www.nature.

com/scientificdata

OPEN UrbanEV: An Open Benchmark

Data Descriptor Dataset for Urban Electric Vehicle
Charging Demand Prediction
1✉
Han Li1,4, Haohao Qu2,4, Xiaojun Tan1, Linlin You , Rui Zhu 3
& Wenqi Fan2

The recent surge in electric vehicles (EVs), driven by a collective push to enhance global environmental
sustainability, has underscored the significance of exploring EV charging prediction. To catalyze
further research in this domain, we introduce UrbanEV — an open dataset showcasing EV charging
space availability and electricity consumption in a pioneering city for vehicle electrification, namely
Shenzhen, China. UrbanEV offers a rich repository of charging data (i.e., charging occupancy, duration,
volume, and price) captured at hourly intervals across an extensive six-month span for over 20,000
individual charging stations. Beyond these core attributes, the dataset also encompasses diverse
influencing factors like weather conditions and spatial proximity. Comprehensive experiments have
been conducted to showcase the predictive capabilities of various models, including statistical, deep
learning, and transformer-based approaches, using the UrbanEV dataset. This dataset is poised to
propel advancements in EV charging prediction and management, positioning itself as a benchmark
resource within this burgeoning field.

Background & Summary

In recent times, there has been a notable surge in the popularity of electric vehicles (EVs), driven by the goals of
diminishing reliance on fossil fuels, ameliorating air quality, addressing global warming concerns, and advancing
the United Nations’ Sustainable Development Goals pertaining to carbon neutrality1,2. The Global EV Outlook
for 2024, as documented by the International Energy Agency (IEA)3, highlights a persistent increase in electric
car sales, projecting a potential volume of approximately 17 million units by 2024. This figure would constitute
over one-fifth of the total global car sales. Despite the manifold advantages associated with the electrification of
vehicles, this widespread adoption presents formidable obstacles to the dependability and robustness of urban
power grids and transportation infrastructures. Amongst these limited charging spaces, particularly in dense
cities, creates the most anxiety for EV users, resulting in critical problems such as long parking cruising time and
additional traffic congestion4. Another illustration lies in the unbalanced electricity consumption caused by reg-
ular EV charging behaviors5,6. This fluctuation in demand during rush hour can impact the stability of the grid,
potentially causing voltage fluctuations, frequency variations, and other issues that may affect the reliability of
the power supply7,8. Given the ongoing expansion of EV charging infrastructure and the limited sustainability of
relying solely on electricity to address associated challenges, many intelligent services have surfaced to manage
EV charging demand9–12. These services include dynamic pricing mechanisms13, collaborative charging resource
allocation strategies14,15, and parking guidance systems16.
The prediction of electric vehicle charging demand, which serves as a cornerstone for enabling intelligent
EV services, has garnered growing attention from academic and industrial communities17–19. Accurate pre-
dictions of regional EV charging demand can enable the provision of tailored parking suggestions to drivers
in transit, alleviating range anxiety concerns, while regulatory bodies can implement dynamic pricing strat-
egies to enhance energy efficiency. To achieve accurate prediction, related studies have attempted to leverage
advanced data analysis techniques and predictive modeling to forecast the future charging demand for elec-
tric vehicles18,20. It typically involves analyzing historical charging data, considering factors such as the num-
ber of electric vehicles, their charging patterns, infrastructure availability, and external variables like weather

1
Sun Yat-sen University, School of Intelligent Systems Engineering, Shen Zhen, 518107, China. 2The Hong Kong
Polytechnic University, Department of Computing, Hong Kong, Hong Kong. 3Institute of High Performance
Computing (IHPC), Agency for Science, Technology and Research (A*STAR), 138632, Queenstown, Republic of
Singapore. 4These authors contributed equally: Han Li, Haohao Qu. ✉e-mail: [email protected]

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 1

www.nature.com/scientificdata/ www.nature.com/scientificdata

Dataset Time Span Granularity Scope #EVSEs Additional Factors*

High-resolution EV Data 33
Nov. 2014‐Oct. 2015 1hour / 105 stations User Habit, Location Type
Perth and Kinross27 Sep. 2016‐Aug. 2019 1min County / /
ACN-Data28 Sep. 2018‐Dec 2019 15min Site 112 piles /
One-Year EV Charging Fleet29 May 2018‐Jul. 2020 10min Site 30 piles /
Multi-faceted EV Charging30 Sep. 2021‐Sep. 2022 15min / 2,119 piles Location Type
UrbanEV39 Sep. 2022‐Feb. 2023 5min/1hour Urban 17,532 piles Price, Weather, POI, Geo-Data

Table 1. Comparison of representative open datasets associated with electric vehicle charging, where “#”
denotes the number of certain headers, “/” denotes that the item is not specified, and EVSE refers to Electric
Vehicle Supply Equipment. * The “Additional Factors” item outlines the feature set provided beyond the
charging data itself.

