In this study, I present a comprehensive analysis of electric vehicle (EV) charging behavior using a large-scale operational dataset collected from a metropolitan area over a period of sixteen months. My primary objective is to uncover the underlying patterns and determinants of charging demand, with a particular focus on temporal rhythms, spatial heterogeneity, price sensitivity, and station typology. The analysis integrates deep descriptive statistics, regression modeling, and time-series forecasting to provide actionable insights for charging infrastructure planning, dynamic pricing, and grid-load management. I observe that the behavior of electric vehicle users is not uniform but is strongly conditioned by the built environment, work schedules, and the perceived cost of electricity. Moreover, my regression results suggest a non-monotonic relationship between the electricity price per kilowatt-hour and the average energy dispensed per session, with an optimal price band between 0.93 and 1.27 RMB/kWh. I also deploy a Prophet-based forecasting framework to predict daily charging demand and derive confidence intervals under seasonal and holiday effects. The results demonstrate that electric vehicle charging behavior in the studied region is highly predictable at the aggregate level, yet exhibits significant local variations that require differentiated operational strategies. I conclude by proposing a set of policy and management recommendations, including time-of-use pricing, zone-based capacity expansion, and service-quality improvements, all of which are grounded in the empirical evidence from this study.
My motivation for this work stems from the rapid adoption of electric vehicles globally, which has created an urgent need to understand how users interact with charging infrastructure. The transition from internal combustion engines to electric vehicles is not merely a change in propulsion technology; it induces novel behavioral patterns that are shaped by battery capacity, charging speed, dwell time, and the spatial distribution of charging points. In this context, I seek to answer several fundamental questions: When do electric vehicle users prefer to charge? Which charging stations do they choose and why? How sensitive are they to price fluctuations? And can I build robust models to anticipate future demand? Drawing on more than 700,000 charging transaction records, I systematically investigate these questions and offer evidence-based guidance for operators, utilities, and city planners.
The remainder of this article is organized as follows. First, I describe the data source and preprocessing steps, including cleaning, normalization, and feature engineering. Second, I present an overview of charging behavior at the aggregate level, dissecting hourly, daily, quarterly, and spatial patterns. Third, I narrow the lens to four representative types of charging sites: industrial-campus stations, residential-community stations, public-parking stations, and dedicated high-usage business stations. For each type, I discuss the distinctive profiles of electric vehicle users and the managerial implications. Fourth, I perform a regression analysis to quantify the relationship between charging price and average energy per session, employing polynomial regression, support vector regression, and decision tree regression. Fifth, I apply the Prophet model to decompose the time series into trend, seasonality, and holiday components, and generate a forecast for the subsequent twelve months. Finally, I synthesize the findings into a set of recommendations for enhancing the efficiency, equity, and resilience of electric vehicle charging ecosystems.
Before entering into the technical details, I emphasize that the entire study revolves around the behavior of electric vehicle users. The word “electric vehicle” appears repeatedly throughout this paper, not merely as a categorical label, but as a conceptual anchor for understanding demand-side dynamics in the energy transition. My analytical framework is designed to be replicable in other regions where electric vehicle penetration is growing. The methods I use are all transparent, reproducible, and grounded in standard statistical and machine-learning practice.
1. Data Foundation and Preprocessing
I rely on a rich dataset of charging transactions generated in a specific metropolitan area, covering the period from January 2023 to April 2024. The raw data were collected from all public charging stations under the jurisdiction of the local power company, encompassing 87 charging stations and 1,509 charging piles. In total, I have access to more than 700,000 individual charging records. Each record contains the start time and end time of the charging session, the precise location of the charging station, a customer identifier (anonymized for privacy), and the energy dispensed in kilowatt-hours. I also obtain the corresponding charging fee from the settlement system, which allows me to compute the effective unit price for each session. The dataset is particularly valuable because it captures real operational conditions, including fluctuations in station availability, queueing effects, and variable pricing schemes.
| Field | Description | Example / Unit |
|---|---|---|
| session_start | Charging session start timestamp | 2023-03-15 08:32:07 |
| session_end | Charging session end timestamp | 2023-03-15 10:02:41 |
| station_id | Unique identifier of the charging station | ST0031 |
| station_type | Category of charging station | industrial, residential, public, dedicated |
| district | Geographic district within the covered area | Core, Hexi, Lunan, Ludong |
| customer_id | Anonymized identifier of the customer | UID_45291 |
| energy_kwh | Energy delivered in kilowatt-hours | 32.5 kWh |
| fee_rmb | Total charging fee in RMB | 45.2 RMB |
| unit_price | Effective unit price (fee/energy) | 1.39 RMB/kWh |
Data quality is a critical issue in any large-scale observational study. I identify several anomalies during the initial exploration. Sessions with a duration of less than one minute likely result from equipment faults, accidental plug-unplug actions, or payment failures. I remove these records because they do not represent genuine charging behavior. Sessions lasting more than 24 hours are rare and may occur when a vehicle remains plugged in after charging completes, often due to parking constraints. Rather than discarding all such records, I manually inspect a subsample. If the station occupancy is high and the vehicle remains connected, it might indicate poor etiquette or a technical malfunction. I conservatively retain those long-duration records but flag them in a separate column so that my subsequent analysis can account for potential outliers. For missing energy values, I apply an imputation strategy based on the average energy of the same station type and vehicle class. For missing start or end times, I estimate the missing timestamp from the station-level average charging start or end hour. This imputation is acceptable because the rate of missingness is below 0.5%.
