019.8932933.
- S. Vaisambhayana and A. Tripathi, “Study of electric vehicles penetration in Singapore and its potential impact on distribution grid,” 2016 Asian Conf. Energy, Power Transp. Electrif. ACEPT 2016, pp. 1–5, 2017, doi: 10.1109/ACEPT.2016.7811513.
- A. F. Botero and M. A. Rios, “Demand forecasting associated with electric vehicle penetration on distribution systems,” 2015 IEEE Eindhoven PowerTech, PowerTech 2015, pp. 1–6, 2015, doi: 10.1109/PTC.2015.7232236.
- E. Mancini, M. Longo, F. Foiadelli, G. Parrotta, and G. Montinaro, “Different penetration of electric vehicles and impact on developments in the electric grid,” 2020 IEEE Veh. Power Propuls. Conf. VPPC 2020 - Proc., 2020, doi: 10.1109/VPPC49601.2020.9330914.
- M. Hemphill, “Electricity distribution system planning for an increasing penetration of plug-in electric vehicles in New South Wales,” 2012 22nd Australas. Univ. Power Eng. Conf. "Green Smart Grid Syst. AUPEC 2012, pp. 1–6, 2012.
- M. Burunkaya and O. F. Demirkol, “Increase in the use of electric vehicles and its potential effects on electricity distribution network and situation analysis for Turkey,” Proc. - 2019 6th Int. Conf. Electr. Electron. Eng. ICEEE 2019, pp. 33–37, 2019, doi: 10.1109/ICEEE2019.2019.00014.
- M. Draz and S. Albayrak, “A Power Demand Estimator for Electric Vehicle Charging Infrastructure,” 2019 IEEE Milan PowerTech, PowerTech 2019, vol. 2019-Janua, pp. 1–6, 2019, doi: 10.1109/PTC.2019.8810659.
- D. Ding, J. Li, P. Tu, H. Wang, T. Cao, and F. Zhang, “Electric vehicle charging warning and path planning method based on spark,” IEEE Access, vol. 8, pp. 8543–8553, 2020, doi: 10.1109/ACCESS.2020.2964307.
- S. Helm, M. Tayyab, I. Hauer, M. Wolter, S. Balischewski, and P. Komarnicki, “Potential analysis of EV and PHEV in weak low-voltage grids,” NEIS 2021 - Conf. Sustain. Energy Supply Energy Storage Syst., no. 3, pp. 68–73, 2021.
- T. Gao et al., “A Grouping Strategy and Day-ahead Scheduling Method of Electric Vehicles for Peak Shaving,” Proc. 2022 IEEE 5th Int. Electr. Energy Conf. CIEEC 2022, pp. 2676–2681, 2022, doi: 10.1109/CIEEC54735.2022.9846201.
- K. Chaudhari, A. Ukil, K. N. Kumar, U. Manandhar, and S. K. Kollimalla, “Hybrid Optimization for Economic Deployment of ESS in PV-Integrated EV Charging Stations,” IEEE Trans. Ind. Informatics, vol. 14, no. 1, pp. 106–116, 2018, doi: 10.1109/TII.2017.2713481.
- T. Ayoub, T. Rasool, and F. Rasheed, “Towards leveraging the role of machine learning and artificial intelligence in precision agriculture and smart farming,” Comput. Electron. Agric., vol. 198, no. June, p. 107119, 2022, doi: 10.1016/j.compag.2022.107119.
- Y. Tao, J. Qiu, S. Lai, X. Zhang, Y. Wang, and G. Wang, “A Human-Machine Reinforcement Learning Method for Cooperative Energy Management,” IEEE Trans. Ind. Informatics, vol. 18, no. 5, pp. 2974–2985, 2022, doi: 10.1109/TII.2021.3105115.
- K. Zheng, H. Xu, Z. Long, Y. Wang, and Q. Chen, “Coherent Hierarchical Probabilistic Forecasting of Electric Vehicle Charging Demand,” IEEE Trans. Ind. Appl., pp. 1–12, 2023, doi: 10.1109/TIA.2023.3344544.
- A. F. M. Jaramillo et al., “Distributed Energy Resources Electric Profile Identification in Low Voltage Networks Using Supervised Machine Learning Techniques,” IEEE Access, vol. 11, no. February, pp. 19469–19486, 2023, doi: 10.1109/ACCESS.2023.3247977.
- Z. N. Bampos, V. M. Laitsos, K. D. Afentoulis, S. I. Vagropoulos, and P. N. Biskas, “Electric vehicles load forecasting for day-ahead market participation using machine and deep learning methods,” Appl. Energy, vol. 360, no. February, p. 122801, 2024, doi: 10.1016/j.apenergy.2024.122801.
- K. Fawagreh, M. M. Gaber, and E. Elyan, “Systems Science & Control Engineering : An Open Access Random forests : from early developments to recent advancements,” vol. 2583, 2014, doi: 10.1080/21642583.2014.956265.
- Y. Wang, D. L. Wu, C. X. Guo, Q. H. Wu, W. Z. Qian, and J. Yang, “Short-Term Wind Speed Prediction Using Support Vector Regression,” pp. 13–18, 2010, doi: 10.1109/PES.2010.5589418.
- B. V. S. Vardhan, M. Khedkar, I. Srivastava, P. Thakre, and N. D. Bokde, “A Comparative Analysis of Hyperparameter Tuned Stochastic Short Term Load Forecasting for Power System Operator,” pp. 1–21, 2023.
- S. Aziz, “Electricity Theft Detection using Empirical Mode Decomposition and K-Nearest Neighbors,” 2020.
- V. Y. Kulkarni, P. K. Sinha, and M. C. Petare, “Weighted Hybrid Decision Tree Model for Random Forest Classifier,” 2014, doi: 10.1007/s40031-014-0176-y.
- S. Tavara, “Parallel Computing of Support Vector Machines : A Survey,” vol. 51, no. 6, 2019.
- C. E. P. Vinícius G. Costa, “Recent advances in decision trees: an updated survey.,” Artif. Intell. Rev., vol. 56, pp. 4765–4800, 2023, doi: https://doi.org/10.1007/s10462-022-10275-5.
- “Population,” Dist. Uttarakhand, Popul. Census, [Online]. Available: https://www.census2011.co.in/census/state/districtlist/uttarakhand.html
- “Substation,” Uttarakhand Power Corporation Limited. https://www.upcl.org/substations/
- “Topographical,” Dehradun Topological Mao, Topographic-map.
- “VAHAN, National Register e-Services,” Ministry of Road Transport & Highways, Government of India. https://vahan.parivahan.gov.in/vahan4dashboard/
- “Electric Chargepoint Analysis 2017,” Department for Transport, GOV.UK. https://www.data.gov.uk/dataset/5438d88d-695b-4381-a5f2-6ea03bf3dcf0/electric-chargepoint-analysis-2017-domestics
No competing interests reported.