conditions and time of day21. For example, a recent study investigated the performance of three conventional
machine learning models, namely Long Short-Term Memory, Auto-Regressive Moving Average, and Multiple
Layer Perceptron, in short-term EV charging forecasting22. Another in-depth research explored the effect of
electricity price on electric vehicle charging demand by conducting correlation tests and estimating its price
elasticity23. More recently inspired by the success of incorporating spatial information with temporal patterns in
traffic prediction, spatial-temporal EV charging demand prediction has emerged as an attractive research topic
in the literature. Representative examples include HSTGCN-EV24 and ChatEV25: The former one incorporated
two heterogeneous graphs (i.e., a demand-based graph and a geographic graph) to improve predictive precision,
while the latter one unified spatial and temporal factors within natural language and harnessed Large Language
Models (LLMs) for regional EV charging prediction.
Although relevant research continues to expand, a well-structured open-source benchmark dataset that
includes a wide array of features and establishes standardized comparison settings for predicting EV charging
demand is still absent. Existing studies face several critical limitations with their data. Table 1 illustrates the
comparison of representative publicly available datasets from various aspects. Firstly, most of them rely solely
on charging data or consider only a limited number of factors, neglecting a comprehensive assessment of other
potential influences26–29. Secondly, although numerous temporal patterns crucial for EV charging demand pre-
diction have been identified, the current datasets are insufficient for delving into spatial analysis in EV charging
behaviors30–33. Lastly, the diverse settings observed across studies introduce substantial variations, hindering a
fair comparison of new techniques, frameworks, and models within relevant research19. These limitations hinder
the advancements of EV charging prediction and related intelligent services in the era of big data.
To fill the gap, we present UrbanEV, an open dataset of EV charging in Shenzhen, China. The dataset
compiles comprehensive information for a total of 1,682 public charging stations with 24,798 charging piles,
shown in Fig. 1. After applying various data processing techniques, we refine the dataset to encompass 1,362
charging stations with 17,532 public charging piles, making them well-suited for charging demand prediction.
Specifically, it provides three charging data (i.e., occupancy, duration, and volume), four dynamic factors (i.e.,
electricity price, service price, weather conditions, and time of day), three spatial attributes (i.e., adjacency,
distance, and coordinates), and four static coefficients (i.e., point of interest, area, pile number, and station
number). Moreover, our dataset covers the period from 1 September 2022 to 28 February 2023, encompassing
six months with hourly granularity. This level of detail enables the exploration of short-, mid-, and long-term
forecasting scenarios. Lastly, based on the station-level information, we further group the data into traffic zones,
offering a new perspective on exploring regional EV charging patterns. As shown in Fig. 2, stations located in
a specific traffic zone are integrated, and an adjacency matrix among neighboring zones can be built corre-
spondingly to represent the spatial relationship. Making this dataset publicly accessible is intended to equip
researchers, policymakers, and industry practitioners with the essential information needed for the effective
and sustainable management of EV charging. This initiative aligns with national priorities and contributes to the
overarching global sustainability objectives.

Methods
To build a comprehensive and reliable benchmark dataset, we conduct a series of rigorous processes from data
collection to dataset evaluation. The overall workflow sequentially includes data acquisition, data processing,
statistical analysis, and prediction assessment. As follows please see detailed descriptions.

Study area and data acquisition. Shenzhen, a pioneering city in global vehicle electrification, has been
selected for this study with the objective of offering valuable insights into electric vehicle (EV) development that
can serve as a reference for other urban centers. This study encompasses the entire expanse of Shenzhen, where
data on public EV charging stations distributed around the city have been meticulously gathered. Specifically,
EV charging data was automatically collected from a mobile platform used by EV drivers to locate public charg-
ing stations. Through this platform, users could access real-time information on each charging pile, including
its availability (e.g., busy or idle), charging price, and geographic coordinates. Accordingly, we recorded the
charging-related data at five-minute intervals from September 1, 2022, to February 28, 2023. This data collection
process was fully digital and did not require manual readings. Furthermore, to delve into the correlation between
EV charging patterns and environmental elements, weather data (i.e., air temperature Ta, atmospheric pressure
P, and relative humidity h) for Shenzhen city were acquired from two meteorological observatories situated in

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 2

www.nature.com/scientificdata/ www.nature.com/scientificdata

Fig. 1 Spatial distribution of 1,682 public charging stations and 24,798 charging piles in the UrbanEV dataset.

(a) N

(b) (c) (d) 1 2 3 4 5 6

14 2 7 14 7
13 1 0 836 1252 1698 994 1083
2
12 13 4 8 4 8 2 836 0 1509 2405 806 1338
11 3 1 3 1252 1509 0 1326 2123 2335
11 1 5 9
3 9 10 6 4 1698 2405 1326 0 2677 2549
10 6 5 5 994 806 2123 2677 0 648
1-hop 2-hop 6 1083 1338 2335 2549 648 0

Fig. 2 Data visualization of the filter areas. (a) illustrates the distribution of charging piles at the regional level.
(b) provides an enlarged view of the CBD dynamic pricing areas. (c) depicts the node graph derived from the
enlarged view, showcasing the center node of the enlarged region, its 1-hop and 2-hop neighbors, as well as their
adjacency relationships. (d) presents the distance matrix (in meters) of the 1-hop neighbors of the center node
in the enlarged region.

the airport and central regions, respectively. These meteorological data are publicly available on the Shenzhen
Government Data Open Platform. Thirdly, point of interest (POI) data was extracted through the Application

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 3

www.nature.com/scientificdata/ www.nature.com/scientificdata

Feature Max Mean Median Min Q1 Q3 Std

o (%) 1 0.24 0.21 0 0.14 0.31 0.17
d (hour) 175.83 9.64 5.00 0.00 1.50 12.50 12.78
Day v (kWh) 13637.50 180.61 37.62 0.00 10.50 111.42 593.53
pe (CNY) 1.80 0.98 0.99 0.23 0.86 1.10 0.18
ps (CNY) 1.45 0.70 0.76 0.00 0.70 0.76 0.15
EV
o (%) 1.00 0.30 0.26 0.00 0.16 0.40 0.19
d (hour) 207.58 13.58 6.00 0.00 1.92 16.58 19.54
Night v (kWh) 16732.50 275.72 46.08 0.00 12.25 147.00 968.24
pe (CNY) 1.80 0.93 0.94 0.23 0.79 1.06 0.23
ps (CNY) 1.45 0.73 0.76 0.00 0.74 0.76 0.14
Ta (°C) 34.70 21.18 21.40 8.30 16.40 26.30 5.92
Weather P (mmHg) 773.80 762.49 762.40 751.10 759.90 765.50 4.18
h (%) 97.00 69.33 72.00 21.00 58.00 82.00 16.49
Number of Piles 373.00 63.75 48.00 4.00 24.00 84.00 56.63
Others Distance (meter) 78680.32 21972.31 20476.39 0.00 12184.82 29764.97 12721.15
Number of Neighbors 10.00 4.36 4.00 0.00 3.00 6.00 1.82