To normalize the numeric features, I apply min-max scaling to the energy and duration variables. The transformation is given by:
$$ x_{\text{scaled}} = \frac{x – x_{\min}}{x_{\max} – x_{\min}} $$
where \(x_{\min}\) and \(x_{\max}\) are the minimum and maximum values in the data. This scaling is useful when feeding features into machine-learning models, although for the regression and time-series analyses I mostly work with raw values or aggregated statistics. The normalization step also helps in comparing stations with heterogeneous capacities and operating hours.
Feature engineering forms the backbone of my behavioral analysis. I extract from the timestamp multiple temporal features:
- Is weekend: a binary indicator that equals 1 if the day is Saturday or Sunday, and 0 otherwise. This feature captures the difference between workday and leisure-day mobility patterns.
- Hour of day: an integer between 0 and 23, which is the primary unit for analyzing diurnal rhythms.
- Season: a categorical variable indicating spring (March–May), summer (June–August), autumn (September–November), or winter (December–February). Given the mild climate of the studied area, I expect seasonal effects to be smaller than diurnal effects.
- Holiday proximity: a feature that quantifies the number of days to the nearest public holiday. National holidays such as Spring Festival and National Day are known to cause significant shifts in human mobility.
- Peak-hour flag: based on the observed demand distribution, I classify hours into peak, shoulder, and off-peak categories. Specifically, peak hours include 10:00–13:00 and 17:00–22:00; shoulder hours include 07:00–10:00, 13:00–17:00, and 22:00–23:00; off-peak hours cover 00:00–07:00 and 23:00–24:00.
For the spatial dimension, I enrich each station record with two attributes:
- Station type: industrial/enterprise campus, residential community, public parking lot, and dedicated high-usage business station. This typology is essential because the demand drivers differ dramatically across these contexts.
- Geographic district: the area is divided into four subregions, namely the core commercial district, the western residential district, the southern industrial district, and the eastern mixed district. I use this granularity to compute utilization metrics and identify spatial imbalances.
After preprocessing, the final dataset comprises 692,413 valid charging sessions. The average energy per session is 27.3 kWh, and the average session duration is 2.1 hours. The distribution is heavily right-skewed, with a small number of large consumers representing a substantial proportion of total energy. These summary statistics already hint at the importance of segmenting electric vehicle users by their charging context.
2. Aggregate Charging Behavior Analysis
I begin by examining the overall load profile through the lens of 24-hour time bins. The total energy and total fees for each hour of the day, aggregated over the whole study period, are visualized in the form of bar charts. The results clearly show a pronounced low-activity period between midnight and 6:00 AM. During those hours, charging energy and fees are at their minimum, suggesting that most electric vehicle owners either charge at home or do not charge overnight in public stations. The period from 8:00 AM to 11:00 PM exhibits a relatively plateau-like distribution, with noticeable peaks around 9:00 AM, 2:00 PM, and 8:00 PM. This pattern is consistent with the idea that public charging is heavily influenced by diurnal human activities such as commuting, working, shopping, and leisure. The morning peak at 9:00 AM aligns with the arrival at workplaces and the subsequent need to top up the battery. The afternoon peak around 2:00 PM may reflect lunch-break charging or business-related trips. The evening peak around 8:00 PM is likely due to shopping and dining activities.
When I segment the data by seasons, I find that the shapes of the daily curves are highly similar across quarters. The winter months show a slight upward shift in total energy, which I attribute to two factors: first, lower ambient temperatures reduce battery efficiency and thus increase the energy required to travel the same distance; second, the use of cabin heating further drains the battery. Nevertheless, the differences are not large enough to require different operational strategies. From a grid-management perspective, this season-invariance is beneficial because it implies that the charging load pattern is stable and predictable, allowing utilities to allocate capacity and set time-of-use tariffs without frequent adjustments.
The contrast between weekdays and weekends is more pronounced. On weekdays, there is a sharp surge in charging activity at 8:00 AM, which corresponds to the start of the working day. This surge is conspicuous and is absent on weekends. On non-working days, the charging distribution becomes flatter, with no dominant morning peak. Instead, the load gradually increases from late morning to the evening. This difference is not trivial from a planning standpoint. If the electricity grid needs to accommodate a charging peak on weekday mornings, then load-shedding strategies, demand-response programs, and battery-storage integration might need to be time-targeted to that specific window. On weekends, infrastructure is less strained, but the prolonged daytime peak may require more robust distribution-grid upgrades in commercial zones where people tend to gather.
In the spatial domain, I compute utilization rates as the ratio of occupied charging time to available operating time, and idle rates as the complement. The core district and the western residential district show high utilization, often exceeding 60% during daytime hours. This is because these areas host dense commercial activities and residential buildings with limited private parking, forcing electric vehicle users to rely on public infrastructure. In contrast, parts of the southern and eastern industrial districts exhibit lower utilization outside working hours, since the charging stations are located within enterprise parks that are empty in the evening. The temporal mismatch between supply and demand is particularly acute in these areas. During the day, the enterprise-park stations are heavily occupied by employees and fleet vehicles; at night, they remain idle. A smarter strategy might involve opening these stations to the public or to ride-hailing drivers during off-hours, provided that security and access-control issues can be resolved.
