Federated learning for smart charging of connected electric vehicles

Actas de las XV Jornadas de Ingeniería Telemática (JITEL 2021), A Coruña (España), 27-29 de octubre de 2021. This work is licensed under a Creative Commons 4.0 International License (CC BY-NC-ND 4.0) Federated learning for smart charging of connected electric vehicles Yaqoob Al-Zuhairi, Prashanth Kannan, M´ onica Aguilar Igartua Department of Network Engineering Universitat Polit´ ecnica de Catalunya (UPC), Barcelona, Spain yaqoob[email protected], [email protected], [email protected] Due to rising concerns over climate change, air pollution and clean energy awareness, the demand for electric vehicles (EVs) and renewable energy generation has increased in recent years. The main objective of this research is to design a decentralized smart charging coordination framework for EVs based on federated learning (FL) algorithms in order to provide an acceptable collaboratively learning model with privacy preservation of EVs, improve charging scenarios, contribute to smart grid stabilization, meet EVs energy requirements wherever and whenever they request, and gain welfare for EV owners. Moreover, FL is introduced with the goal of bringing machine learning (ML) down to the edge level in vehicular networks. Ultimately, a multimetric routing protocol is also used to predict the best route for transmitting messages among EVs, infrastructures, charging stations (CSs), and central servers. Key words—FL, Edge Computing, VANET, V2X, Multimetric Routing Protocol I. INTRODUCTION In recent years, the transport industry accounts for the bulk of greenhouse gas emissions and pollution to the environment [1]. To tackle this issue, governments are implementing policies to support use of renewable energies, decrease the dependency on crude oil, reduce CO2and pollutant emissions, and promote transition to more sustainable mobility. Current electric vehicles (EVs) as a part of the intelligent transport systems (ITS) have noticeably attracted substantial attention recently because they are very friendly to the environment as well as pollution-free vehicles. The recent prominent progress in the construction of charging infrastructure accelerates the penetration of EVs in the market. In 2018, the global EV fleet in the world exceeded 5.1 million units and it is expected to rise to 250 million units by 2030 [2]. Commercialization of the fifth-generation (5G) communication technologies and the emergence of vehicular networks and edge computing can promote a superior performance of the charging management [3]. In the meantime, future intelligent vehicles, which are at the heart of high mobility networks, are increasingly equipped with a wide variety of sensors to help the vehicle perceive the surrounding environment as well as monitor its own operational status in real time. Together with high performance computing and storage devices, these sensing technologies are transforming vehicles from a simple transportation facility to a powerful computing and networking hub with intelligent processing capabilities [4]. However, privacy concerns arise from the exchange of data with different parties. Privacy, in recent years, has been one of the most important concerns in vehicular environments. Thus, EVs may fail to exchange data among themselves and other parties due to privacy restrictions in situations where drivers are unwilling to provide their personal data due to the risk of data misuse and leakage [5]. In order to solve the aforementioned issue, federated learning (FL), which is a technique that enables distributed vehicles to collaboratively learn a shared machine learning (ML) model without sharing their raw data, would be a good solution to preserve the privacy of the local data [6]. Vehicular networks, e.g., vehicular ad hoc networks (VANETs) and cellular V2X (C-V2X), have gained popularity in recent years. In such networks vehicles equipped with wireless communication devices form vehicular networks [7]. However, with the evolution of technology and sudden growth in the number of smart vehicles, this will result in unprecedented pressure on communication infrastructures. Moreover, traditional VANET faces several technical challenges in deployment and management due to less flexibility, less resources, scalability, poor connectivity, and inadequate intelligence. Merging vehicular networks with edge computing, an emerging paradigm which moves computing tasks and services from the core to the network edge, i.e. closer to end users, is an appropriate solution for these types of challenging networks. Vehicular networks show special features such as high vehicles’ mobility, difficult network connectivity, which is specially challenging for real-time applications 212 Al-Zuhairi, Kannan, Aguilar, 2021. requiring low latency. The notion of exploiting vehicles as infrastructures could make the best use of several unused resources of vehicles to meet the ever increasing requirement in communication and computational capabilities [8]. FL can utilize vehicular big data generated from a large number of vehicles, and build a global ML model that could be used by any vehicle. FL can also be easily incorporated with edge computing where the edge vehicles provide an underlying infrastructure for FL [5][9]. Vehicular communications are crucial for exchanging data either between vehicles through vehicle-to-vehicle (V2V) or via vehicle-to-infrastructure (V2I) connections [10]. Although road side units (RSUs) enlarge the network communication capacity, they are really expensive and difficult to fully deploy along roads, particularly on a large scale such as over a whole city. Furthermore, future vehicles will need to communicate with everything around them in what is known as vehicle-to-everything (V2X) [10]. To enable hybrid vehicular communications, dedicated short-range communication (DSRC), which is based on IEEE 802.11p, and cellular V2X (C-V2X) can be adopted. In this paper, we present the topics that will be studied in the