Table 2. Data statistics of the UrbanEV Dataset. This table presents data across three dimensions: EV, Weather,
and Others. First, UrbanEV provides EV charging-related data, i.e., occupancy ratio (o), charging duration
(d), charging volume (v), and charging price (including electricity price pe and service fee ps), with statistics at
the traffic zone level. In the weather dimension, it offers three representative features that have the potential
to influence charging behaviors, namely air temperature (Ta), atmospheric pressure (P), and relative humidity
(h) across the entire study area. Lastly, information on pile number, adjacency, and distance within or between
traffic zones is also incorporated into the dataset.

Programming Interface Platform of AMap.com, along with three primary types: food and beverage services, busi-
ness and residential, and lifestyle services. Lastly, the spatial and static data were organized based on the traffic
zones delineated by the sixth Residential Travel Survey of Shenzhen34. The collected data contains detailed spati-
otemporal information that can be analyzed to provide valuable insights about urban EV charging patterns and
their correlations with meteorological conditions.

Processing raw information into well-structured data. To streamline the utilization of UrbanEV
dataset, we harmonize heterogeneous data from various sources into well-structured data with aligned temporal
and spatial resolutions. An overview of the descriptive statistics of the processed data is presented in Table 2. This
process can be segmented into two parts: the reorganization of EV charging data and the preparation of other
influential factors.

EV charging data. The raw charging data, obtained from publicly available EV charging services, pertains to
charging stations and predominantly comprises string-type records at a 5-minute interval. To transform this raw
data into a structured time series tailored for prediction tasks, we implement the following three key measures:

• Initial Extraction. From the string-type records, we extract vital information for each charging pile, such
as availability (designated as “busy” or “idle”), rated power, and the corresponding charging and service fees
applicable during the observed time periods. First, a charging pile is categorized as “active charging” if its
states at two consecutive timestamps are both “busy”. Consequently, the occupancy within a charging station
can be defined as the count of in-use charging piles, while the charging duration is calculated as the product
of the count of in-use piles and the time between the two timestamps (in our case, 5 minutes). Moreover, the
charging volume in a station can correspondingly be estimated by multiplying the duration by the piles’ rated
power. Finally, the average electricity price and service price are calculated for each station in alignment with
the same temporal resolution as the three charging variables.
• Error Detection and Imputation. Data quality is crucial for decision-making, advanced analytics, and
machine learning. Inaccuracies, often referred to as dirty data, can significantly undermine the reliability
of analysis or modeling efforts35. To improve the quality of our charging data, we identified several errors,
notably negative values for charging fees and inconsistencies between counts of occupied, idle, and total
charging piles. Records containing these anomalies were removed and treated as missing data. A two-step
imputation process was employed for missing values: forward filling replaced missing values using preceding
timestamps, followed by backward filling to fill gaps at the beginning of each time series. Additionally, out-
liers, which could significantly impact prediction performance, were detected using the interquartile range
(IQR) method36 for metrics such as charging volume (v), charging duration (d), and the rate of active charging
piles (o). To retain more original data and minimize the impact of outlier correction, we set the coefficient
to 4, instead of the default 1.5. Each outlier was then replaced with the mean of its adjacent valid values. This
preprocessing pipeline transformed the raw data into a structured and analyzable dataset.

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 4

www.nature.com/scientificdata/ www.nature.com/scientificdata

• Aggregation and Filtration. Building upon the station-level charging data that has been extracted and
cleansed, we further organize the data into a region-level dataset with an hourly interval providing a new per-
spective for EV charging behavior analysis. This is achieved by two major processes: aggregation and filtration.
First, we aggregate all the charging data from both temporal and spatial views: a. Temporally, we standardize
all time-series data to a common time resolution of one hour, as it serves as the least common denominator
among the various resolutions. This aims to establish a unified temporal resolution for all time-series data,
including pricing schemes, weather records, and charging data, thereby creating a well-structured dataset.
Aggregation rules specify that the five-minute charging volume (v) and duration (d) are summed within each
interval (i.e., one hour), whereas the occupancy (o), electricity price (pe), and service price (ps) are assigned
specific values at certain hours for each charging pile. This distinction arises from the inherent nature of these
data types: volume (v) and duration (d) are cumulative, while (o), (pe), and (ps) are instantaneous variables.
Compared to using the mean or median values within each interval, selecting the instantaneous values of (o),
(pe), and (ps) as representatives preserves the original data patterns more effectively and minimizes the influ-
ence of human interpretation. b. Spatially, stations and piles are aggregated based on the traffic zones delin-
eated by the sixth Residential Travel Survey of Shenzhen34. After aggregation, the resulting dataset includes
331 regions (also referred to as traffic zones) and 4344 timestamps. Variance tests and zero-value filtering
were then applied to exclude regions with negligible or no variation in charging data. Specifically, regions
with an occupancy variance below 0.001 or with more than 30% zero values were removed. As a result, 275
traffic zones, comprising 1,362 charging stations and 17,532 charging piles, were retained for further analysis,
as depicted in Fig. 2.

Other influential factors. Apart from the EV charging data, we also constructed a set of variables that might
influence charging behaviors37,38. These variables can be categorized into three classes, namely temporal factors,
spatial attributes, and static features. First and foremost, the temporal factors include three weather conditions:
air temperature (Ta), relative humidity (h), and atmospheric pressure (P). The raw weather data is collected from
two meteorological observatories located in the airport and central regions of Shenzhen, and they were further
organized into numeric data with the same hourly interval as the structured charging data. Notably, weather
data is shared across all charging stations and traffic zones. Spatial information, including the adjacency matrix
and distances, is computed using ArcGIS tools. Specifically, adjacency is determined by evaluating whether two
traffic zones share a boundary, based on the distance between their geometric centers. Additionally, UrbanEV
provides static features, such as Points of Interest (POI), area, and road length for each traffic zone. Only those
relevant to charging activities within the 275 selected zones, aligned with the structured charging data, are
retained.