In Table 2, I summarize the average daily energy and the peak hour for each district and station type. This table helps to highlight where the bottlenecks are and where new investments should be directed.
| District | Dominant station type | Avg daily energy (MWh) | Peak hour | Weekend peak hourly load (relative to weekday peak) |
|---|---|---|---|---|
| Core | Public parking | 3.42 | 12:00–13:00 | 0.91 |
| Hexi | Residential | 2.28 | 19:00–20:00 | 1.02 |
| Lunan | Industrial | 2.91 | 09:00–10:00 | 0.45 |
| Ludong | Mixed / Dedicated | 3.77 | 10:00–11:00 | 0.78 |
3. Behavioral Patterns of Electric Vehicle Users
When I analyze the behavior of individual electric vehicle users, a very consistent picture emerges. A large fraction of customers charge during daytime, and their charging sessions are strongly correlated with commuting periods. For example, the proportion of sessions starting between 7:00 AM and 10:00 AM is significantly higher on weekdays than on weekends. Conversely, the share of sessions starting after 10:00 PM is relatively low, indicating that most users are not deliberately scheduling overnight public charging. This could be due to a lack of overnight public charging options or to user habits that have not yet adapted to smart charging.
Customer loyalty manifests in the repeated usage of the same set of charging stations. I calculate for each customer the number of distinct stations visited per month. The distribution shows that about 45% of active monthly users visit only one station, 27% visit two stations, and the remaining 28% visit three or more. The single-station users are likely those who have access to a charger at their workplace or residential community. Multi-station users are typically ride-hailing or commercial fleet drivers who need to top up opportunistically whenever the battery level drops below a threshold. The station choice behavior is influenced by three main factors: price, convenience, and service quality. Price is a dominant factor because electric vehicle users are generally conscious of operating costs. Convenience includes the location of the station relative to their regular routes and the waiting time involved. Service quality involves the reliability of the charging equipment, the rated power, and additional amenities such as sheltered parking, lighting, and nearby toilets.
I also investigate the price elasticity of demand at the session level. Because the effective unit price varies across stations and over time, I exploit this natural variation to estimate how the average energy per session responds to the unit price. The scatter diagram in Figure (inserted above) indicates an inverted-U shape: at very low prices, the average energy is not the highest; it reaches an optimum at intermediate prices, and then declines as prices rise. This counterintuitive pattern at low prices may be due to the fact that very cheap stations are located in less accessible places or are older and have lower-power chargers, which discourages high-energy sessions. Alternatively, some low-price stations are reserved for fleet vehicles whose charging needs are predetermined and not necessarily large.
The customer typology analysis suggests that electric vehicle users can be broadly divided into four segments:
- Workplace commuters: They charge during office hours, usually at industrial campuses or corporate parking lots. They are price-sensitive but value the convenience of charging while working. They exhibit high routine and rarely need fast charging, because their vehicles remain parked for several hours.
- Residential overnight chargers: They charge at home or in residential communities, usually in the evening after returning from work. They have limited access to workplace chargers and rely on slow overnight charging. Their demand is regular and periodic.
- Opportunistic public users: These are shoppers, visitors, and errand-runners who charge while parked at shopping malls or public parking lots. Their sessions are short and often use fast chargers. They have high requirements for charging speed and payment convenience.
- Commercial fleet drivers: This group encompasses ride-hailing drivers, taxi operators, and logistics vehicles. They need high energy throughput during the day and often charge at dedicated fast-charging hubs. They are highly rational in their station choices because time is money.
Each of these segments exhibits a different willingness to pay. Fleet drivers often accept higher prices if the charger power is high and the waiting time is negligible, because a loss of 30 minutes of driving time costs more than the additional charging fee. Residential users, by contrast, prefer the lowest possible price, even if that means charging overnight. These insights are fundamental for designing incentive schemes and for the deployment of storage systems to shave peak loads.
4. Case Studies of Representative Charging Stations
To make my analysis more tangible, I select four representative stations that epitomize each station category. These stations are not necessarily the largest by number of piles, but they have the highest total energy within their category during the study period. I describe their customer profiles, load shapes, and operational circumstances in detail. The four selected stations are: an industrial park station (designated as industrial), a residential community station, a public parking station, and a dedicated high-usage station. I do not disclose their exact names or addresses, but I describe them in generic terms that allow operators to recognize similar stations in their own portfolios.
4.1 Industrial Campus Station
The industrial campus station is located inside a large high-technology manufacturing site. Its customers are predominantly employees of the factories and offices within the campus, together with a smaller number of delivery and logistics vehicles. I observe a bimodal daily load curve with a prominent peak between 7:00 and 9:00 AM, when employees arrive and plug in their cars before starting their shifts, and a second broader peak between 4:00 and 6:00 PM, when employees unplug and leave for home. There is also a minor peak around 1:00 PM, likely corresponding to logistics vehicles that charge during the drivers’ lunch break. The total number of sessions is 9,119 over the observation period, with an average energy per session of 29.8 kWh and an average duration of 1.9 hours. The total energy supplied is approximately 272 MWh, and the total revenue is about 354 thousand RMB.
