Large-Scale EV Fleets Gain New Role in Grid Stability
As the global push toward electrification accelerates, electric vehicles (EVs) are no longer seen merely as a sustainable alternative to internal combustion engines. A groundbreaking study from researchers at Hefei University of Technology reveals that EVs, when intelligently aggregated, can serve as a pivotal resource in stabilizing power grids through multi-scenario auxiliary services. This new modeling approach, grounded in real-world data and advanced machine learning techniques, redefines how EVs interact with energy systems, positioning them as dynamic, dispatchable assets rather than passive loads.
The research, led by Wang Yangyang, a doctoral candidate at the Ministry of Education’s Research Center for Photovoltaic Systems Engineering, introduces a novel dual-layer clustering method to model the aggregation schedulable capacity (ASC) of large-scale EV fleets. This capacity—defined as the upper and lower limits of energy or power exchange between EVs and the grid under normal usage conditions—is critical for virtual power plants (VPPs) seeking to participate in peak shaving, frequency regulation, and voltage support. The findings, published in the journal Automation of Electric Power Systems, offer a scalable framework for integrating millions of EVs into provincial-level power dispatch systems.
With China’s EV fleet projected to reach nearly 100 million vehicles by 2030, storing an estimated 4 terawatt-hours of energy, the implications of unmanaged charging are profound. Uncoordinated EV charging could exacerbate grid imbalances, especially as renewable energy penetration increases. However, the study demonstrates that through vehicle-to-grid (V2G) technologies and smart aggregation, EVs can become a flexible, responsive resource capable of delivering gigawatt-scale regulation capacity.
Traditional approaches to modeling EV fleets have often treated them as monolithic entities or relied on probabilistic assumptions about charging behavior. These methods, while useful for small-scale applications, fall short when applied to regional or provincial grids where spatial distribution, charging infrastructure diversity, and user behavior variability play critical roles. The existing models typically assume uniform charging patterns or rely on synthetic data, limiting their real-world applicability.
Wang and his team address these limitations by proposing a data-driven, two-tiered clustering framework. Instead of treating EVs and charging stations separately, the model integrates both vehicle and charger characteristics into a unified “generalized energy storage system” (EV-GESS). This holistic view acknowledges that the dispatchability of an EV is not solely determined by its battery size or state of charge, but also by the type of charger it connects to, its geographic location, and the user’s charging habits.
The first layer of the model uses the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm to analyze over 1.78 million real charging records from 4,181 charging stations across a Chinese province in 2021. Unlike k-means or other centroid-based clustering methods, DBSCAN does not require a predefined number of clusters and is robust to outliers—making it ideal for handling the noisy, irregular nature of real-world charging data. The algorithm identifies distinct charging patterns, such as “morning-type,” “noon-type,” and “evening-type” behaviors, based on start times, durations, energy consumed, and idle time ratios.
One of the key innovations is the introduction of the “idle time ratio,” a metric that quantifies the proportion of time an EV remains connected to a charger without actively charging. A high idle time ratio indicates greater flexibility—such EVs can delay charging or even discharge back to the grid without inconveniencing the user. For example, overnight parkers who plug in after dinner but don’t need their vehicle until the next morning represent an ideal candidate for V2G services.
The second layer of the model applies an enhanced version of the Self-Organizing Map (SOM) neural network, modified with dimension selection and autoencoder-based feature reduction (AE-DSSOM). This stage clusters the charging stations themselves based on the distribution of charging profiles and fixed infrastructure parameters such as rated power and geographic coordinates. In peak shaving and frequency regulation scenarios, spatial location may be less critical, but for voltage regulation, the physical position of the charger directly influences which part of the distribution network can be stabilized.
By combining these two clustering stages, the model creates distinct EV-GESS aggregators (EV-GESSA), each representing a unique combination of temporal behavior and spatial-technical characteristics. This allows grid operators to target specific aggregators for different services—for instance, mobilizing high-idle-time, centrally located EVs for real-time frequency response, while reserving large-capacity, high-power stations for bulk energy shifting during peak hours.
The practical implications of this research are substantial. The study estimates that, based on 2021 data, the modeled EV fleet could provide an average adjustable power range of [-39.7, 10.5] megawatts at a 1-minute dispatch interval—sufficient to support regional frequency stability. When scaled to projected 2030 EV adoption rates, the potential grows to several gigawatts, rivaling the output of conventional peaking power plants.
Moreover, the model accounts for the dynamic nature of dispatchable capacity. As EVs charge or discharge in response to grid signals, their available headroom for further adjustment changes. The researchers incorporate this feedback loop into their ASC calculations, ensuring that real-time dispatch decisions are based on accurate, up-to-date capacity estimates. This is particularly important for ancillary services that require sustained power output over several minutes, such as frequency regulation, where the ability to maintain a set power level for 3 to 30 minutes is essential.