Data Records
To enable further in-depth predictive analyses by researchers, the 1-hour resolution region-level dataset is pro-
vided as the primary dataset, with the 5-minute resolution version also made available in the Dryad repository39,
offering comprehensive access to time-series data at varying granularities. For consistency, all data are stored
as comma-separated value (.csv) files along with their corresponding header descriptions stored in .txt files.
Moreover, this dataset provides the geometry information of the studied areas in ArcGIS format (e.g., .shp, shx,
and .dbf files). Lastly, we have developed a benchmarking code framework for EV charging forecasting, compris-
ing program files or scripts written in Python (.py). Here is a detailed overview of these files:

• (occupancy.csv, duration.csv, and volume.csv) provide the EV charging occupancy ratio, duration, and volume,
in the studied areas, measured in %, hours, and kWh, respectively. The volume in volume.csv is derived from
the rated power of charging piles and may deviate from actual charging volumes. Nevertheless, it serves as the
foundation for validation in subsequent analyses, with volume-11kW.csv providing a vehicle-side estimation
as an alternative.
• (e_price.csv, s_price.csv) describe the electricity price and service fee, respectively, with a granularity of hour.
Both of them are units in Yuan/hour.
• (weather_central.csv and weather_airport.csv) store the weather data obtained from two different meteorolog-
ical stations located in the central area and the airport of Shenzhen city, respectively. Their header informa-
tion is presented in the file titled weather_header.txt.
• (Shenzhen.shp, Shenzhen.shx, Shenzhen.dbf) store geographic information in Shenzhen city in ArcGIS format,
using the WGS 1984 Albers projected coordinate system.
• (adj.csv, distance.csv) depict the adjacency relationships between traffic zones, along with their respective dis-
tances. The distances are computed as the Euclidean distance between the centroids of the zones, measured in
meters. In the adjacency file, a value of 1 indicates that two traffic zones are adjacent, otherwise 0.
• (inf.csv) contains several basis information for each zone, including pile capacity, longitude, latitude, the area
and perimeters of the zone (in meters).
• (poi.csv) contains information about Points of Interest throughout the studied city, collected in December
2022.
• (volume-11kW.csv) provides an alternative vehicle-side estimation of charging volume to mitigate potential
overestimation in volume.csv. Specifically, for direct current charging stations, the volume is calculated using
the standard power of the most commonly used electric vehicle, Tesla Model Y (11kW), instead of the rated
power of the charging pile.
• (main.py, models.py, utils.py, preprocessing.py) are the code files used in this work.

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 5

www.nature.com/scientificdata/ www.nature.com/scientificdata

Technical Validation
In order to validate UrbanEV’s efficacy in EV charging demand prediction, we are conducting a comprehensive
benchmarking test covering forecasting methods specifically designed for EV charging demand as well as meth-
ods supporting general time-series forecasting tasks. It is noted that the validation relies on a one-hour resolu-
tion dataset. Through a thorough comparison and analysis of these baselines, we seek to address three crucial
questions: First, Q1: Does the dataset effectively capture the temporal patterns in EV charging behaviors? Second,
Q2: Can UrbanEV accurately depict the spatial interplay among different areas? Finally, Q3: Are the identified
correlated factors instrumental in enhancing prediction accuracy?
In this validation, we compare three traditional forecasting methods, five deep learning models, and two
state-of-the-art Transformer-based predictors. The three conventional models include the last observation (LO),
Auto-regressive (AR), and Auto-regressive Integrated Moving Average(ARIMA) model. The six deep learning
models are listed as follows: a fully connected neural network (FCNN) is a classical network that has been used
to capture the non-linearity in time series; Long Short-Term Memory (LSTM), a representative recurrent neural
network, has been recently utilized for predicting electric vehicle charging demand22,40; Graph Convolutional
Network (GCN), a typical graph learning model, has also been employed for electric vehicle (EV) forecast-
ing tasks24. Expanding on the aforementioned achievements, there has been a recent integration of graph and
recurrent models to enhance predictive performance for EV charging demand. Accordingly, GCN-LSTM41 as
a hybrid model is included in the evaluation. Moreover, one advanced time-series forecasting method, namely
the Attention-Based Spatial-Temporal Graph Convolutional Network (ASTGCN)42, is utilized as well in our
study. Lastly, we evaluate two Transformer-based forecasting Models in our investigation, i.e., TimeNet43, and
TimeXer44. Comparing and analyzing the performance of these baselines can assist us in evaluating whether
UrbanEV can act as a competitive benchmark dataset for EV charging prediction tasks. This validation employed
time-series cross-validation to address the temporal characteristics of a six-month dataset, which, accordingly,
supports a six-fold approach. Specifically, each fold incrementally included one additional month of data, with
80% allocated to training and the remaining 20% equally divided between validation and testing sets. Model per-
formance was assessed using four complementary metrics45,46: 1) Root Mean Squared Error (RMSE), 2) Mean
Absolute Percentage Error (MAPE), 3) Relative Absolute Error (RAE), and 4) Mean Absolute Error (MAE).
Finally, the evaluation objective is established as the distribution prediction (spatial-temporal prediction) for
EV charging data, supplemented by experiments on node prediction to answer, and insights gained from the
factorial experiment.

