Federated Learning Boosts EV Charging Forecast Accuracy
As the global push for sustainable transportation accelerates, electric vehicles (EVs) are no longer a niche market but a central pillar of future mobility. With millions of EVs hitting the roads annually, the strain on power grids from charging demand has become a critical challenge for utilities and urban planners alike. Accurately predicting short-term EV charging load is essential to ensure grid stability, optimize energy dispatch, and support the expansion of charging infrastructure. However, traditional forecasting models have long struggled with a dual dilemma: they require detailed user behavior data to improve accuracy, yet such data is highly sensitive and cannot be freely shared due to privacy concerns.
A groundbreaking study published in High Voltage Engineering introduces a novel solution that bridges this gap—by leveraging federated learning (FL), a privacy-preserving artificial intelligence framework, to enhance the precision of EV charging load forecasts without compromising user data security.
Led by Professor Yang Ting and his team at the School of Electrical and Information Engineering, Tianjin University, the research presents a comprehensive approach that integrates real-world user charging behaviors—such as start and end times, battery state of charge (SOC), battery capacity, and selected charging power—into a machine learning model, all while ensuring that personal and operational data remains localized and protected.
The significance of this work lies in its practical response to one of the most pressing issues in smart grid development: how to harness valuable behavioral data for predictive analytics without violating user privacy. As EV adoption continues to surge, especially in densely populated urban centers, the ability to forecast charging demand with high accuracy becomes not just a technical goal, but a necessity for maintaining reliable electricity supply.
Current methods for EV load forecasting fall into two broad categories. The first relies on historical load data and machine learning algorithms such as XGBoost, random forests, or LSTM networks. While these models can capture temporal patterns, they often lack granular behavioral context, limiting their predictive power. The second category involves probabilistic or travel-chain-based models that simulate user behavior using assumptions about driving patterns, trip durations, and charging preferences. Although theoretically sound, these models are difficult to scale and calibrate in real-world settings, particularly when user behavior is highly variable.
What sets the Tianjin University team’s approach apart is its integration of both data-driven learning and behavioral realism within a privacy-conscious architecture. Instead of collecting raw user data in a central server—a practice fraught with cybersecurity and regulatory risks—the researchers implemented a horizontal federated learning framework. In this setup, each electric vehicle charging operator (EVCO) trains a local model using its own dataset. Only the model parameters—such as weights and biases—are shared with a central aggregator (e.g., a grid company), which then fuses them into a global model and redistributes it for further refinement.
This decentralized training mechanism ensures that sensitive information, such as individual charging sessions or user identities, never leaves the operator’s secure environment. At the same time, the collaborative nature of federated learning allows the global model to learn from diverse datasets across multiple regions and user demographics, significantly enriching its generalization capability.
The core of the proposed model is a bidirectional long short-term memory (BiLSTM) neural network, chosen for its ability to capture both past and future dependencies in time series data. Unlike standard LSTM models that process sequences in a single direction, BiLSTM analyzes data forward and backward, enabling it to detect complex temporal patterns in charging behavior—such as peak usage hours, charging duration trends, and SOC recovery curves—with greater sensitivity.
To prepare the input data, the team constructed a composite feature set combining historical load measurements with behavioral indicators. For instance, the total required charging energy at any given time was calculated based on the number of active users, their current SOC, and battery capacities. Similarly, average charging power and the SOC at which users typically end their sessions were extracted to reflect real-world usage patterns. These features were then fed into the BiLSTM model through a rolling sliding window strategy, allowing the system to continuously update its predictions based on the most recent data.
One of the key innovations in the study is the use of model parameters—not gradients—as the communication payload between local and central systems. While conventional federated learning often transmits gradient updates, which can sometimes be reverse-engineered to infer original data, transmitting final weights and biases adds an additional layer of obfuscation. Even if intercepted, these parameters are far less likely to reveal identifiable user information, thereby enhancing the overall security of the system.
The experimental validation was conducted using real-world charging data from nine EVCOs in a major Chinese city, covering a full month of operations in December 2022. The dataset included 25,920 sampling points, capturing both load dynamics and detailed user behavior. Each operator’s data was split into training, validation, and test sets in a 4:1:1 ratio, and all inputs were normalized to ensure consistency across sources.
Performance was evaluated using two standard metrics: mean absolute percentage error (MAPE) and root mean square error (RMSE). The results were compelling. When user behavior data was incorporated into the model, both RMSE and MAPE showed significant reductions across all operators. For example, one participant saw a 5.57% drop in RMSE and a 0.55% improvement in MAPE, demonstrating that behavioral features meaningfully enhance predictive accuracy.
