In recent years, the adoption of electric vehicles has surged globally, with China emerging as a key market due to supportive policies and technological advancements. As consumers increasingly prioritize energy efficiency and driving range, accurately assessing the real-world energy consumption of electric vehicles becomes critical. Traditional methods rely on standardized driving cycles like the China Light-duty vehicle Test Cycle (CLTC) and the Worldwide harmonized Light vehicle Test Cycle (WLTC), which are derived from mixed data of internal combustion engine and electric vehicles. However, these cycles may not fully capture the unique driving behaviors of electric vehicle users, such as rapid acceleration and regenerative braking, leading to discrepancies between laboratory tests and actual energy use. To address this, we leverage cloud-based big data to construct user-specific driving cycles that reflect real-road conditions, enabling more accurate energy consumption analysis and development guidance for electric vehicles.
Our approach involves collecting anonymized vehicle data from a diverse fleet of electric vehicles across different regions, models, and time periods. This data-driven methodology ensures that the constructed driving cycles represent authentic user behaviors, accounting for variations in driving styles, road types, and environmental factors. By analyzing these cycles, we aim to bridge the gap between standardized tests and real-world performance, ultimately enhancing the development of energy-efficient electric vehicles. In this article, we detail the entire process—from data preprocessing and cycle synthesis to energy simulation and validation—providing insights into how cloud big data can revolutionize electric vehicle energy management.
The foundation of our study lies in the collection and preprocessing of massive datasets from electric vehicles. We obtain time-series data, including vehicle speed and timestamps, from cloud servers after rigorous desensitization to protect user privacy. This data spans multiple electric vehicle models, regions, and seasons, allowing us to investigate the impact of these factors on driving patterns. For instance, we consider datasets from spring and winter in various Chinese regions to account for seasonal variations. The raw data often contains anomalies, such as sudden speed spikes or drops, which we address through preprocessing. Specifically, we identify and correct outliers by replacing them with the average speed from adjacent time points, and we smooth idle segments by setting fluctuating speeds to zero during prolonged stops. This ensures the integrity of the data for subsequent analysis.
To construct representative driving cycles, we segment the preprocessed data into short trip segments. Each segment is defined as the interval between two consecutive idle points, encompassing acceleration, deceleration, cruising, and idle phases. This segmentation aligns with methodologies used in standard cycles like CLTC but is adapted for electric vehicle specifics. We further categorize these segments into three road scenarios based on speed profiles: urban roads (low-speed), urban expressways or elevated roads (medium-speed), and highways (high-speed). Each scenario is subdivided into congested and uncongested states to capture traffic conditions. This hierarchical classification allows for a nuanced representation of real-world driving, as summarized in the table below.
| Scenario Type | Road Description | Traffic State | Symbol |
|---|---|---|---|
| Scenario 1 | Urban Roads | Congested | 1.1 |
| Scenario 1 | Urban Roads | Uncongested | 1.2 |
| Scenario 2 | Urban Expressways | Congested | 2.1 |
| Scenario 2 | Urban Expressways | Uncongested | 2.2 |
| Scenario 3 | Highways | Congested | 3.1 |
| Scenario 3 | Highways | Uncongested | 3.2 |
Next, we extract 16 feature parameters from each short trip segment to quantify its kinematic characteristics. These parameters include time-based metrics (e.g., total time, acceleration time), speed-based metrics (e.g., maximum speed, average speed), and acceleration-based metrics (e.g., maximum acceleration, average deceleration). The full list of parameters is provided in the table below, which serves as the basis for subsequent dimensionality reduction and clustering. For electric vehicles, parameters like relative positive acceleration are particularly important as they relate to the powertrain’s energy demands during dynamic driving.
| Parameter Symbol | Unit | Description |
|---|---|---|
| T | s | Total duration of the segment |
| Ta | s | Total time spent in acceleration |
| Td | s | Total time spent in deceleration |
| Te | s | Total time spent in cruising |
| Ti | s | Total time spent idling |
| S | km | Total distance traveled |
| vmax | km/h | Maximum speed observed |
| amax | m/s² | Maximum acceleration |
| amin | m/s² | Maximum deceleration (negative value) |
| vm | km/h | Average speed over the segment |
| vme | km/h | Average speed while moving |
| aa | m/s² | Average acceleration during acceleration phases |
| ad | m/s² | Average deceleration during deceleration phases |
| vsd | km/h | Standard deviation of speed |
| asd | m/s² | Standard deviation of acceleration |
| arpa | m/s² | Relative positive acceleration, indicating energy demand |
Given the high dimensionality and correlation among these parameters, we apply Principal Component Analysis (PCA) to reduce the data complexity. PCA transforms the original 16 features into a set of uncorrelated principal components, retaining most of the variance while minimizing redundancy. We standardize the parameters to have zero mean and unit variance before computing the covariance matrix and its eigenvectors. The eigenvalues are sorted in descending order, and we select the top k components that cumulatively explain over 95% of the total variance. As shown in the table below, the first six principal components achieve a cumulative contribution rate of 95.68%, making them suitable for subsequent clustering analysis. This step not only improves computational efficiency but also enhances the robustness of the driving cycle construction.
| Principal Component | Eigenvalue | Variance Contribution Rate (%) | Cumulative Contribution Rate (%) |
|---|---|---|---|
| 1 | 7.709 | 48.18 | 48.18 |
| 2 | 4.544 | 28.40 | 76.58 |
| 3 | 1.142 | 7.14 | 83.72 |
| 4 | 1.008 | 6.30 | 90.02 |
| 5 | 0.555 | 3.47 | 93.49 |
| 6 | 0.350 | 2.19 | 95.68 |
With the reduced feature set, we perform cluster analysis using the K-Means algorithm to group similar short trip segments within each scenario. For every road scenario (urban, expressway, highway), we partition the segments into two clusters: congested and uncongested. This unsupervised learning approach initializes cluster centers randomly and iteratively assigns segments to the nearest center based on Euclidean distance in the principal component space. The process continues until convergence, resulting in six distinct categories (e.g., urban-congested, urban-uncongested). This clustering captures the intrinsic patterns in electric vehicle driving data, such as frequent stops in congested urban areas or sustained high speeds on uncongested highways.