Distribution Prediction. The results presented in Table 3 reveal several key observations regarding the pre-
dictive performance across different models. The performance can be categorized into three distinct categories,
consistent with the baseline model classification.
1) Statistical Models: This category includes models such as LO, AR, and ARIMA, which rely on simple
linear transformations to capture temporal dynamics. These models are computationally efficient due to their
simplicity but exhibit limited predictive accuracy compared to more advanced approaches. Their inability to
model complex nonlinear patterns in the data constrains their utility in scenarios requiring high precision.
2) Conventional Deep Learning Models: The second category encompasses models like Fully Connected
Neural Networks (FCNN), Long Short-Term Memory (LSTM) networks, and Graph Convolutional Networks
(GCN). These models incorporate nonlinear temporal modeling capabilities, leading to significant improve-
ments in predictive accuracy over statistical methods. Additionally, leveraging spatial information-such as
using the charging demand from surrounding regions-further enhances prediction performance. Models like
GCN-LSTM and ASTGCN demonstrate the benefits of joint spatiotemporal feature modeling, achieving supe-
rior results by capturing complex dependencies across both dimensions.
3) Transformer-Based Models: Representing the highest-performing category, Transformer-based architec-
tures dynamically capture intricate spatiotemporal interactions, effectively addressing limitations observed in
other methods. By leveraging attention mechanisms, these models exhibit a transformative potential for spatio-
temporal prediction tasks, delivering state-of-the-art performance. Their ability to adaptively focus on relevant
temporal and spatial features provides a robust framework for capturing the nuanced dynamics of EV charging
demand.
These findings underscore the pivotal role of accurately extracting temporal and spatial features to enhance
forecasting accuracy. More importantly, the results demonstrate that the dataset exhibits pronounced spati-
otemporal characteristics, and the application of nonlinear modeling techniques proves effective in predict-
ing EV charging-related metrics. Compared to independently modeling temporal or spatial features, the joint
modeling of spatiotemporal features significantly improves predictive performance. However, the performance
differences among various spatiotemporal prediction models highlight the inherent complexity of the dataset’s
spatiotemporal characteristics. This complexity necessitates the design of specialized models tailored to capture
these intricate patterns effectively. Hence, we advocate for a deeper investigation into spatiotemporal forecasting
models to unveil the underlying patterns in EV charging behaviors within the UrbanEV dataset.

Node Prediction. The results presented in Table 4 highlight the performance of various models across three
key metrics-charging occupancy (o), duration (d), and volume (v). First, the Transformer-based model, TimeXer,
demonstrates superior performance among the models evaluated, achieving the lowest RMSE, MAPE, RAE, and
MAE values across all metrics. Specifically, it achieves an RMSE of 0.07 for o, 2.73 for d, and 43.66 for v, signif-
icantly outperforming traditional statistical models (e.g., LO, AR, ARIMA) and deep learning methods (e.g.,
FCNN, LSTM). Second, the recurrent model, LSTM, shows competitive results to TimeXer for charging duration
d and volume v and outperforms other compared models in most cases. These two observations indicate that the

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 6

www.nature.com/scientificdata/ www.nature.com/scientificdata

RMSE(×10−2) MAPE(%)
Model 3h 6h 9h 12h Average 3h 6h 9h 12h Average
LO 9.75 12.52 14.65 15.45 13.09 25.39 39.07 50.92 56.70 43.02
AR 13.08 13.00 12.89 12.67 12.91 58.30 59.12 60.93 61.74 60.02
ARIMA 13.76 13.88 13.44 12.79 13.47 58.63 58.89 59.56 59.10 59.05
FCNN 9.47 10.74 10.95 9.79 10.24 40.59 50.12 52.67 46.22 47.40
LSTM 9.37 10.96 11.05 9.74 10.28 36.17 46.44 49.81 43.54 43.99
GCN 8.91 10.63 10.93 10.08 10.14 39.93 50.32 51.76 46.92 47.23
GCNLSTM 8.41 9.67 10.65 9.39 9.53 35.96 45.01 50.12 43.26 43.59
ASTGCN 9.15 10.61 10.92 9.83 10.13 35.67 46.02 49.37 47.52 44.64
TimesNet 9.00 9.59 9.92 9.64 9.54 31.65 35.19 37.58 36.37 35.20
TimeXer 8.32 9.38 9.89 9.39 9.24 26.13 33.20 36.47 35.14 32.74

RAE( × 10−2) MAE( × 10−2)

Model 3h 6h 9h 12h Average 3h 6h 9h 12h Average
LO 36.62 54.05 68.08 74.05 58.20 4.91 7.26 9.17 9.98 7.83
AR 67.54 67.07 66.66 65.65 66.73 8.99 8.98 8.96 8.83 8.94
ARIMA 70.82 70.90 69.21 65.78 69.18 9.46 9.52 9.31 8.84 9.28
FCNN 45.62 54.62 56.02 49.08 51.34 6.11 7.33 7.54 6.62 6.90
LSTM 43.51 54.57 55.85 48.27 50.55 5.82 7.32 7.52 6.50 6.79
GCN 45.61 55.69 57.34 52.64 52.82 6.11 7.48 7.73 7.11 7.11
GCNLSTM 41.79 50.13 55.86 48.12 48.97 5.59 6.73 7.52 6.48 6.58
ASTGCN 42.70 52.90 55.32 49.91 50.21 5.71 7.10 7.45 6.73 6.75
TimesNet 40.78 45.05 47.28 45.86 44.74 5.48 6.03 6.31 6.12 5.99
TimeXer 35.01 42.24 46.43 44.12 41.95 4.71 5.66 6.20 5.89 5.61

Table 3. Performance comparison of representative forecasting methods on the spatial-temporal occupancy

prediction task using the UrbanEV dataset. The result showcases that models incorporating both spatial and
temporal patterns can achieve superior predictive accuracy. This observation suggests that the UrbanEV dataset
exhibits pronounced spatiotemporal dependencies within EV charging data. The best and second best results in
each column are marked by Bold and underlined, respectively.

