Abstract
Federated Learning (FL) is widely regarded as a "win-win" solution for improving machine learning model accuracy while preserving data privacy. However, this study reveals a hidden economic consequence in competitive markets: FL can lead to decreased profits. By improving the Quality of Service (QoS) for all competitors, FL levels the playing field, intensifies price competition, and potentially triggers a "race to the bottom." Using a game-theoretic approach applied to Electric Vehicle (EV) charging stations, we demonstrate that the technical benefits of FL do not always translate into economic gains.
Introduction
Electric Vehicle (EV) charging stations compete on two primary dimensions: price and Quality of Service (QoS). High QoS, such as accurate wait-time predictions, requires robust machine learning models. However, individual stations often face "data scarcity," limiting their ability to build high-quality models independently. While Federated Learning (FL) offers a technical solution to this data scarcity without compromising privacy, its economic impact on market competition remains under-explored. This paper investigates whether the boost in service quality from FL is sufficient to offset the intensified price wars that occur when competitors become less differentiated.
Methodology: A Three-Stage Game
The interaction between two competing charging stations (a duopoly) and EV owners is modeled as a three-stage game, solved via backward induction.
- Stage I (FL Participation): Stations decide whether to join the FL collaboration.
- Stage II (Pricing): Stations set charging prices based on their respective QoS.
- Stage III (User Selection): EV owners choose a station based on a Hotelling model, which balances charging utility (QoS), price, and travel costs.
The resulting optimization problem is non-concave, which the authors address by decomposing it into a piece-wise concave program. This allows for a full characterization of the equilibrium across various market scenarios, including monopoly and bifurcated markets.
Experimental Setup
The study utilizes a real-world dataset of 20,000 EV charging sessions from Dundee, focusing on demand prediction (kWh consumption). To simulate realistic market conditions, a Dirichlet distribution was used to control data heterogeneity (non-IIDness) between stations. Model performance was evaluated using Root Mean Squared Error (RMSE).
Key Insights: The Similarity Trap
The core finding is that FL acts as a double-edged sword. While it consistently reduces RMSE (improving accuracy), it also reduces the differentiation between stations. When two stations offer nearly identical high-quality services, they are forced to compete primarily on price, which erodes profit margins.
A station's optimal price is fundamentally tied to the difference between its QoS and its competitor's QoS. By shrinking this gap, FL can lead to lower profits even if the collaboration itself is cost-free. Stations are most likely to participate in FL when their data is "mildly dissimilar"—enough to gain accuracy but not so much that price competition becomes ruinous.
Results and Conclusion
Numerical results confirm that across all tested levels of data heterogeneity ($\beta$), FL improved technical accuracy. However, economic analysis reveals that for specific data distributions, both stations' profits were lower with FL than without it. Stations naturally opt out of FL when data is either highly skewed or nearly identical to protect their market position.
This research highlights that the decision to adopt Federated Learning must go beyond technical metrics. In a competitive landscape, the strategic impact on market dynamics and pricing power is just as critical as the improvement in model performance. Future work will extend this analysis to multi-provider scenarios and more complex charging service types.
Summary updated