The study also examines long-term trends in EV flexibility. Weekly, monthly, and annual analyses reveal seasonal and behavioral patterns. For instance, dispatchable capacity tends to dip on weekends, likely due to reduced commuting and charging demand. However, national holidays such as Labor Day and National Day see a surge in activity, possibly linked to increased travel and longer parking durations. These insights enable more accurate forecasting and strategic planning for grid operators.
Another significant contribution is the validation of algorithmic choices. The team compares DBSCAN, k-means, SOM, and DSSOM across both clustering layers, evaluating performance using the Davies-Bouldin Index (DBI), a measure of cluster separation and compactness. Results show that DBSCAN achieves the best balance of speed and accuracy in the first layer, processing over 1.7 million records in under 15 seconds with a DBI of 0.7247—outperforming k-means (0.8262) and standard SOM (0.7354). In the second layer, where data dimensionality is higher but volume is lower, AE-DSSOM achieves the lowest DBI (0.8556), demonstrating superior clustering quality despite a slightly longer computation time.
The integration of autoencoders in the second stage proves particularly effective. By reducing the dimensionality of input features, the model avoids the “curse of dimensionality” and enhances the neural network’s ability to discern meaningful patterns. This is crucial when dealing with 12-dimensional data that includes both behavioral ratios and geographic coordinates. The autoencoder learns a compressed representation that captures the most salient features, improving both clustering speed and accuracy.
From a policy and economic standpoint, the research underscores the financial viability of EV-based grid services. Current compensation rates for ancillary services in China range from 300 to 800 yuan per megawatt-hour. Even with the 2021 dataset—representing only about 30% of the province’s total charging activity—the potential hourly revenue runs into tens of thousands of yuan. By 2030, with full fleet integration, the economic value could reach billions annually, creating a strong incentive for utilities, automakers, and consumers to participate in V2G programs.
However, the authors acknowledge limitations. The model is based on historical charging data and does not yet incorporate real-time driver responses to dispatch signals. User behavior may change when financial incentives or mandatory grid requests are introduced. Additionally, battery degradation concerns and charger availability remain barriers to widespread V2G adoption. Future work will focus on predictive modeling of ASC and field trials to validate the theoretical framework.
The broader significance of this research lies in its scalability and adaptability. While the study focuses on a single province, the methodology can be applied to any region with sufficient charging infrastructure data. As smart meters, connected vehicles, and cloud-based energy platforms become ubiquitous, the inputs required for such models—start and end times, energy delivered, charger type—will be increasingly available.
This shift represents a fundamental change in how we view transportation and energy systems. EVs are no longer just consumers of electricity; they are mobile storage units, capable of providing grid services at scale. The transition from passive loads to active participants requires sophisticated modeling, robust communication protocols, and market mechanisms that fairly compensate users. This study provides a critical piece of that puzzle.
Grid operators are already exploring pilot programs that leverage EV fleets for frequency regulation. In Europe, projects like eV2g in Denmark and the UK’s EV Energy Systems have demonstrated the technical feasibility of V2G. In the United States, utilities such as Pacific Gas & Electric and Southern California Edison are testing bidirectional charging with fleets of electric school buses and municipal vehicles. The Chinese study adds a new dimension by showing how large-scale, heterogeneous EV populations can be systematically organized and optimized for multiple services.
The dual-layer clustering approach also has implications beyond grid services. It could inform urban planning, helping cities identify optimal locations for fast-charging hubs based on actual usage patterns. It could support utility investment decisions, guiding the deployment of transformers and distribution equipment in areas with high EV concentration. And it could enable personalized energy services, where drivers receive tailored charging recommendations based on their historical behavior and grid needs.
As the energy transition progresses, the line between transportation and power systems will continue to blur. Vehicles will not only be powered by electricity but will also help balance the grid that powers them. The work of Wang Yangyang and his colleagues at Hefei University of Technology provides a roadmap for making this vision a reality—transforming millions of individual charging events into a coordinated, intelligent, and highly valuable resource for the modern power grid.
The study sets a new benchmark for EV-grid integration research, combining rigorous data science with practical engineering insights. It moves beyond theoretical simulations to deliver a model grounded in real-world operations, validated against actual charging behavior. By doing so, it bridges the gap between academic research and industry application, offering a tool that can be directly used by grid operators, aggregators, and policymakers.
In conclusion, the future of grid stability may well depend on the intelligent coordination of electric vehicles. As battery capacities grow, charging infrastructure expands, and digital connectivity improves, the potential of EVs as distributed energy resources will only increase. The research presented here is not just a technical achievement—it is a step toward a more resilient, efficient, and sustainable energy future.
Wang Yangyang, Mao Meiqin, Yang Cheng, Zhou Kun, Du Yan, Nikos D. Hatziargyriou, Hefei University of Technology, Automation of Electric Power Systems, DOI: 10.7500/AEPS20230627012