RMSE MAPE
Model o d v o d v
LO 0.10 4.35 68.53 0.41 0.57 0.58
AR 0.08 5.60 74.29 0.11 0.89 0.88
ARIMA 0.13 5.86 76.41 0.59 0.89 0.89
FCNN 0.11 3.62 55.52 0.54 0.56 0.57
LSTM 0.09 3.20 45.14 0.46 0.52 0.52
TimeXer 0.07 2.73 43.66 0.29 0.55 0.66

RAE MAE
Model o d v o d v
LO 0.79 0.78 0.78 0.07 3.17 51.71
AR 1.06 1.07 1.07 0.07 4.51 59.18
ARIMA 1.10 1.09 1.09 0.10 4.67 63.81
FCNN 1.05 0.78 0.79 0.09 2.89 44.44
LSTM 0.92 0.72 0.71 0.07 2.48 34.12
TimeXer 0.76 0.70 0.71 0.05 2.04 33.81

Table 4. Performance comparison of six representative forecasting methods in node prediction. It is evident
that the Transformer-based model, TimeXer, and the RNN-based model, LSTM, stand out with superior
performance. This observation indicates that the charging data offered by UrbanEV encompasses ample
temporal features. The best and second best results in each column are marked by Bold and underlined,
respectively.

UrbanEV dataset can serve as a suitable and trustworthy benchmark dataset for EV charging prediction, as the
compared models are appropriately ranked: namely, TimeXer > LSTM > others.

Factorial experiment. To investigate the influence of the five features mentioned above both individ-
ually and in combination, we conducted factorial experiments in various feature groups, including pairwise

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 7

www.nature.com/scientificdata/ www.nature.com/scientificdata

RMSE(×10−2) MAPE(%)
Model FCNN LSTM ASTGCN TimeXer Average FCNN LSTM ASTGCN TimeXer Average
None 9.47 9.37 9.15 7.95 8.99 40.59 36.17 35.67 24.81 34.31
P 9.59 9.25 8.96 8.41 9.05 39.24 35.77 32.57 26.19 33.44
Ta 9.48 9.28 8.87 8.47 9.02 36.32 34.88 32.41 27.06 32.67
h 9.44 9.23 8.98 8.48 9.03 38.70 34.62 32.41 26.97 33.18
pe 9.80 9.19 8.91 8.50 9.10 39.19 35.68 33.08 27.77 33.93
ps 9.86 9.21 8.97 8.69 9.18 39.80 35.75 32.20 28.39 34.04
P + ps 9.23 9.03 9.03 8.47 8.94 37.25 34.08 32.86 26.45 32.66
Ta + pe 9.12 9.03 8.88 8.51 8.88 36.07 34.16 33.29 26.46 32.50
h + Ta 9.20 9.09 9.04 8.59 8.98 36.58 33.42 34.78 26.57 32.84
ps + p e 9.24 9.02 8.93 8.49 8.92 38.05 35.26 33.22 26.60 33.28
all 9.12 9.11 9.22 8.63 9.02 36.87 36.21 39.09 27.44 34.91

RAE(×10 )−2
MAE(×10 )−2

Model FCNN LSTM ASTGCN TimeXer Average FCNN LSTM ASTGCN TimeXer Average
None 45.62 43.51 42.70 32.53 41.09 6.11 5.82 5.71 4.37 5.50
P 46.08 42.76 40.81 35.51 41.29 6.17 5.72 5.46 4.78 5.53
Ta 45.06 43.16 40.22 35.65 41.02 6.04 5.77 5.38 4.80 5.50
h 45.07 42.37 40.64 35.58 40.92 6.03 5.67 5.43 4.81 5.49
pe 47.41 42.83 40.75 35.31 41.57 6.34 5.73 5.45 4.78 5.58
ps 47.58 42.96 40.59 36.15 41.82 6.37 5.74 5.43 4.90 5.61
P + ps 43.84 41.64 41.43 36.01 40.73 5.87 5.57 5.54 4.85 5.46
Ta + p e 43.01 41.80 41.04 35.96 40.45 5.75 5.59 5.49 4.84 5.42
h + Ta 43.52 41.65 41.62 36.04 40.71 5.82 5.58 5.57 4.87 5.46
ps + pe 44.18 41.92 41.03 34.86 40.50 5.91 5.61 5.49 4.72 5.43
all 43.37 43.18 46.13 35.92 42.15 5.80 5.78 6.17 4.86 5.65

Table 5. Results of factorial experiments with ten factorial groups on EV charging demand. Utilizing the
occupancy ratio in each traffic zone as the target data, our analysis indicates that individual features make
a minimal contribution to prediction accuracy, and may even hinder performance in certain scenarios.
In contrast, the combination of features, notably air temperature (Ta) and electricity price (pe), can lead to
significant enhancements. The best and second best results in each column are marked by Bold and underlined,
respectively.

combinations (i.e., P + ps, Ta + pe, h + Ta, and ps + pe) to assess whether joint factors affect charging occupancy.
Additionally, we integrated all five features to explore whether the collective effect of external factors exhibits
consistent influences on o. First, as presented in Table 5, it indicates that the inclusion of individual features yields
minimal improvement in predicting EV charging demand and, in some cases, even deteriorates the prediction
accuracy. However, combinations of features prove to be significantly more effective in enhancing demand fore-
casting. Notably, pairings that include pe and ps, such as Ta + pe and ps + pe, demonstrate the strongest auxiliary
effects on prediction accuracy. This suggests that external factors like temperature and current charging costs
influence users’ charging decisions. For example, extreme temperatures-whether hot or cold-reduce the likeli-
hood of travel, subsequently lowering the demand for charging. Similarly, elevated electricity prices or service fees
may prompt users to either seek alternative charging stations or forgo charging altogether.
Second, it can be observed that the combination of pe and ps is particularly impactful, as these factors col-
lectively represent the total cost incurred during charging. The interplay between these two features effectively
captures the influence of charging costs on user behavior, which cannot be fully captured by either feature alone.
Consequently, the joint variation of pe and ps reflects users’ sensitivity to charging costs, making it a superior
predictor compared to single-factor variations.
Finally, integrating all five features to assist in predicting o does not necessarily improve the prediction accu-
racy. Although these features encompass various dimensions, such as weather conditions and price fluctuations,
providing diverse information, the prediction performance can deteriorate if the model fails to effectively pro-
cess these inputs. This highlights the importance and potential of developing advanced prediction models capa-
ble of handling multidimensional auxiliary information to further enhance forecasting accuracy.

