Trusted AI Framework Enhances Vehicle-to-Grid Scheduling
A groundbreaking advancement in smart grid technology has emerged from a collaborative research effort aimed at improving the reliability and efficiency of vehicle-to-grid (V2G) systems. As electric vehicles (EVs) become increasingly integrated into power networks, the challenge of coordinating their charging and discharging activities without compromising data privacy or decision accuracy has grown more complex. In response, a team of researchers has introduced a novel trustworthy federated learning method specifically designed for V2G scheduling, offering a robust solution that aligns the interests of EV users, charging station operators, and grid operators while ensuring secure and verifiable model training.
The study, led by Yunhua He and Yuhang Cheng from the School of Information Science and Technology at North China University of Technology, in collaboration with Xiaodong Yuan and Yajuan Guo from the Electric Power Research Institute of State Grid Jiangsu Electric Power Co., Ltd., and Jianlin Li from the National Use-Side Energy Storage Innovation Research and Development Center, presents a comprehensive architecture that addresses two critical challenges in current V2G implementations: the subjectivity of training data labels and the lack of verification in model parameter aggregation. Their findings were published in the December 2024 issue of Distributed Energy, a peer-reviewed journal recognized for its contributions to energy systems innovation.
At the heart of modern V2G operations lies the need for intelligent scheduling models capable of guiding thousands of EVs to charge during off-peak hours and discharge during periods of high demand. This process, known as peak shaving and valley filling, helps stabilize the electrical grid, reduces strain on infrastructure, and lowers overall energy costs. However, training such models traditionally requires access to vast amounts of sensitive data—ranging from individual driving patterns to real-time electricity pricing—which many stakeholders are reluctant to share due to privacy and competitive concerns.
Federated learning has been proposed as a promising approach to overcome this data-sharing barrier. Unlike conventional machine learning, where data is centralized, federated learning allows multiple parties to collaboratively train a model without exposing their raw data. Each participant trains a local model using their own dataset and shares only the updated model parameters with a central aggregator. These parameters are then combined to form a global model, which is redistributed for further refinement. While this method enhances privacy, it introduces new vulnerabilities. The quality of the final model heavily depends on the accuracy and objectivity of the training labels used, which are often determined subjectively by the model developer. Moreover, the central aggregator, typically a load aggregator in V2G contexts, could potentially manipulate the aggregation process, leading to incorrect or biased scheduling decisions.
Recognizing these limitations, the research team developed a three-part architecture designed to ensure both the integrity of the training process and the trustworthiness of the outcomes. The first component is a label generation module that creates objective, optimal training labels by considering the interests of all key stakeholders. Instead of relying on arbitrary assumptions, the model calculates labels based on a multi-objective optimization framework that simultaneously maximizes operator revenue, minimizes EV user charging costs, and reduces grid-side load fluctuations.
This approach marks a significant departure from previous methods, which often prioritized a single objective—such as minimizing grid load variance—without adequately accounting for economic incentives for users and service providers. By integrating these diverse goals into a unified mathematical model, the researchers ensure that the resulting scheduling decisions are not only technically sound but also economically viable and socially acceptable. For instance, when a large number of “early-out, late-back” EVs dominate the fleet, the model automatically adjusts pricing strategies to reflect increased nighttime demand, ensuring that operators can maximize profits while still offering competitive rates to users.
The second component of the architecture is a verifiable federated learning module that employs homomorphic encryption to protect the privacy of local model updates. Homomorphic encryption allows computations to be performed directly on encrypted data, meaning that the load aggregator can sum up the encrypted model parameters from each charging station without ever decrypting them. This eliminates the risk of the aggregator reverse-engineering sensitive information about individual stations’ operations or user behavior, a vulnerability present in many existing systems.
However, encryption alone does not guarantee correctness. A malicious or malfunctioning aggregator could still submit an incorrect aggregated result, either intentionally or due to computational errors. To address this, the researchers introduced a third component: a secure record and verification mechanism built on blockchain technology and smart contracts. After the aggregator performs the encrypted parameter summation, it constructs a binary tree structure—referred to as an aggregation tree—where each leaf node represents a charging station’s encrypted update, and internal nodes represent partial sums. This tree is stored off-chain using the InterPlanetary File System (IPFS) to minimize blockchain storage costs, while a reference pointer is recorded on-chain.
Each participating charging station can then independently verify the correctness of the aggregation by executing a lightweight smart contract that checks the mathematical consistency of the tree. Because the verification follows a hierarchical path from the root to the leaves, the number of computational steps required grows logarithmically with the number of participants, rather than linearly. This design dramatically improves efficiency, especially in large-scale deployments involving hundreds or even thousands of charging stations.
