Federated Learning for Electric Vehicle Routing and Charging Station Location

Federated Learning for Electric Vehicle Routing and Charging Station Location

K P Suprith Patil

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology Bengaluru, India, suprithpatil@ieee.org

Siddarth T

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology Bengaluru, India

Siddarthsiddu373@gmail.com

Phalani Kumar

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology Bengaluru, India, Phalanikumar09@gmail.com

Prof. Anu D

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology anu.d@nmit.ac.in

S Tejaswi Naidu

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology Bengaluru, India, tejasompalli07@gmail.com

Pavan saiteja

Dept. of Artificial Intelligence and Data Science Nitte Meenakshi Institute of Technology Bengaluru, India, pavansaiteja29@gmail.com

AbstractThe Electric Vehicle Routing and Charging Station Location (EVRCSL) problem is pivotal for smart mobility, but the use of conventional centralized optimization techniques jeopardizes gross privacy and scalability challenges through the need for exposure to sensitive user data. This work introduces a new FL-based solution to address these problems. Our approach trains two concurrent prediction models, one for route cost and another for charging station utility, on distributed data sources without the need for central aggregation. The models are incorporated into a recommendation engine ranking charging options by a composite score. Experimental outcomes from a simulated setup reveal that our FL-based method attains prediction accuracy on par with centralized baselines while ensuring user data privacy. This paper illustrates the feasibility of FL as an underlying technology for developing scalable, privacy- preserving, and efficient intelligent transportation systems and sets the stage for more advanced real-time applications.

Index TermsFederated Learning, EV Routing, Smart Mobil-

ity, Privacy, Intelligent Systems, Edge Computing

1. Introduction

   Shifting towards electric vehicles (EVs) globally is one of the main strategies in combating climate change and clear- ing city smogs. Though their environmental implications are obvious, widespread adoption of EVs is plagued with severe operational disadvantages. Among them are a scarcity of and frequently nonexistent charging station facilities, variable and extended charging times, and limited driving rangeall of which pose logistically challenging issues to fleet managers and driver range anxiety [1]. The Electric Vehicle Routing and Charging Station Location (EVRCSL) problem neces- sitates sophisticated optimization methods beyond minimum pathfinding [16]. Our suggested system design for this problem

   mathematical programming methods such as Mixed-Integer Linear Programming (MILP) can yield optimal solutions for small cases [6]. Such conventional solutions, however, have an essential weakness: they rely on access to a central database of individual data such as journey history, car performance data, and location data. This centralized approach has two inherent problems: (1) significant privacy concerns due to mixing individual data, and (2) scalability issues in large-scale field deployments with millions of cars and charging spots generating data. tigating In order to bypass these limitations, researchers are inves- decentralized, privacy-guaranteeing solutions. Feder- ated Learning (FL) has been a potential technique for joint model training across distributed devices without raw data sharing [9]. Previous work has realized FLs potential in usage such as personalized station recommendations [12], yet no systemic framework has been established employing FL for the joint training of both charging station value prediction models and time-varying route costs. This limitation drives our research, as dual modeling is necessary for end-to-end informed suggestions. The main contributions of this paper are :

   1. A new dual-model FL architecture learning route cost and charging station utility predictors in parallel, build- ing a strong recommendation base

   2. Simulation testing demonstrating our system operates with near-centralized performance without sacrificing privacy

   3. A scalable, secure architecture explicitly addressing the scalability and privacy limitations of centralized EV routing systems The remainder of the paper is structured as follows. Literature on the issues of EV routing and federated learning is reviewed in Section II. Section III illustrates our suggested methodology and system architecture in detail. Section IV

   presents and discusses experimental findings. Finally, Sec- tion V concludes the paper and outlines future research di- rections.

   Fig. 1. Proposed system architecture for EV routing and charging station location. The figure shows the whole pipeline from data acquisition up to federated model training and final recommendations.

2. Literature Survey

   The EVRCSL problem has been addressed across several optimization paradigms, largely driven by its NP-hard nature that makes the computation of exact solutions intractable. Metaheuristics are the most prevalent class of techniques, as they aim to discover good, near-optimal solutions within the time constraints of real-world applications. For instance, Genetic Algorithms (GA) have proven effective in navigating the inherently high-dimensional and combinatorial solution space, optimizing simultaneously for multiple objectives such as minimizing energy consumption and travel time [2]. Despite their strong exploration capabilities, GAs are often hindered by slow convergence and a reliance on centralized data sources to evaluate solution fitness, which assumes a global system view.Similarly, Ant Colony Optimization (ACO) has demon- strated excellent performance in dynamic routing scenarios. The algorithms emergent behavior is particularly well-suited to pathfinding in environments where conditions change over time [3], [24]. However, the major drawback of ACO lies in the computational expense associated with pheromone trail updates, especially in large-scale, densely connected road networks. Other metaheuristics such as Simulated Annealing (SA) [4] and Bat Optimization (BO) [5] have also been employed. These methods are capable of escaping local optima effectively, but they typically require finely tuned, problem- specific parameters, which can be challenging to generalize across different scenarios. To mitigate individual limitations,

   hybrid approaches have been proposed that combine the local search strengths of ACO with the global search capacity of GA. These hybrid methods aim to accelerate convergence and enhance overall solution quality [7], [8],

   [17].