Code availability
Our code used in this paper for the dataset setup, data analysis, and experiments can be found in a GitHub
repository at (https://github.com/IntelligentSystemsLab/UrbanEV).

Received: 29 January 2025; Accepted: 20 March 2025;

Published: xx xx xxxx

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 8

www.nature.com/scientificdata/ www.nature.com/scientificdata

References
1. Tran, M., Banister, D., Bishop, J. D. & McCulloch, M. D. Realizing the electric-vehicle revolution. Nature climate change 2, 328–333
(2012).
2. Crabtree, G. The coming electric vehicle transformation. Science 366, 422–424 (2019).
3. Agency, I. E. Global ev outlook 2024, https://www.iea.org/reports/global-ev-outlook-2024 (2024).
4. Pal, A., Bhattacharya, A. & Chakraborty, A. K. Planning of ev charging station with distribution network expansion considering
traffic congestion and uncertainties. IEEE Transactions on Industry Applications 59, 3810–3825 (2023).
5. Hussain, M. T., Sulaiman, N. B., Hussain, M. S. & Jabir, M. Optimal management strategies to solve issues of grid having electric
vehicles (ev): A review. Journal of Energy Storage 33, 102114 (2021).
6. Muratori, M. Impact of uncoordinated plug-in electric vehicle charging on residential power demand. Nature Energy 3, 193–201
(2018).
7. Chen, Q. et al. Afml: An asynchronous federated meta-learning mechanism for charging station occupancy prediction with biased
and isolated data. IEEE Transactions on Big Data 1–16, https://doi.org/10.1109/TBDATA.2024.3484651 (2024).
8. You, L. et al. Fmgcn: Federated meta learning-augmented graph convolutional network for ev charging demand forecasting. IEEE
Internet of Things Journal 11, 24452–24466, https://doi.org/10.1109/JIOT.2024.3369655 (2024).
9. Al-Ogaili, A. S. et al. Review on scheduling, clustering, and forecasting strategies for controlling electric vehicle charging: Challenges
and recommendations. Ieee Access 7, 128353–128371 (2019).
10. Vandet, C. A. & Rich, J. Optimal placement and sizing of charging infrastructure for evs under information-sharing. Technological
Forecasting and Social Change 187, 122205 (2023).
11. Gaete-Morales, C., Kramer, H., Schill, W.-P. & Zerrahn, A. An open tool for creating battery-electric vehicle time series from
empirical data, emobpy. Scientific data 8, 152 (2021).
12. Barbar, M., Mallapragada, D. S., Alsup, M. & Stoner, R. Scenarios of future indian electricity demand accounting for space cooling
and electric vehicle adoption. Scientific Data 8, 178 (2021).
13. Zhao, Z. & Lee, C. K. Dynamic pricing for ev charging stations: A deep reinforcement learning approach. IEEE Transactions on
Transportation Electrification 8, 2456–2468 (2021).
14. Aveklouris, A., Vlasiou, M. & Zwart, B. A stochastic resource-sharing network for electric vehicle charging. IEEE Transactions on
Control of Network Systems 6, 1050–1061 (2019).
15. Ji, N., Zhu, R., Huang, Z. & You, L. An urban-scale spatiotemporal optimization of rooftop photovoltaic charging of electric vehicles.
Urban Informatics 3, 4, https://doi.org/10.1007/s44212-023-00031-7 (2024).
16. Liu, S. et al. Reservation-based ev charging recommendation concerning charging urgency policy. Sustainable Cities and Society 74,
103150 (2021).
17. Zhang, X. et al. Deep-learning-based probabilistic forecasting of electric vehicle charging load with a novel queuing model. IEEE
transactions on cybernetics 51, 3157–3170 (2020).
18. Yaghoubi, E., Yaghoubi, E., Khamees, A., Razmi, D. & Lu, T. A systematic review and meta-analysis of machine learning, deep
learning, and ensemble learning approaches in predicting ev charging behavior. Engineering Applications of Artificial Intelligence
135, 108789 (2024).
19. Akshay, K., Grace, G. H., Gunasekaran, K. & Samikannu, R. Power consumption prediction for electric vehicle charging stations and
forecasting income. Scientific Reports 14, 6497 (2024).
20. Yi, Z., Liu, X. C., Wei, R., Chen, X. & Dai, J. Electric vehicle charging demand forecasting using deep learning model. Journal of
Intelligent Transportation Systems 26, 690–703 (2022).
21. Qu, H., Kuang, H., Wang, Q., Li, J. & You, L. A physics-informed and attention-based graph learning approach for regional electric
vehicle charging demand prediction. IEEE Transactions on Intelligent Transportation Systems (2024).
22. Wang, S. et al. Short-term electric vehicle charging demand prediction: A deep learning approach. Applied Energy 340, 121032
(2023).
23. Kuang, H. et al. Unraveling the effect of electricity price on electric vehicle charging behavior: A case study in shenzhen, china.
Sustainable Cities and Society 115, 105836 (2024).
24. Wang, S., Chen, A., Wang, P. & Zhuge, C. Predicting electric vehicle charging demand using a heterogeneous spatio-temporal graph
convolutional network. Transportation Research Part C: Emerging Technologies 153, 104205 (2023).
25. Qu, H. et al. Chatev: Predicting electric vehicle charging demand as natural language processing. Transportation Research Part D:
Transport and Environment 136, 104470 (2024).
26. Orzechowski, A. et al. A data-driven framework for medium-term electric vehicle charging demand forecasting. Energy and AI 14,
100267 (2023).
27. Team, O.-D. Electric vehicle charging station usage. Perth and Kinross Open Data https://data.pkc.gov.uk/ (2020).
28. Lee, Z. J., Li, T. & Low, S. H. ACN-Data: Analysis and Applications of an Open EV Charging Dataset. In Proceedings of the Tenth
International Conference on Future Energy Systems, e-Energy ’19, https://ev.caltech.edu/dataset (2019).
29. Obusevs, A., Domenico, D. D. & Korba, P. One year recordings of electric vehicle charging fleet. IEEE Dataport https://
doi.org/10.21227/fkap-fr63 (2021).
30. Baek, K., Lee, E. & Kim, J. A dataset for multi-faceted analysis of electric vehicle charging transactions. Figshare https://
doi.org/10.6084/m9.figshare.22495141.v1 (2023).
31. Baek, K., Lee, E. & Kim, J. A dataset for multi-faceted analysis of electric vehicle charging transactions. Scientific Data 11, 262 (2024).
32. Asensio, O. I., Lawson, M. C. & Apablaza, C. Z. Electric vehicle charging stations in the workplace with high-resolution data from
casual and habitual users. Scientific Data 8, 168 (2021).
33. Asensio, O. I., Lawson, M. C. & Apablaza, C. Z. High-resolution electric vehicle charging data from a workplace setting. Harvard
Dataverse https://doi.org/10.7910/DVN/QF1PMO (2021).
34. Zhou, J. & Ma, L. Analysis on the evolution characteristics of shenzhen residents? travel structure and the enlightenment of public
transport development policy. Urban Mass Transit 24, 63–68, https://doi.org/10.16037/j.1007-869x.2021.07.014 (2021).
35. Chu, X., Ilyas, I. F., Krishnan, S. & Wang, J. Data cleaning: Overview and emerging challenges. In Proceedings of the 2016
international conference on management of data, 2201–2206 (Association for Computing Machinery, New York, NY, USA, 2016).
36. Nouvellet, P. et al. Reduction in mobility and covid-19 transmission. Nature communications 12, 1–9 (2021).
37. Zhu, R. et al. Multi-sourced data modelling of spatially heterogenous life-cycle carbon mitigation from installed rooftop
photovoltaics: A case study in singapore. Applied Energy 362, 122957, https://doi.org/10.1016/j.apenergy.2024.122957 (2024).
38. Schober, P., Boer, C. & Schwarte, L. A. Correlation coefficients: appropriate use and interpretation. Anesthesia & analgesia 126,
1763–1768 (2018).
39. Li, H. et al. UrbanEV: An open benchmark dataset for urban electric vehicle charging demand prediction. Dryad https://
doi.org/10.5061/dryad.np5hqc04z (2025).
40. Shanmuganathan, J., Victoire, A. A., Balraj, G. & Victoire, A. Deep learning lstm recurrent neural network model for prediction of
electric vehicle charging demand. Sustainability 14, 10207 (2022).
41. Kim, H. J. & Kim, M. K. Spatial-temporal graph convolutional-based recurrent network for electric vehicle charging stations
demand forecasting in energy market. IEEE Transactions on Smart Grid 15, 3979–3993, https://doi.org/10.1109/TSG.2024.3368419
(2024).