The employees who charge at this station tend to be extremely regular. Weekday charging dominates; the weekend activity is less than 15% of the weekday level. The station is effectively idle on Sundays. Because the charging piles are relatively low-power (typically alternating-current piles with 7 kW to 22 kW), the vehicles remain plugged in for several hours, which is perfectly compatible with the work schedule. These users are aware of the time-of-use tariff and would shift their charging to earlier hours if the price were lower, but many start charging immediately upon arrival because they want to ensure a full battery for the drive home.
From a managerial perspective, the industrial campus station would benefit from expanding the number of piles if the parking lot has available space, because the utilization ratio is almost 90% during workdays. Another strategy is to introduce a reservation system that allows employees to book a charging slot in advance, reducing the morning scramble. A third idea is to install some high-power DC piles near the canteen or main gate to serve logistics vehicles and employees who are in a hurry.
4.2 Residential Community Station
The residential community station is located in a residential complex that does not have sufficient private garage parking for all residents. The customers are almost entirely local residents, with two subgroups: conventional commuters and ride-hailing drivers. Commuters return home in the evening and charge overnight or until the next morning. Ride-hailing drivers may come back at unconventional times because their schedules vary, and they often take advantage of low electricity prices during the deep night. The average session at this station lasts 3.6 hours, which is longer than at any other type of station, because the installed power is generally 7 kW alternating-current chargers that deliver at a slow rate. The average energy per session is 25.3 kWh. Over the observation period, the station recorded 3,667 sessions, supplying 93 MWh and generating 130 thousand RMB in revenue.
The load profile is sharply concentrated in the evening hours. From 18:00 to 23:00, the utilization rates rise steeply, because returning residents compete for a limited number of piles. Once a vehicle is plugged in, it often remains until the battery is full, which may be early in the morning. This causes a problem: a resident who arrives home at 22:00 may find all piles occupied by vehicles that have finished charging but have not yet been moved. This behavior, sometimes called “charger hogging,” is a major source of dissatisfaction. I notice that the average charging duration is longer than necessary because many drivers plug in when they still have a high state of charge, and they do not return until the next morning.
To improve the efficiency of the residential station, I recommend implementing a time-of-use pricing schedule that strongly discourages parking after the charging process ends. Another effective mechanism is an idle fee: if a vehicle remains connected after reaching 100% state of charge for more than 15 minutes, the per-minute fee is charged at a high rate. In addition, the installation of smart charging piles with load-balancing capability could prevent overloading of the local distribution transformer. Possibly, a small-scale battery storage unit could be combined with the charging piles to store solar energy generated during the day and discharge it to the vehicles in the evening.
4.3 Public Parking Station
The public parking station is one of the busiest in the entire network. It is located in a commercial area near shopping malls, office buildings, and a subway station. The customers are highly heterogeneous, including shopping mall customers, nearby workers, and ride-hailing drivers who use the DC fast chargers that are available. This station has a mix of 60 kW DC fast chargers and some 7 kW AC chargers. The average session duration is only 0.8 hours, which reflects the high effectiveness of the DC chargers. The station records 49,432 sessions over the observation period, with an average energy of 27.6 kWh. The total energy supplied is 1,365 MWh, and the total revenue is about 1.92 million RMB.
Because the station is located in an area with heavy traffic, the load profile is broad and includes two daily peaks: one in the late morning (11:00–14:00) and one in the evening (18:00–21:00). Weekend peaks are actually higher than weekday peaks, because shoppers and leisure visitors arrive in large numbers on Saturday and Sunday. This public station also serves as a surrogate for home charging for many residents who live in apartment buildings without dedicated chargers. The ride-hailing drivers often visit in the mid-afternoon, when the station is less crowded, to avoid queues. In contrast to industrial and residential stations, a significant share of public-station users are first-time visitors. They are less loyal and more likely to be influenced by real-time availability shown on mobile applications.
The main challenge for the public parking station is to manage the peak-hour waiting time. Once queues form, some drivers become frustrated and may never return. An effective solution is to increase the number of high-power charging piles, but this is limited by the available space and the grid connection capacity. Another strategy is to coordinate with nearby shopping malls to offer reduced parking fees for drivers who charge during off-peak hours, thus shifting some demand to the afternoon or late evening. Dynamic pricing that varies by the minute based on real-time occupancy could also smooth the arrival pattern by encouraging drivers to defer charging by a couple of hours when the price is lower.
4.4 Dedicated High-Usage Station
The dedicated high-usage station is operated under contract with a large enterprise, serving a fleet of electric vans and shockingly also several ride-hailing vehicles. It is, in essence, a fleet depot rather than a public station, though it may be accessible under specific agreements. The station has high-power charging equipment and ample parking spaces. Over the observation period, it records 49,077 sessions, with an exceptionally high average energy per session of 42.7 kWh. The average session duration is 0.9 hours, implying that the station achieves a very fast throughput. The total energy supplied is 2,097 MWh, and the total revenue is 2.63 million RMB.