   Mathematical programming provides an avenue to provable optimal solutions as a heuristic approach to the methods stated previously. The Mixed-Integer Linear Programming (MILP) paradigm has been used to formulate the EVRCSL problem and gives optimally guaranteed solutions to limited sized problems [6]. Subsequent development in this area have been addressed towards having more advanced and realistic constraints, like grid aware coordination for power load management [14], and human centered variables such as passenger satisfaction to encourage and maximize user take up [25]. Nonetheless, while mthematically sound as MILP, its primary limitation is its computational complexity is exponential, making it not viable in the large scale, real- time nature of modern transport systems. All these centred approaches, whether mathematical or metaheuristic, share they have a common limitation: a central controller with full access to all the relevant data. This means they assumed access to information like current vehicle locations, battery charge, driver behavior and charging station queue positions. Not only is this assumption a major scalability problem, but significant privacy problems, making these models impractical.

   To deal with these major scalability and privacy issues, the field of research has pivoted towards decentralized learning systems. A leading decentralized learning system is Federated Learning (FL), which allows several parties to collaboratively build machine learning models on distributed data sources without sharing user raw data [9], [10]. As an emerging area, extending FL to EV use is gaining more research interest [11].For example, Li and Wang [12] introduced a federated meta-reinforcement learning approach to real-time personalized charging station recommendation. This is an ex- ample of FL capability for flexibility and privacy guaranteeing and realistic complexity/communication cost bounds. Besides, to deal with data heterogeneity characteristic of distributed systems, Vertuzzo et al. [18] examined context-aware models in enabling FL personalization and performance on non-IID data.

   Although these works confirm the possibility of FL, they tend to solve only one aspect of the EVRCSL problem, viz., station recommendation in isolation. A complete solution demands a more holistic approach that weighs how desirable a charging point is (its inherent usefulness) against travel cost to it. Consequently, there is a substantial research gap for the creation of one, dual-model FL framework that can learn predictive models jointly and in private for dynamic route prices and static charging point value. Our contribution seeks to close this crucial gap by introducing and investigating exactly such a system, providing key to an entirely scalable and privacy-conscious smart transport solution.

   In addition to that, an entire EVRCSL model needs to account for something other than distance. Dynamic and

   stochastic real-world environments are also being integrated bythe researchers. They integrate routing with soft time windows

   [21] and optimization based on Vehicle-to-Grid (V2G) [15]. Planning for the charging infrastructure itself, such as optimal location and load balancing, is another pressing line of research [22], [23]. For the purpose of determining the efficacy and strength of such advanced models, most of the studies employ benchmarked data sets and realistic simulations to provide evidence to their testing process [19].

   I. Methodology

   1. Local Client Update: In each training round, each client k receives the current global model weights w. The client then trains the model on its local dataset Dk by minimizing a

      loss function Lk using an optimizer like Stochastic Gradient Descent (SGD). For a single data point, the local model

      weights are updated as:

      w w Lk(w) (4)

      where is the learning rate. For our predictive tasks, we use the Mean Squared Error (MSE) as the loss function:

      Our solution follows a federated learning (FL) scheme L = 1 2

      where two domain-specific predictive models are trained. The system is implemented in Python, employing Flower for FL

      k

      |Dk| (xi,yi)Dk

      (yi f (xi; w))

      (5)

      orchestration and PyTorch for the model design.

      1. Experimental Setup

         Experiments on this project were performed on a local sys- tem with Windows 10, an Intel Core i5 (11th Gen) processor,16 GB of RAM, and an NVIDIA GeForce GTX 1650 GPU.

      2. Mod Formulation and EV Constraints

         Any EV routing model to be of relevance needs to work within some fundamental physical constraints. Our predictive models are designed to learn implicitly subject to these con- straints.

         1. State-of-Charge (SoC) Dynamics: The battery level on reaching a destination node j from an origin node i is determined by energy consumed, Eij:

            SoCj = SoCi Eij (1)

            The energy consumption itself is often modeled as a function of distance dij and a consumption rate :

            Eij = · dij (2)

         2. SoC Bounds: The battery level should be within a protected operation envelope along the route:

            SoCmin SoCj SoCmax j Route (3)

      3. erated Learning Framework

         The core of our methodology is the federated training pro- cess. We employ the Federated Averaging (FedAvg) algorithm, which consists of local client updates followed by server-side aggregation.