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 9

www.nature.com/scientificdata/ www.nature.com/scientificdata

42. Guo, S., Lin, Y., Feng, N., Song, C. & Wan, H. Attention based spatial-temporal graph convolutional networks for traffic flow
forecasting. In Proceedings of the AAAI conference on artificial intelligence, vol. 33, 922–929 (2019).
43. Wu, H. et al. Timesnet: Temporal 2d-variation modeling for general time series analysis. In Proceedings of the Eleventh International
Conference on Learning Representations (2023).
44. Wang, Y. et al. Timexer: Empowering transformers for time series forecasting with exogenous variables. In Proceedings of the Thirty-
eighth Annual Conference on Neural Information Processing Systems(2024).
45. Li, J., Qu, H. & You, L. An integrated approach for the near real-time parking occupancy prediction. IEEE Transactions on Intelligent
Transportation Systems 24, 3769–3778 (2022).
46. Esling, P. & Agon, C. Time-series data mining. ACM Computing Surveys (CSUR) 45, 1–34 (2012).

Acknowledgements
This work was supported in part by the National Key Research and Development Program of China
(2023YFB4301900), Research Funds from the Department of Science and Technology of Guangdong Province
(2021QN02S161), and the GuangDong Basic and Applied Basic Research Foundation (2023A1515012895).”
Thank you for your time and support.

Author contributions
L.Y. and H.Q. conceived the experiment(s), H.L. conducted the experiment(s), and H.Q. and H.L. analyzed the
results. X.T. and R.Z. reviewed the manuscript and participated actively in its editing. All authors contributed to
the development of the manuscript and provided their approval for the final version.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to L.Y.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-
NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribu-
tion and reproduction in any medium or format, as long as you give appropriate credit to the original author(s)
and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed mate-
rial. You do not have permission under this licence to share adapted material derived from this article or parts of
it. The images or other third party material in this article are included in the article’s Creative Commons licence,
unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative
Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted
use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit
http://creativecommons.org/licenses/by-nc-nd/4.0/.

© The Author(s) 2025

Scientific Data | (2025) 12:523 | https://doi.org/10.1038/s41597-025-04874-4 10