The load pattern at this station is determined by the operational schedules of the fleet. The vehicles tend to return to the depot in waves, for example around noon and later in the evening. During these waves, the station struggles to keep pace; between waves, the chargers are idle. Because the fleet operator has a strong incentive to maximize vehicle utilization, the charging schedule is optimized to ensure that every truck is fully charged before its next departure. This requires not only sufficient charging power but also sophisticated queue-management software. The fleet operator is interested in vehicle-to-grid (V2G) technology, because the depot’s vehicles are parked at predictable times and could provide ancillary services to the grid. I see a large potential for integrating such a dedicated station into a virtual power plant, as the station has a single counterparty and standardized power electronics.
In Table 3, I summarize the key metrics of the four representative stations. The contrast among the four categories is stark and supports the need for differentiated planning.
| Station category | Number of sessions | Avg energy per session (kWh) | Avg duration (h) | Total energy (MWh) | Total revenue (thousand RMB) |
|---|---|---|---|---|---|
| Industrial campus | 9,119 | 29.8 | 1.9 | 272 | 354 |
| Residential community | 3,667 | 25.3 | 3.6 | 93 | 130 |
| Public parking | 49,432 | 27.6 | 0.8 | 1,365 | 1,920 |
| Dedicated high-usage | 49,077 | 42.7 | 0.9 | 2,097 | 2,630 |
The important insights are as follows. First, the dedicated high-usage station and the public parking station deliver much higher total energy than the other two categories. Second, the average duration is shortest at the public parking station, thanks to DC fast charging. Third, the residential station has the lowest throughput per charger and suffers from inefficient use of chargers after the vehicles are fully charged. Fourth, the industrial station is a weekday phenomenon and is underutilized at weekends. A balanced portfolio of charging stations would strategically combine these types so that the system as a whole exhibits a smoother load profile.
5. Price and Demand: Regression Analysis
Price is a central instrument for influencing the charging behavior of electric vehicle users. In this section, I delve deeper into the relationship between the unit price (in RMB per kilowatt-hour) and the average energy dispensed per session. The reason I focus on the average energy per session instead of total energy is that the latter is heavily affected by the number of chargers available at each price point. By using the average energy per session, I effectively normalize the station size and consider the behavioral response of an individual user facing a particular price level. The unit price is calculated as the total fee divided by the energy dispensed. The prices in the dataset range from about 0.5 RMB/kWh to more than 2.5 RMB/kWh, although the majority of observations fall within 0.7 and 1.8 RMB/kWh. I aggregate the sessions into bins of 0.05 RMB/kWh and compute the average energy per session within each bin.
The resulting scatter plot (Figure 4 in the original version) reveals an asymmetric pattern. For prices between 0.8 and 1.3 RMB/kWh, the average energy is high, between 30 and 35 kWh. As the price increases beyond 1.3 RMB/kWh, the average energy gradually declines, falling to about 10 kWh at 2.0 RMB/kWh. Interestingly, for prices below 0.8 RMB/kWh, the average energy also falls, which suggests that extremely low prices are associated with limited charging sessions or with older/malfunctioning chargers. The peak average energy occurs around a unit price of 1.27 RMB/kWh, according to the polynomial regression. This observation is important because it tells us that a moderate pricing level maximizes the energy delivered per session.
5.1 Polynomial Regression
I model the relationship using a quadratic polynomial:
$$ E(p) = \beta_0 + \beta_1 p + \beta_2 p^2 + \varepsilon $$
where \(p\) is the unit price, \(E\) is the average energy per session, and \(\varepsilon\) is the error term. I fit the model by ordinary least squares. The estimated coefficients are such that the fitted curve opens downwards. I determine the optimal price by setting the derivative to zero:
$$ \frac{dE}{dp} = \beta_1 + 2\beta_2 p = 0 \quad \Rightarrow \quad p^* = -\frac{\beta_1}{2\beta_2} $$
The fitted optimal price is 1.27 RMB/kWh. The coefficient of determination \(R^2\) for the polynomial regression over the range 0.5–1.8 RMB/kWh is 0.89, indicating a strong fit. However, the quadratic shape is too light-tailed because it cannot perfectly capture the sharp drop at the right tail. I could extend the model to a cubic term, but the qualitative interpretation will be unchanged.
5.2 Support Vector Regression
To relax the global functional form, I apply Support Vector Regression (SVR) with a radial basis function kernel. SVR searches for a tube of width \(\epsilon\) around the true function, within which errors are not penalized. The objective is to minimize:
$$ \min_{\mathbf{w},b} \frac{1}{2}\|\mathbf{w}\|^2 + C \sum_{i=1}^{N} (\xi_i + \xi_i^*) $$
subject to the constraints that the prediction error is inside the \(\epsilon\)-tube except for slack variables. I set \(\epsilon = 0.1\) and \(C = 1.0\). The fitted SVR model predicts an optimal price of 1.19 RMB/kWh, which is close to the polynomial result. The advantage of SVR is that it captures the local plateau and the change in curvature. However, the SVR curve exhibits local fluctuations that may not be statistically meaningful.
