New AI Model Identifies EV Charging Patterns with Limited Data
A groundbreaking study has introduced a novel artificial intelligence-driven model capable of identifying electric vehicle (EV) charging behavior patterns using minimal user data. This advancement, developed by researchers from North China Electric Power University and State Grid Smart Internet of Vehicles Co., Ltd., promises to revolutionize how utilities and charging operators engage EV owners in grid-supportive activities. As the global EV fleet surpasses 20 million units, the strain on power distribution networks intensifies, making intelligent management of charging behavior more critical than ever. Traditional methods of analyzing user habits often require extensive datasets, raising concerns over privacy, data security, and regulatory compliance—especially under strict frameworks like the EU’s GDPR and California’s CCPA. However, this new approach circumvents these challenges by leveraging a hybrid framework combining cloud modeling and fuzzy Petri nets, enabling accurate user classification without the need for deep personal data access.
The research, published in the October 2024 issue of Electric Power Construction, addresses a persistent barrier in vehicle-to-grid (V2G) integration: low user participation. Despite the theoretical benefits of V2G—such as load balancing, frequency regulation, and peak shaving—real-world adoption remains limited. Many EV owners are hesitant to participate due to concerns about battery degradation, lack of financial incentives, or simply because the process feels too complex or intrusive. The authors argue that one of the root causes is the one-size-fits-all approach to user engagement. Without a clear understanding of individual charging habits, utilities struggle to design effective incentive programs. Some users may be willing to shift their charging to off-peak hours for a small discount, while others might respond better to gamified rewards or priority access to fast chargers. The key, the researchers suggest, lies in segmentation—identifying distinct user archetypes and tailoring strategies accordingly.
To achieve this, the team proposed a three-tier classification system based on observable charging patterns. The first category, “energy-responsive users,” are those who consistently charge during peak market periods but have limited flexibility in adjusting their timing. These users typically plug in when they arrive at work or home and leave their vehicles connected for extended durations, making them ideal candidates for long-term energy market participation. The second group, “power-responsive users,” exhibit high elasticity in both charging time and energy consumption. They frequently adjust their behavior in response to price signals or availability of renewable energy, making them well-suited for demand response programs and real-time grid balancing. The third and most valuable segment, “frequency-responsive users,” demonstrate exceptional responsiveness, capable of rapidly modulating their charging or even discharging back to the grid during critical moments. These users are essential for ancillary services such as frequency regulation, where milliseconds matter.
What sets this model apart is its ability to classify users accurately despite limited data inputs. Unlike deep learning models that require vast amounts of granular information—such as GPS trajectories, driving history, or battery state-of-charge—the proposed method relies on a handful of high-level indicators. These include the overlap between a user’s charging window and predefined market target periods, the controllability of the charging infrastructure, and the elasticity of charging duration and power draw. By focusing on these metrics, the model respects user privacy while still extracting meaningful behavioral insights. The researchers emphasize that this aligns with the growing emphasis on data minimization in smart grid applications, where collecting only what is necessary is not just a technical choice but an ethical imperative.
The core of the methodology lies in its dual-layer analytical engine. The first layer employs a cloud model to translate qualitative descriptions—such as “high overlap” or “low responsiveness”—into quantitative measures. Cloud models are particularly effective in handling uncertainty and ambiguity, common traits in human behavior. For instance, the concept of “frequent charging” is not a fixed number but a fuzzy range that varies between individuals. The cloud model captures this variability by defining each behavioral trait with three parameters: expectation (the central tendency), entropy (the degree of uncertainty), and hyper-entropy (the volatility of uncertainty). This allows the system to represent not just what users do, but how consistently and predictably they do it.
The second layer uses a fuzzy Petri net (FPN) to perform logical inference based on the outputs of the cloud model. FPNs are a type of graphical modeling tool that combines the state-transition logic of Petri nets with the uncertainty-handling capabilities of fuzzy logic. In this context, the FPN acts as a rule-based decision engine, where each node represents a behavioral condition (e.g., “high time overlap”) and each transition represents a logical rule (e.g., “if time overlap is high and power elasticity is low, then the user is energy-responsive”). The beauty of this structure is its interpretability—unlike black-box AI models, the reasoning process is transparent and can be audited or adjusted by domain experts. This is crucial for building trust among stakeholders, including regulators, utility operators, and end-users.
