🤖 machine learning

Federated Learning for Early Prediction of EV Charging Demand

This article proposes a framework for federated learning that enables accurate, privacy-compliant early prediction of electric vehicle charging demand with minimal session data while achieving performance comparable to centralized models, all while keeping data localized at the charging stations.

Original authors: Vasilis Perifanis, Foteini Nikolaidou, Nikolaos Pavlidis, Panagiotis Thomakos, Andreas Sendros

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vasilis Perifanis, Foteini Nikolaidou, Nikolaos Pavlidis, Panagiotis Thomakos, Andreas Sendros

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Guessing the Future Before the Movie Ends

Imagine you walk into a cinema and take a seat. You haven't seen the movie yet, but the cinema manager wants to know: "How much popcorn and soda will this person buy by the time the credits roll?"

Normally, the manager must wait until the movie ends to know the answer. But in this paper, the researchers try to solve a different problem: Can they guess the total amount of energy an electric vehicle (EV) needs by looking only at the first 10 minutes of the charging process?

This is crucial for charging station operators. If they know early on that a car needs a massive amount of energy, they can immediately adjust the energy flow to keep the entire power grid stable. If they wait until the car is almost full, it is too late to make changes.

The Challenge: The "Privacy" Wall

To teach a computer to make these predictions, you normally need a lot of data. In the past, companies collected all charging data from every station, stored it in a huge central database, and trained a model there.

But there is a problem: Privacy and logistics.

Privacy: Station operators do not always want to send raw data about their customers to a central cloud.
Logistics: Moving terabytes of data from hundreds of stations to one location is slow and expensive.

The Solution: The "Secret Recipe" Exchange (Federated Learning)

The authors propose a clever solution called Federated Learning.

Imagine a group of chefs, each working in their own separate kitchen (a charging station). All of them want to learn how to prepare the perfect "Energy Prediction Soup."

The old way: Each chef sends their ingredients and cooking logs to a central headquarters. The head chef mixes everything together to create a single giant recipe.
The new way (Federated Learning): The head chef sends a "base recipe" to every kitchen. Each chef cooks a little using their own local ingredients. Instead of sending their ingredients back, they send only the changes they made to the recipe (e.g., "I added a pinch more salt"). The head chef collects these changes, updates the main recipe, and sends the new version back.

In this paper, the "chefs" are the individual charging stations, and the "recipe" is the AI model that predicts energy demand. The raw data never leaves the station; only the learned "insights" travel.

What They Did

The Data: They used real data from the Caltech charging park (54 different charging stations). They examined sessions where they had only information from the moment of connection and the first 10 minutes of the charging process.
The Clues: They taught the AI to look for specific clues available in those first 10 minutes:
- Time: Is it a weekend? What time is it?
- User Clues: Did the driver say, "I need to leave in 2 hours" or "I need 50 miles of range"?
- Early Behavior: How fast is the car charging right now? Is the power rising or falling?
The Test: They trained many different types of AI models (from simple mathematical formulas to complex "neural networks" that think like brains) to predict the total energy required. They compared two scenarios:
- Centralized: All data in one place.
- Federated: Data stays at the stations; only model updates are shared.

The Results: The Magic of Collaboration

Here is what they found, translated into simple language:

Early guessing works: Even with only the first 10 minutes of data, the AI could predict the total energy demand with surprising accuracy. It is like guessing the total length of a movie by watching only the first 10 minutes.
The "Secret Recipe" works: The Federated Learning models (the chefs exchanging recipes) performed almost as well as the centralized models (the giant database).
- The best "centralized" model (a complex AI called a Transformer) had a low error rate.
- The best "federated" model (the same Transformer) had a slightly higher error rate, but was close enough to be useful in the real world.
Not all models are equal: Some simple models (like basic mathematical formulas) struggled when data was distributed across stations. However, the more complex AI models (like the Transformer and MLP) were very good at handling the differences between stations and learned effectively without seeing everyone's data.
Efficiency: The complex models were somewhat heavier (requiring more computer memory), but the simpler models were very light. This means operators can choose a model that fits their hardware, whether it is a powerful server or a small device directly at the charger.

The Conclusion

The paper proves that you do not need to steal the private data of everyone to build an intelligent system for electric vehicles. By using Federated Learning, charging stations can collaborate to build an intelligent "crystal ball" that predicts energy demand early in the charging process. This keeps data private, saves bandwidth, and still provides operators with the accurate information they need to manage the grid efficiently.

In short: They showed that a group of strangers can learn to predict the future together without ever showing each other their private notebooks.