5.3 Decision Tree Regression
I also fit a decision tree regression with the mean squared error as the splitting criterion and a maximum depth of 5. The tree learns a piecewise constant function. The optimal price according to the decision tree is 0.93 RMB/kWh. The tree is useful in identifying the discrete regimes of charging behavior. For example, in the node with price lower than 0.93 RMB/kWh, the average energy is around 20 kWh, while in the node from 0.93 to 1.4 RMB/kWh, the average energy jumps to 34 kWh. Above 1.4, the average energy falls below 15 kWh. The decision tree highlights the knife-edge nature of the price response. The discontinuity is plausible because users may have a mental anchor of a “fair price” around 1 RMB/kWh. When the price exceeds that threshold, they reduce the charged energy significantly, perhaps because they only charge enough to reach their next destination.
5.4 Model Comparison
Table 4 reports the coefficient of determination and the mean absolute percentage error for the three models over the same test data.
| Model | Optimal price (RMB/kWh) | R² (0.5–1.8 range) | MAPE (%) |
|---|---|---|---|
| Polynomial (degree 2) | 1.27 | 0.89 | 4.8 |
| SVR (RBF kernel) | 1.19 | 0.87 | 5.1 |
| Decision tree (depth 5) | 0.93 | 0.82 | 7.3 |
In practice, I recommend that pricing strategies operate within the window of 0.93–1.27 RMB/kWh, as the three models agree that this band contains the optimum. The precise position within the band should be determined by ancillary objectives: if the goal is to maximize revenue, I would lean toward the upper end (1.25–1.30), because the average energy is near the top and the multiplier effect on the total fee will be beneficial. If the goal is to maximize social welfare and electric vehicle adoption, the lower end (0.95–1.10) may be preferable because customers perceive the price as fair and will be more inclined to use public stations regularly.
My analysis of price elasticity also reveals an interesting asymmetry: the response to a price increase from the optimum is much more pronounced than the response to a decrease. This implies that operators should avoid raising the price above 1.3 RMB/kWh unless there are overwhelming grid constraints. Instead, they should use temporal differentiation: keeping the price moderate during off-peak hours and high during on-peak hours can flatten the load curve without significantly reducing total energy demand.
6. Forecasting Future Charging Demand Using Prophet
Accurate forecasting of the electric vehicle charging load is fundamental for the stable operation of the distribution network. In this section, I construct a time-series model using the Prophet framework introduced by Taylor and Letham. The Prophet model is suitable for daily or hourly data with strong weekly seasonality and holiday effects. It assumes the time series is the sum of three components:
$$ y(t) = g(t) + s(t) + h(t) + \varepsilon_t $$
where \(g(t)\) is a trend function, \(s(t)\) captures periodic seasonality (e.g., weekly, yearly), \(h(t)\) captures holiday effects, and \(\varepsilon_t\) is the idiosyncratic error. The trend \(g(t)\) is specified as a piecewise linear or logistic growth curve with change points. The seasonality is represented as a Fourier series:
$$ s(t) = \sum_{n=1}^{N} \left( a_n \cos\left(\frac{2\pi n t}{P}\right) + b_n \sin\left(\frac{2\pi n t}{P}\right) \right) $$
where \(P\) is the period (for weekly effects, \(P=7\) days; for annual effects, \(P=365.25\) days), and \(N\) is the number of Fourier terms. I set the maximum change points to 5 for the trend, allowing the trend to adapt to structural breaks such as the rollout of new subsidies or the opening of additional charging stations. Holiday effects are modeled with dummy variables for each important holiday date, including a window before and after the event.
6.1 Decomposition Results
The model is fitted to the daily total charging energy from January 1, 2023, to April 30, 2024. The decomposition yields three compelling findings.
Trend component: The overall trend of daily energy shows a gentle increase over the observed period, followed by a slight stabilization. The average daily energy in the first quarter of 2023 is about 34.5 MWh, while in the same quarter of 2024 it is 37.0 MWh, an annual growth of about 7%. This growth is not purely linear; it accelerates starting in the third quarter of 2023, when the total fleet of electric vehicles in the service area grew quickly. In the long run, the trend is expected to continue upward as electric vehicle adoption grows, but the saturation effect will likely appear as the market approaches maturity.
Seasonal component: The weekly seasonality is strong. The model estimates that charging energy on weekdays is 15% to 20% higher than the weekly average, while on Saturdays the deviation is around +5% and on Sundays around −12%. This is consistent with the observation that Sunday is the day of lowest charging activity, likely because many electric vehicle users stay at home or avoid the traffic. There is also a yearly seasonal pattern: during December, January, and February, the daily energy is about 10–15% higher than during June, July, and August. The winter peak in energy consumption is caused by reduced battery efficiency and the use of resistive heaters, which increases the kilowatt-hours needed per kilometer driven.
Holiday component: The biggest negative holiday effect occurs during the Chinese Spring Festival, when a large fraction of the population leaves the city for family reunions in their hometowns. The charging energy during the Spring Festival week drops by 30–40%. Similarly, the National Day holiday causes a drop of around 35%. Short holidays such as the Qingming Festival (Tomb-Sweeping Day) or Labour Day are less disruptive, with a decrease of approximately 10%. Public holidays also shift the load from weekdays to weekends because the official holidays are often positioned immediately before or after a weekend.