         .

         where yi is the true label and f (xi; w) is the models predic- tion for input xi.

         k

   2. Server-Side Aggregation: After local training, each client k sends its updated weights w(t+1) to the central server. The server aggregates these updates to form the new global model for the next round t + 1:

   n k

   k=1

   where nk is the number of data samples on client k.

   1. Core Predictive Models

      Under the above described FL framework, we train two identical Multi-Layer Perceptron (MLP) models as shown in (Fig. 2):

      • Station Utility Model: It is a model U (xs; wu), that takes a vector of static station features xs and predicts a utility score. The parameters of the model are wu.

      • Route Cost Model: This model, C(xr; wc), takes a vector of route features xr as input and predicts the associated travel cost (e.g., time). Its parameters are wc.

   2. Data Pre-processing

      As demonstrated in Phase 1 of our architecture (Fig. 1), data for the Station Utility Model is collected from actual station listings, cleaned and normalized. Data for the Route Cost Model is produced synthetically to provide a diverse set of routing situations to train on.

   3. Integrated Recommendation

   The final recommendation is produced by a script that loads both trained models. Based on the location of a user, it determines a final score for a possible charging station using a subtractive utility-cost function:

   Scorefinal = U (xs; wu) · C(xr; wc) (7)

   where is a variable parameter to maximize the relative weighting of station quality in comparison to travel cost. The stations are ranked subsequently on this combined score to return the highest-recommended choices to the user.

   Fig. 2. The core federated learning (FL) training workflow, detailing the iterative cycle of local client training and central server aggregation.

   Fig. 3. The end-to-end recommendation architecture, showing how trained models and external APIs are integrated to serve the end-user .

3. Results and Discussion

   Our first observation is that Federated Learning (FL) is a tremendous opportunity in electric vehicle routing and charg- ing station utility estimation application use. Our experiments on a synthetically generated dataset observed FL models matching performance of baseline centralized models such as Random Forest and XGBoost, especially for station utility estimation.

   While FL-based route cost estimation lags behind central- ized baselines by a hair, this is within the parameters of what can be reasonably expcted from synthetic data as well as not having access to real-world temporal traffic patterns. Most notably, models were trained within a privacy-preserving fed- erated setting without ever exposing raw datathus demon- strating FL performs flawlessly even within restrictive settings.

   Fig. 4. Graph 1

   Fig. 5. Graph 2

   These results show that, as they learn from more compre- hensive real-world data, FL models are not only competitive but actually superior to baseline methods. The potential of collaborative learning with no data centralization presents a promising way forward for large-scale privacy-preserving intelligent mobility systems.

   Fig. 6. result

   The most important point to take away from our experiment outcome is not that the federated learning model will necessar- ily do better than its centralized equivalents, but that it gives equivalent performance with a data privacy guarantee inherent in the system. This trade-off is the crux of the contribution of this paper. In a real-world deployment scenario of intelligent mobility where user travel patterns, car performance data, and current positions are highly sensitive in nature, their privacy is of highest importance. The minimal loss of prediction accuracy by the route cost model is a price worth paying for a fundamentally secure, scalable, and user confidentiality- sustaining system.

   Our experiments unequivocally show that federated learning

   is a viable candidate as a basis for the next generation of intel- ligent transportation systems. Dual model training of stationary station benefit and dynamic route cost efficiently validates our main assumption of the designed architecture. It proves that challenging, multi-dimensional recommendation tasks can be efficiently solved by decomposition and decentralized solving. Also, the loss in performance in route cost estimation is symptomatic of one of the more commonly known issues for federated learning: non-IID data distribution. The synthetic data were split to simulate this real-world situation in which each client (vehicle) would have different driving patterns and represent a distinct geography. Though the baseline FedAvg algorithm employed here is adequate as a point of comparison, results suggest that performance could be further optimized by utilizing more advanced FL algorithms like FedProx or FedAvgM, which have been designed to provide better con- vergence and robustness in the presence of heterogeneous data.

   Lastly, this project also exists as a proof-of-concept critique. It demonstrates that subsequent efficient smart, large-scale EV routing does not necessarily have to be at the expense of privacy. With continued optimization on optimized federated algorithms and being exposed to larger real-world datasets, FL-based systems can be an even better alternative than the

   1. M. Zhang, Hybrid GA-ACO for Large-Scale EV Routing, Journal of Heuristics, vol. 29, no. 3, pp. 210225, 2023.

   2. P. Kumar, Improved Metaheuristics for EV Routing, Soft Computing, vol. 28, no. 4, pp. 567578, 2024.

   3. F. Yuan, A Comprehensive Survey of Federated Meta-Learning, ACM Computing Surveys, vol. 55, no. 2, pp. 134, 2023.

   4. J. Wang, Review of Federated Meta-Learning in Smart Mobility, IEEE Communications Surveys, vol. 25, no. 1, pp. 5675, 2023.