Even more striking was the comparison between standalone local models and the federated learning approach. In every case, the federated model outperformed individual local models, with RMSE reductions ranging from 9.34% to 11.07% and MAPE improvements between 2.13% and 5.29%. This consistent uplift underscores the value of collaborative learning: by pooling insights from multiple operators, the global model gains exposure to a broader spectrum of charging behaviors, leading to more robust and accurate predictions.
Perhaps the most impressive demonstration of the model’s strength came in a generalization test involving a new, previously unseen EVCO—referred to as C10. This operator had limited historical data, having only recorded 24 hours of charging activity. When trained locally, its model performed poorly, yielding an RMSE of 65.3 kW and a MAPE of 19.2%—unacceptably high for operational planning. However, when the pre-trained federated model was applied to C10’s data, the RMSE plummeted to 29.4 kW and the MAPE dropped to 6.8%, representing a 54.9% and 12.4% improvement, respectively.
This result highlights a crucial advantage of federated learning: it enables rapid deployment and accurate forecasting even for new or data-scarce operators. In practical terms, this means that emerging charging networks can immediately benefit from the collective intelligence of established players, accelerating their integration into the broader energy ecosystem without the need for years of data accumulation.
From a computational standpoint, the model was optimized using the Adam algorithm, an adaptive learning rate method that adjusts parameter updates based on historical gradients. This not only speeds up convergence but also helps avoid local minima, ensuring stable and efficient training. The entire federated learning process was configured with 100 global communication rounds, 5 local epochs per round, and a batch size of 128, striking a balance between model accuracy and computational efficiency.
The implications of this research extend beyond technical performance. As governments and utilities grapple with the integration of distributed energy resources, including EVs, solar panels, and home storage systems, the ability to forecast demand with high fidelity becomes a cornerstone of grid resilience. Traditional forecasting tools, often built on aggregated and anonymized data, are increasingly inadequate in the face of dynamic, decentralized energy consumption patterns.
The federated learning framework proposed by Yang Ting and his colleagues offers a scalable, secure, and intelligent alternative. It aligns with modern data governance principles, such as those outlined in the EU’s General Data Protection Regulation (GDPR) and China’s Personal Information Protection Law (PIPL), by minimizing data exposure and decentralizing control. At the same time, it empowers operators and grid managers with more accurate, actionable insights.
Moreover, the model’s architecture is inherently extensible. Future iterations could incorporate additional data sources, such as weather conditions, electricity pricing signals, or traffic congestion levels, to further refine predictions. The same framework could also be adapted for other applications, such as demand response management, vehicle-to-grid (V2G) coordination, or renewable energy forecasting.
For charging operators, the benefits are clear: improved load forecasting enables better capacity planning, reduces peak demand charges, and enhances customer satisfaction by minimizing wait times and service disruptions. For grid operators, accurate forecasts support more efficient unit commitment, frequency regulation, and voltage control—critical functions in maintaining power quality and reliability.
The study also opens new avenues for collaboration between public and private stakeholders. By establishing a federated learning consortium among EVCOs, utilities, and research institutions, cities can build shared forecasting platforms that serve the common good without compromising competitive advantage or user privacy.
In an era where data is often described as the new oil, this research reminds us that value does not always come from centralization. Sometimes, the most powerful insights emerge not from hoarding data, but from learning together—safely, ethically, and intelligently.
As the world transitions toward electrified transportation, the challenges of grid integration will only grow more complex. But with innovative approaches like federated learning, the path forward is becoming clearer. By respecting user privacy while unlocking the predictive power of behavioral data, this work sets a new standard for smart grid analytics—one that balances technological ambition with ethical responsibility.
The success of this project also reflects the growing maturity of AI applications in energy systems. No longer confined to theoretical simulations or controlled lab environments, machine learning models are now being deployed in real-world, multi-stakeholder settings where data privacy, model robustness, and operational feasibility are paramount. The fact that this model was tested on data from multiple independent operators speaks to its readiness for real-world implementation.
Looking ahead, the research team plans to explore vertical federated learning scenarios, where different types of data—such as user demographics, vehicle types, and charging station configurations—may be held by different entities and require more complex alignment mechanisms. They are also investigating the use of differential privacy techniques to further strengthen data protection, adding mathematical guarantees against re-identification attacks.
In conclusion, the work by Yang Ting, Qin Xiaobing, Feng Xiangwei, and Xu Zheming represents a significant leap forward in the field of EV load forecasting. It demonstrates that high accuracy and strong privacy protection are not mutually exclusive goals, but complementary objectives that can be achieved through thoughtful system design. As electric mobility continues to evolve, such innovations will be essential to building a sustainable, resilient, and user-centric energy future.
Yang Ting, Qin Xiaobing, Feng Xiangwei, Xu Zheming, School of Electrical and Information Engineering, Tianjin University. High Voltage Engineering, DOI: 10.13336/j.1003-6520.hve.20230765