6.2 Out-of-Sample Forecasting
After fitting the Prophet model, I generate a forecast for the period from May 1, 2024, to April 29, 2025, along with 95% prediction intervals. The forecast and historical data are shown in Figure (not reproduced, but the corresponding numerical summaries are). The yearly total charging energy for 2024 is predicted to be 14.28 GWh, with a 95% confidence interval from 13.85 GWh to 14.62 GWh. This is nearly identical to the 14.30 GWh generated in 2023, an increase of only 0.1%. However, the first four months of 2024 were already 8% above the same period in 2023, suggesting that the full-year rate of growth may be higher than my initial estimate. One caveat in modelling is that the dataset ends in April 2024; the forecast for the later months is subject to greater uncertainty. The average daily charging energy in 2024 is 39,000 kWh, with the 95% prediction interval roughly between 36,000 and 42,000 kWh.
The maximum single-day energy is predicted to occur around December 20, 2024, with an expected value of 62,000 kWh and the 95% interval between 58,000 and 66,000 kWh. This peak is driven by the combination of cold weather, pre-holiday travel, and the fact that December has fewer public holidays than February. The minimum daily energy is predicted to occur around September 17, 2024, with an expected 13,000 kWh (95% interval: 11,000–15,000 kWh). This minimum may be explained by a U-shaped pattern of moderate weather and no school activities.
To measure the forecast accuracy, I backtest the model on the historical period using a sliding window. The mean absolute percentage error (MAPE) of the daily energy forecast is 4.2%, and the empirical coverage of the 95% intervals is 92.3%, which is close to the nominal level. Hence, I consider the forecast reliable for planning purposes.
The output of the Prophet model also helps in evaluating how different operational decisions affect the charging demand. For example, if the electricity price is lower during the night, some demand may shift from the daytime to the night. However, capturing this endogenous shift would require coupling the time-series model with the price-elasticity model of Section 5. I leave such an integrated framework for future work.
7. Policy Recommendations and Strategic Implications
In light of the insights revealed by the analysis, I offer several evidence-based recommendations for charging station operators, electric power utilities, and urban planners.
First, update the spatial layout to rebalance the utilization of chargers. The industrial-area stations are heavily used during the day but idle at night; the residential stations show the opposite pattern. One practical solution is to create “shared charging corridors” between adjacent industrial parks and residential communities. By making the enterprise charging piles available to off-hours users through a smart locking mechanism, the underutilized capacity could be leveraged without undermining the primary user base. Furthermore, the core district of the city would benefit from increasing the density of charging piles, as their utilization is high throughout the day. However, I must be cautious about the limitation of the grid connection; I should combine expansion plans with energy storage systems to smooth out the peak demand. The dataset demonstrates that the current distribution of stations does not perfectly align with the true charging demand, so a systematic reallocation of mobile or modular chargers may be necessary.
Second, implement time-varying prices to shape the charging behavior of electric vehicle users. My regression analysis shows that a high price significantly reduces the energy charged per session. Therefore, it is wise to avoid price spikes above 1.5 RMB/kWh unless they are restricted to the true peak hours of the grid. A well-designed time-of-use tariff has three tiers: a generous off-peak price below 1.0 RMB/kWh for overnight charging (00:00–06:00), a moderate shoulder price between 1.0 and 1.2 RMB/kWh for midday and late evening, and an on-peak price between 1.3 and 1.5 RMB/kWh for the system peak periods. For peak hours that coincide with the system load (early afternoon and evening), the price premium should be visible to the customer, but it should not be so high that it creates social exclusion. My models show that the optimal price is close to the 1.27 RMB/kWh threshold; pricing above that value is likely to suppress the average charging volume dramatically.
Third, strengthen the demand-response infrastructure and integrate vehicle-to-grid (V2G) elements. Since the charging load is predictable at the daily and weekly timescales, a battery storage system at the charging station can mitigate the local grid impact. For dedicated fleet stations, V2G is especially compelling. The fleet vehicles are parked for several hours each day, and their batteries are large and often idle. By allowing the grid operator to draw electricity from these vehicles during peak load periods and charging them during off-peak periods, the station can become a prosumer that supports the grid. This requires a control system and an appropriate market mechanism, but the technology is mature. In the long run, the wide adoption of V2G could convert electric vehicles into mobile energy-storage units, reducing the need for stationary storage.
Fourth, adopt a differentiated service strategy for each station type. For industrial park chargers, I recommend shifting from a first-come-first-served system to a booking or reservation system. This will ease the morning surge and allow employees to plan their charging schedule. For residential stations, the installation of smart charging piles that enable scheduled charging is crucial. The users can set their desired departure time and the charger will optimize the charging rate, ensuring that the car is fully charged by then without exceeding the contractual power limit. For public parking stations, I recommend adding more DC fast chargers, because demand is time-sensitive. I also recommend improving the user experience by providing easy-to-read charging status screens and an integrated mobile-payment system. For the dedicated fleet station, I recommend a periodic revision of the charging timetable to accommodate changes in fleet routes and to balance the charging load throughout the day, thereby avoiding simultaneous charging waves that stress the transformer.
Fifth, use data analytics for predictive maintenance. My dataset reveals that some charging piles produce repeated sessions of very short duration (less than 1 minute), which are often due to communication errors or broken connectors. Monitoring the distribution of session durations per pile can help identify faulty equipment and reduce downtime. The big-data infrastructure can be used to generate an automated alert when a pile’s failure rate exceeds a specified threshold. In this way, the reliability of charging facilities increases, contributing to higher customer satisfaction. Reliability is also a key determinant of electric vehicle user trust; it is not just about the number of chargers but about their consistent availability.