To validate their approach, the researchers conducted a case study using real-world data from 100 EV users at a charging facility in Beijing. The dataset included basic charging records—start and end times, energy delivered, and station power ratings—but no personal identifiers or location histories. Using the proposed model, the team successfully classified users into the three defined categories with high confidence. For example, User 1, who consistently charged between 5 PM and 8 PM with little variation in duration, was identified as an energy-responsive user with a confidence score of 0.95. User 2, who frequently adjusted charging times and power levels in response to external cues, was classified as power-responsive with a score of 0.897. User 4, who not only charged during peak hours but also demonstrated the ability to modulate power rapidly, was labeled frequency-responsive with a confidence of 0.941.
The implications of these findings are far-reaching. From a utility perspective, the ability to identify high-potential users means incentive budgets can be allocated more efficiently. Instead of offering blanket discounts to all EV owners, operators can target specific groups with tailored offers. Energy-responsive users might receive long-term subscription-based benefits, such as discounted overnight charging rates. Power-responsive users could be enrolled in dynamic pricing programs, where they earn credits for shifting consumption away from peak periods. Frequency-responsive users, being the most valuable for grid stability, could be offered premium rewards, such as free charging sessions or priority access to high-power chargers.
Moreover, the model enables a more sustainable business model for V2G services. By demonstrating that participation can be both profitable and non-intrusive, it may encourage more EV owners to opt in. The researchers suggest that future iterations could integrate feedback loops, where user responses to incentives are continuously monitored and used to refine the classification model. This would create a self-improving system that adapts to changing behaviors and market conditions.
Another significant advantage is scalability. Because the model does not rely on proprietary or sensitive data, it can be deployed across different regions and charging networks with minimal customization. This is particularly important in markets where data governance laws vary significantly. A utility in Germany, for instance, could apply the same framework as one in California, ensuring consistency in user engagement strategies while complying with local regulations.
The study also highlights the importance of infrastructure readiness. While the model can identify users with high response potential, actual participation depends on the availability of bidirectional chargers and supportive grid architecture. The authors note that current deployment of vehicle-to-grid-capable hardware remains limited, which constrains the practical impact of even the most sophisticated behavioral models. They call for coordinated investment in charging infrastructure, particularly in commercial and fleet depots where vehicles are parked for extended periods and can serve as distributed energy resources.
From a policy standpoint, the research underscores the need for regulatory frameworks that facilitate user participation. Current market rules in many jurisdictions impose strict thresholds on minimum power output, duration of availability, and response speed, effectively excluding individual EV owners from ancillary service markets. The authors advocate for the creation of aggregated bidding mechanisms, where multiple EVs are grouped into virtual power plants managed by third-party aggregators. This would lower the entry barrier for individual users while still delivering the grid services needed by system operators.
The human factor cannot be overlooked. Even with the best technology and policies in place, user adoption hinges on trust and perceived value. The researchers stress that transparency is key—users should understand how their data is being used, what benefits they stand to gain, and how their participation contributes to broader energy goals such as decarbonization and grid resilience. Incentive programs should be designed not just to reward behavior, but to educate and empower users, turning them from passive consumers into active participants in the energy transition.
Looking ahead, the model could be extended to incorporate additional behavioral dimensions, such as user preferences for renewable energy, willingness to accept delayed charging, or sensitivity to battery health concerns. Integration with smart home systems could further enhance personalization, allowing the model to consider factors like household electricity consumption or solar generation patterns. The ultimate vision is a seamless, user-centric energy ecosystem where EVs are not just transportation devices but integral components of a flexible, responsive, and sustainable power grid.
In conclusion, the work presented by Shi Tianchen, Yang Ye, Liu Mingguang, Wang Wen, Wang Jiani, and Liu Dunnan offers a practical and privacy-preserving solution to one of the most pressing challenges in the EV revolution. By moving away from data-intensive AI models and embracing a more nuanced, rule-based approach, they have paved the way for smarter, fairer, and more inclusive grid integration. As the world accelerates toward electrified transportation, tools like this will be essential for ensuring that the transition is not only rapid but also resilient and equitable.
Shi Tianchen, Yang Ye, Liu Mingguang, Wang Wen, Wang Jiani, Liu Dunnan, North China Electric Power University and State Grid Smart Internet of Vehicles Co., Ltd., Electric Power Construction, DOI: 10.12204/j.issn.1000-7229.2024.10.007