Technical Summary: Federated Learning for Early Prediction of Electric Vehicle Charging Power

Problem Definition
The work addresses the challenge of predicting the total energy demand of an electric vehicle (EV) charging session as early as possible. In depot-based environments, operational decisions regarding load distribution, prioritization, and power allocation often need to be made shortly after vehicle plugging-in, before the session concludes. Traditional prediction methods frequently rely on data only available after session completion or require centralized data aggregation, raising privacy and governance concerns. The specific problem formulated here is a supervised regression task: given the plug-in context and measurements from an initial observation window (the first 10 minutes of charging), predict the total energy delivered by the session ( $E_{tot}$ ).

Methodology
The authors construct a session-level dataset using the Adaptive Charging Network (ACN) framework, focusing on a single operational depot (Caltech) with 54 charging stations.

Feature Engineering: The input vector combines static session metadata (e.g., connection timestamp, user-requested energy, departure time) with dynamic features extracted from the first 10 minutes of charging. These dynamic features include summary statistics (mean, maximum, minimum, standard deviation, last value) as well as linear trend coefficients for charging current and pilot signal. Additionally, interaction features such as the utilization ratio (current vs. pilot capacity) and an approximate energy estimate for the early window (via trapezoidal integration) are derived.
Federated Learning (FL) Setup: The study employs an intra-depot FL setup where each charging station acts as an independent client. Raw data remain local at each station; only model updates are exchanged with a central server using the FedAvg algorithm. This setup models a realistic scenario where data are distributed across stations within a single depot.
Data Heterogeneity: The authors analyze statistical heterogeneity across the 54 station clients using Jensen–Shannon (JS) divergence. They characterize the data distribution as "slightly non-IID," with most stations exhibiting moderate divergence, although some stations with small sample sizes act as outliers.
Model Families: The study evaluates a diverse range of models under both centralized and federated conditions:
- Baselines: Mean-only and Gaussian dummy predictors.
- Machine Learning: Linear Regression (LR) and XGBoost (XGB).
- Neural Networks: Multi-Layer Perceptron (MLP), 1D-Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU), and a Tabular Transformer.

Main Contributions

Problem Formulation: The work defines an early-session regression problem where total delivered energy is predicted exclusively using plug-in context and behavior from the first 10 minutes of the session, thereby avoiding information leakage from later phases.
Dataset Construction: A realistic session-level dataset was created by integrating ACN metadata with early-window charging dynamics, resulting in a compact tabular representation suitable for edge deployment.
Federated Evaluation: The work compares centralized training against station-partitioned FL across multiple model families. It demonstrates that privacy-enhancing distributed learning can achieve performance levels comparable to centralized training while preserving data locality.

Results

Centralized Performance: In the centralized setting, XGBoost achieved the highest accuracy (MAE: 3.36 kWh, RMSE: 4.84 kWh). Among neural networks, the Transformer performed best (MAE: 3.55 kWh), significantly outperforming linear models (LR MAE: 4.33 kWh), indicating that the relationship between early-session descriptors and total energy is non-linear.
Federated Performance: Federated models generally converged to performance levels close to their centralized counterparts.
- Neural Models: The Transformer (MAE: 3.69 kWh) and MLP (MAE: 3.79 kWh) showed the least degradation compared to centralized training. They also converged relatively quickly (approximately 160–223 communication rounds).
- Tree/Linear Models: XGBoost and Linear Regression performed worse in the federated setting than their centralized versions and required more communication rounds to converge (XGB had not fully converged within 400 rounds).
- Stability: Validation and test curves remained close together, indicating stable optimization without significant overfitting across rounds.
Resource Efficiency: A trade-off between accuracy and resource usage was observed. While the Transformer offered the best accuracy, it had the largest model size (2.33 MB) and highest memory requirements. MLP and GRU provided a balanced compromise with competitive accuracy and a smaller footprint, making them suitable for resource-constrained edge hardware.

Significance and Claims
The work claims that reliable estimates of EV charging demand can be obtained early in a session using minimal data (plug-in context + 10 minutes of behavior). It argues that Federated Learning offers a practical pathway for scalable, privacy-conscious analytics in EV charging networks. The study concludes that while centralized training delivers the highest theoretical performance, FL enables distributed infrastructures to achieve comparable prediction accuracy without transmitting raw data, thereby supporting privacy-compliant operations. The authors suggest this approach is particularly suitable for edge-oriented deployments where lightweight models (such as MLP or GRU) can balance prediction quality with deployability. Future work is identified as extending to inter-depot settings and integrating complementary signals such as photovoltaic production or local storage forecasts.