One of the most compelling aspects of the proposed method is its ability to detect and locate errors in the aggregation process. If a discrepancy is found during verification, the smart contract can pinpoint the exact level and position within the aggregation tree where the inconsistency occurred. This error tracing capability is crucial for maintaining system integrity, as it enables operators to identify and rectify faulty components without having to re-run the entire training process.
To evaluate the performance of their framework, the researchers conducted a series of simulations involving 100,000 EVs categorized into three distinct usage patterns: “early-out, late-back,” “regular routine,” and “night-shift.” These categories reflect common real-world driving behaviors, such as commuters who leave home early in the morning and return late at night, office workers with standard 9-to-5 schedules, and individuals working overnight shifts. By varying the proportion of each type, the team analyzed how different fleet compositions affect operator revenue, user costs, and grid stability.
The results revealed that the dominant EV type significantly influences economic outcomes. When “early-out, late-back” EVs were in the majority, the minimum charging cost per vehicle was 1.2 yuan, slightly higher than the 1.152 yuan observed when “regular routine” EVs dominated. This difference stems from the concentrated nighttime charging demand associated with early commuters, which drives up electricity prices during off-peak hours. In contrast, when “night-shift” EVs—whose charging windows overlap with peak daytime loads—were predominant, the minimum charging cost rose to 1.584 yuan, reflecting the higher market prices during periods of high demand.
Despite these variations in cost and revenue, the grid-side load variance remained consistently low across all scenarios. This outcome underscores a key strength of the proposed label generation model: its ability to maintain grid stability regardless of fleet composition. By optimizing charging schedules to minimize load fluctuations, the system ensures that the integration of EVs contributes positively to grid management rather than exacerbating existing challenges.
Beyond the economic and operational insights, the study also provides critical technical benchmarks for scalability and efficiency. The researchers measured the time and storage overhead associated with constructing the aggregation tree under different network sizes, ranging from 8 to 256 charging stations. They found that tree construction time increased from less than 1 millisecond for 8 stations to approximately 23 milliseconds for 256 stations—a manageable latency for real-time applications. Storage requirements, while growing exponentially due to the binary tree structure, remained within practical limits thanks to the use of IPFS for off-chain data storage.
Perhaps the most striking finding relates to the efficiency of the verification process. When compared to traditional methods that require checking every individual parameter update, the proposed smart contract-based verification reduced computational costs—measured in Ethereum’s “Gas” units—by more than 96% in large networks. For example, with 256 charging stations, the Gas savings exceeded 96%, making the approach highly suitable for deployment on public blockchain platforms where transaction fees can be prohibitive.
The implications of this research extend far beyond academic interest. As utilities and governments worldwide push for greater electrification of transportation, the ability to manage millions of EVs as distributed energy resources will become essential. Current pilot programs often rely on centralized control systems that lack transparency and are vulnerable to single points of failure. The trustworthy federated learning framework proposed by He, Cheng, Yuan, Guo, and Li offers a decentralized alternative that enhances security, promotes fairness, and fosters stakeholder trust.
Moreover, the integration of blockchain and smart contracts introduces a new level of auditability and accountability into the V2G ecosystem. Every aggregation event is immutably recorded, and any deviation from the expected protocol can be automatically detected and flagged. This transparency is particularly valuable in regulatory environments where compliance and data integrity are paramount.
The research also highlights the importance of interdisciplinary collaboration in addressing complex energy challenges. By combining expertise in machine learning, cryptography, power systems engineering, and blockchain technology, the team has created a solution that is greater than the sum of its parts. Their work exemplifies how cutting-edge computer science techniques can be applied to real-world infrastructure problems, paving the way for smarter, more resilient energy systems.
Looking ahead, the framework could be extended to incorporate additional variables such as renewable energy generation, dynamic pricing signals, and user preferences. Future iterations might also explore hybrid models that combine federated learning with edge computing to further reduce latency and improve responsiveness. As 5G and beyond networks roll out, enabling faster and more reliable communication between EVs and charging infrastructure, the potential for real-time, adaptive V2G coordination will only grow.
In conclusion, the study represents a significant leap forward in the development of trustworthy, scalable, and efficient V2G systems. By addressing the dual challenges of data privacy and model integrity, the researchers have laid the foundation for a new generation of intelligent grid management tools. Their work not only advances the state of the art in federated learning applications but also demonstrates the transformative potential of integrating artificial intelligence with secure, decentralized technologies in the energy sector.
Yunhua He, Yuhang Cheng, Xiaodong Yuan, Yajuan Guo, Jianlin Li, North China University of Technology and State Grid Jiangsu Electric Power, Distributed Energy, DOI: 10.16513/j.2096-2185.DE.2409608