Table 5 consolidates the main findings and corresponding actions for each station type based on the empirical evidence obtained in this study.
| Station type | Observed behavior | Core problem | Recommended action |
|---|---|---|---|
| Industrial campus | Bimodal weekday peaks; high morning surge | Idle on weekends; narrow peak window | Facilitate booking; expand capacity; share with commute-neighbors |
| Residential community | Long sessions; evening congestion | Charger hogging; low turnover | Install smart AC charging piles; add idle fee; time-of-use scheduling |
| Public parking | Short sessions; high throughput; weekend peaks | Queue formation at peak; heterogenous users | Increase DC fast piles; dynamic pricing; offline app integration |
| Dedicated high-usage | Very high average energy; predictable waves | Transformer capacity; scheduling constraints | Roll out V2G; optimize scheduling; integrate battery storage |
8. Study Limitations and Future Work
This study is based on a rich dataset from a single metropolitan area, which may limit the generalization of the results to other regions with different climate, urban form, and subsidy policies. The data cover only public charging stations and do not include private home charging sessions that might represent a substantial segment of electric vehicle charging. Therefore, my conclusions are biased toward customers who lack access to home chargers or who are on the road. Access to private charging data is restricted for privacy reasons, but an integrated analysis of public and residential charging would provide a more complete picture of electric vehicle usage.
The price regression is associative, not causal. Charging stations with different prices also differ in location, power, and age. Although I have accounted for some observable variables, there may be unobserved confounders, such as the surrounding parking availability. A quasi-experimental design, such as a staggered price change across stations, would be the next step to strengthen the causal claim of price elasticity. My polynomial and SVR models are suitable for interpolation, but their extrapolation beyond the observed price range is uncertain. In particular, if the unit price exceeds 2.5 RMB/kWh, the average energy per session could become close to zero, but I lack data to estimate the curve at that extreme.
The Prophet forecasting model is a univariate time-series model that does not explicitly include the growth in the electric vehicle population. I implicitly account for it through the trend component, but a structural model would link the future charging demand to the anticipated number of new electric vehicle registrations. If a large fleet operator signs a contract with a new set of vehicles and charges at the same station, my forecast would under-predict the future demand. Similarly, the introduction of new charging stations in the region will change the demand distribution across the existing stations. Forecasting at the network level is more robust, but station-level forecasting is subject to reallocation shocks that are difficult to model purely from past data.
My recommendation regarding pricing is based on the observation that the average energy per session peaks at 1.27 RMB/kWh. However, the objective function of a station operator is not necessarily to maximize energy per session. If the operator aims to maximize profit or social welfare, the optimal price could be lower or higher. The profit function would take into account the marginal cost of electricity, the fixed costs of the charging equipment, and the opportunity cost of charger time. I leave the full profit-maximization analysis for future work, possibly using a revenue-management framework.
Another limitation is that I do not measure the waiting time at charging stations. Waiting time is a critical factor in customer satisfaction. In a station with 20 dozens of chargers, the perceived congestion is a function of the queue length and the charging speed of each pile. My available dataset lacks precise queueing information, but the utilization rate and the session duration can provide an indirect measure. I observe that the residential station has an average duration of 3.6 hours and nearly 90% utilization during peak hours. This suggests an extremely high waiting time for late arrivals, which cannot be gleaned from the transaction logs alone. Future research should combine transaction records with occupancy-sensor data to estimate the waiting time more accurately.
Despite these limitations, I believe that the methodology proposed in this paper is robust and applicable. The use of large-scale charging data, combined with regression and time-series models, provides a solid foundation for the planning and operation of electric vehicle charging networks. The novel aspects of this work, such as the multiscenario segmentation by station type, the price-demand curve, and the time-series decomposition, offer actionable insights that are not commonly found in simpler analyses. The main contribution of the paper is to show that electric vehicle charging behavior is neither random nor chaotic; it is a structured, policy-responsive, and predictable phenomenon that can be effectively managed with the aid of big data.
In the context of the global push toward carbon neutrality, the electrification of transportation is inevitable. Yet the speed of adoption hinges on the convenience, affordability, and reliability of charging infrastructure. By understanding the behavior of electric vehicle users, we can design a charging network that is not only efficient but also equitable. I hope that the insights derived from this study will inspire more data-driven research and practical actions in the field of electric mobility.
To conclude, I emphasize that every element of my analysis—from data cleansing to model building—is centered on the behavior of the electric vehicle. The term “electric vehicle” is not just a subject; it is the core of a socio-technical transition. Charging behavior is the observable manifestation of how early adopters and mainstream users cope with the constraints and opportunities of this transition. As the market expands, the behavior will evolve, and the models will need to be updated. However, the fundamental principles of price sensitivity, temporal routines, spatial dependence, and heterogeneity across segments will likely retain their explanatory power.
I hope this study provides a scientific basis for charging infrastructure investment, encourages the adoption of dynamic demand management, and ultimately contributes to the sustainable and resilient development of the electric vehicle industry in China and around the world.