   5. Y. Li, Survey on Federated Meta-Learning for EV Routing, AI Review, vol. 47, no. 3, pp. 123145, 2024.

   6. Y. Li and X. Wang, Real-Time Distributed Charging Station Recom- mendation for Electric Vehicles: A Federated Meta-RL Approach, IEEE Trans. Intell. Transp. Syst., 2024.

   7. J. Gao, M. Zhang, and H. Wu, Genetic Algorithm-Based Optimization of EV Charging Station Placement on Long-Distance Routes, in Proc. IEEE Int. Conf. on Smart Transportation, 2024.

   8. L. Zhang and P. Chen, Integrated Model for Routing and Charging Coordination with Power-Aware Operations, Sustain. Cities Soc., 2024.

   9. H. Liu and A. Zhang, Electric Vehicle Path Optimization Based on Charging and Switching Methods under V2G, Sci. Rep., vol. 14, no. 1, Art. 30843, Dec. 2024.

   10. F. Meng, Q. Li, and S. Wang, The Method of Route Optimization of Electric Vehicle, Adv. Oper. Res., vol. 2020.

   11. R. Smith et al., Optimizing Electric Vehicle Routing using ACO, GA and SA Algorithms, Int. J. Electr. Comput. Eng., vol. 13, no. 4, pp. 1 10, 2023.

   12. A. Vertuzzo, L. Bianchi, and M. Costa, Personalized FL with Context Modulation and Meta Learning, Proc. FedML Workshop, 2023.

   13. S. Vani and R. Patel, An Efficient Optimization Algorithm for Electric Vehicle Routing Problem, IET Power Electron., vol. 16, no. 5, pp. 12341242, 2023.

       traditional centralized alternative.

4. CONCLUSION

This paper presented a new federated learning paradigm to

2. X. Chen, Y. Liu, and A. Al-Dujaili, Federated Meta-Learning with Fast Convergence and Efficient Communication, arXiv:1802.07876, 2018.

   Y. Meng and T. Liu, Route Optimization of Electric Vehicle Consider- ing Soft Time Windows and Two Ways of Power Replenishment, Adv. Oper. Res., 2020.

   solve the twin problem of electric vehicle route planning and

   station selection in a privacy-preserving way. Our two-model solution, where predictors for route cost and station utility are learned end-to-end in unison, proved to be a workable solution. The experimental result demonstrated that our FL model is able to achieve predictive accuracy levels on par with centralized baselines, placing it within reach of real-world deployment without sacrificing user data. Major limitations of this work, including the utilization of synthetic data and the baseline FedAvg algorithm, are clear areas for future work. Key areas for future research involve incorporating real- time traffic data, investigating more advanced FL algorithms such as FedProx to improve heterogeneity robustness of the data, and developing API-level connectivities with industrial navigation systems. Finally, this paper gives a proof-of-concept foundation, demonstrating federated learning to be a robust and scalable technology that can facilitate the next generation of ITS

3. P. Gupta, S. Rao, and L. Sharma, Electric Vehicle Charging Sta- tions: Model, Algorithm, Simulation, Location and Capacity Planning, Transp. Res. Part C, 2024.

4. J. Lee and K. Kim, An In-Depth Exploration of Electric Vehicle Charging Station Infrastructure, Energy Policy, 2024.

5. X. Wang et al., EV Charging Route Planning for Shortest Travel Time Based on Improved ACO, Sensors, vol. 25, no. 176, 2025.

6. A. Brown and D. Green, Optimizing Long-Distance EV Routes Based on Passenger Satisfaction Model, Transp. Res. Part A, 2024.

References

1. J. Lopez, A Comprehensive Survey on Electric Vehicle Routing, IEEE Transactions on Smart Transportation, vol. 12, no. 3, pp. 123134, 2023.

2. Y. Wang, Multi-Objective Genetic Algorithms for EV Routing, Ap- plied Soft Computing, vol. 145, pp. 112123, 2023.

3. L. Chen, Improved Ant Colony Optimization for EV Routing, Swarm Intelligence, vol. 29, no. 2, pp. 201215, 2024.

4. H. Li, Simulated Annealing for Multi-Objective EV Routing, Opti- mization Letters, vol. 17, no. 1, pp. 5570, 2023.

5. R. Sharma, Hybrid Bat Optimization for Charging Station Scheduling, Energy Reports, vol. 10, pp. 12341245, 2024.

6. X. Guo, MILP Models for EV Charging Station Placement, IEEE Transactions on Smart Grid, vol. 15, no. 2, pp. 345356, 2024.