To synthesize the final driving cycles, we compute the centroid of each cluster—representing the average feature vector—and select the short trip segments closest to these centroids. The selection is proportional to the time distribution of each category in the original data, ensuring that the synthesized cycle reflects real-world usage. For example, if congested urban driving accounts for 30% of the total time, we include segments accordingly. The selected segments are concatenated to form a continuous speed-time profile, resulting in a user driving cycle that encompasses low-speed, medium-speed, and high-speed phases with varying traffic conditions. This method generates multiple cycles for different electric vehicle models, regions, and seasons, enabling comparative analysis.
We analyze the characteristics of the constructed user driving cycles by comparing them with standard cycles like CLTC and WLTC. The user cycles exhibit higher maximum acceleration and deceleration values, indicating more aggressive driving behavior typical of electric vehicles. For instance, the maximum acceleration in user cycles reaches $$2.63 \, \text{m/s}^2$$ compared to $$1.47 \, \text{m/s}^2$$ for CLTC and $$1.58 \, \text{m/s}^2$$ for WLTC. Similarly, the average acceleration during acceleration phases is $$0.74 \, \text{m/s}^2$$ for user cycles, versus $$0.45 \, \text{m/s}^2$$ for CLTC and $$0.54 \, \text{m/s}^2$$ for WLTC. These differences highlight the need for cycles that better represent the dynamic performance of electric vehicles. The table below summarizes key parameters for a representative user cycle and the standard cycles, demonstrating the heightened kinematic intensity in user-driven data.
| Parameter | User Cycle (Example) | CLTC | WLTC |
|---|---|---|---|
| Total Time (s) | 1800 | 1800 | 1800 |
| Maximum Speed (km/h) | 119.59 | 114 | 131.3 |
| Maximum Acceleration (m/s²) | 2.63 | 1.47 | 1.58 |
| Maximum Deceleration (m/s²) | -2.21 | -1.47 | -1.49 |
| Average Acceleration (m/s²) | 0.74 | 0.45 | 0.54 |
| Average Deceleration (m/s²) | -0.78 | -0.49 | -0.58 |
| Relative Positive Acceleration (m/s²) | 0.28 | 0.17 | 0.15 |
To evaluate the energy consumption implications, we develop a simulation model using AVL Cruise software. The model includes components such as the vehicle body, powertrain, battery system, and tires, parameterized with real-world data from a typical electric vehicle. Key parameters include a curb mass of 2300 kg, a wheelbase of 2800 mm, a drag area of $$0.75 \, \text{m}^2$$, and low-rolling-resistance tires. The battery is a lithium-ion ternary type, and the drivetrain uses a fixed-ratio reducer. We simulate energy consumption over the user driving cycles, CLTC, and WLTC to quantify differences. The model computes the battery state of charge (SOC) dynamically, as described by the equation: $$\Delta SOC = -\frac{\int P_{\text{bat}} \, dt}{C_{\text{bat}}}$$ where $$P_{\text{bat}}$$ is the battery power and $$C_{\text{bat}}$$ is the battery capacity. The SOC profile shows fluctuations during acceleration and regeneration, with an overall decreasing trend.
We investigate the influence of factors like electric vehicle model, region, and season on energy consumption by simulating multiple user cycles. For instance, cycles derived from different models in the same region and season show minimal energy variation—less than 2% from the mean. Similarly, cycles from the same model in different regions or seasons exhibit negligible differences, indicating that user driving behavior is the dominant factor. This robustness allows us to generalize the user cycles for broader applications in electric vehicle development. The simulation results reveal that the user cycles consume more energy than CLTC but less than WLTC. Specifically, the user cycle energy consumption is approximately 9.7% higher than CLTC and 8.4% lower than WLTC, aligning with the more aggressive driving patterns captured in the user data.
To validate the simulation findings, we conduct real-world energy consumption tests on a chassis dynamometer following the GB/T 18386.1-2021 standard. The test vehicle is an electric model equipped with data loggers to measure current and voltage at the battery terminals. Tests are performed under controlled conditions (23 ± 5°C) after vehicle soak, and the dynamometer simulates road load based on coast-down data. We run the CLTC cycle and the user driving cycle sequentially, recording distance traveled and energy consumed to calculate kWh/100 km. The results confirm that the user cycle consumes 8.8% more energy than CLTC, consistent with the simulation trend. This discrepancy stems from the higher kinematic intensity of user cycles, which increases energy losses during frequent acceleration and deceleration events.
In conclusion, our study demonstrates the value of cloud big data in constructing realistic driving cycles for electric vehicles. By leveraging large-scale, diverse datasets, we create cycles that accurately reflect user behavior, leading to more reliable energy consumption estimates. The user cycles show significant differences from standard cycles like CLTC and WLTC, with higher energy demands due to more dynamic driving. This approach can guide the development of electric vehicles by providing targets for energy efficiency optimization and range prediction. Future work could explore the integration of additional factors, such as weather and topography, to further enhance the accuracy of electric vehicle energy models. As the adoption of electric vehicles grows, especially in markets like China, data-driven methodologies will play a crucial role in achieving sustainable mobility.