first authors’ doctoral thesis. In section II we highlight some related works. Section III summaries the basics of FL framework for vehicular networks. Since we consider the presence of EVs in our proposals, section IV depicts innovative ways to charge EVs, which will be taken into account in the design of our FL framework. Section V lists our ongoing work, whereas section VI shows which simulators are we going to use to evaluate the performance of our proposals. Finally, section VII concludes the paper. II. LITERATURE SURVEY The authors in [11] propose a scheme for charging moving EVs in a wireless fashion. An unsupervised ML algorithm is used to estimate the current charging status of each EV. Moreover, a routing protocol using distance vector information was used to advertise participating EVs about their state of charge (SoC), vehicle ID, distance information, and location. Authors mentioned that sharing such information among EVs causes concerns related to privacy reduction. Results in this work show to be reliable in terms of dynamic wireless charging in both static and dynamic scenarios. In [9], a survey of technical challenges, possible solutions, open problems, and future research directions for applying FL in vehicular networks are discussed. In [6], an FL approach for learning on edge devices is studied. The paper analyzes the effect of participating clients and the importance of client selection strategies in FL models. The authors investigate the performance of the FL model focusing on the effect of various parameters on its accuracy and training time, as well as comparing the performance to a traditional ML technique. In the paper [12], a new proposal of a novel probabilistic multimetric routing protocol is presented. The proposal takes better forwarding decisions that guarantee the packet delivering to destination with the highest probability, while keeping the average packet delay low. Four designed metrics are considered in this article (distance to destination, vehicles’ density, position of vehicle, and available bandwidth). OMNeT++, VEINS, and SUMO are used to conduct simulations in a realistic urban scenario. III. IMPLEMENTATION OF AN FL FRAMEWORK Instead of sending raw data to a central server, which is common in the traditional centralized ML approach, FL can be beneficial in vehicular environments to support cooperation of multiple EVs with the central server. This way, distributed EVs will train a partial model using own local data to ensure privacy protection. The global model in the central server will be updated from the parameters’ aggregation of those partial models. The basic steps for an FL framework are as follows: •Client selection: Initially, the central server must specify those EVs that should be involved in the model training. •Model dissemination: Once the EVs are selected, the central server broadcasts an initial learning model for the training to the selected EVs. •Distributed learning: Each EV trains the model based on its own local dataset, and then calculates its local updates of the global model. •Global model aggregation: After a predefined training period to aggregate the new version of the global model, all EVs send only their updates of the common global model to either the central server or to an aggregator. An EV can be designated as the aggregator looking to be close to participating EVs. The aggregator EV will manage the training process and ultimately send the result to the central server [13]. •Model testing: The central server tests the aggregated global model. According to the testing results, the central server could tune some hyper-parameters to repeat the training process, or continue with the next step which is the model update. •Model update: The server updates the shared model according to the aggregated result from the EVs. Then, the server sends the updated global model to the EVs. These steps are repeated until the central server achieves a satisfactory global model. While designing an FL framework, it is significant to select a variety of parameters and compare multiple scenarios to attain the trade-offs between privacy, efficiency, and accuracy in this environment [5]. Furthermore, different from conventional decentralized ML approaches, FL in vehicular environments is expected to achieve many advantages, such as low communication overhead, better privacy, larger data for training, better efficiency, better utility, and shorter response time. IV. FUTURE WAYS TO CHARGE ELECTRIC VEHICLES Currently, the most common form of charging EVs is via a plug-in charging station. However, there are many This work is licensed under a Creative Commons 4.0 International License (CC BY-NC-ND 4.0) 213 Federated learning for smart charging of connected electric vehicles issues reducing drivers’ willingness to use EVs instead of traditional vehicles, such as limited number of charging stations (CSs). Furthermore, the nearest CS is perhaps at an inconvenient site in relation to an EV’s route. In this point, the driver may be feeling anxiety over the need to recharge the EV during a journey. A further issue may happen after the EV reaches the plug-in CS, the driver may find the charging slots already taken by other EVs. These issues can effectively be solved by adopting a new structure of wireless charging technology for EVs based on the magnetic resonant coupling wireless power transfer technique, which was introduced in the last decade [11]. This technique is used to exchange charge among paired EVs in a static position as well as in dynamic position (motion). This structure can work together with plug-in charge EVs or operate independently in order to increase charging opportunities for EVs. There are mainly two ways for wireless charging: [14] •Inductive charging: It includes a transmitter coil, which is embedded in the floor of the charging area and connected to the power supply (grid), while a receiver coil which is embedded in the EV’s chassis and connected to the EV’s battery. •Wireless vehicle-to-vehicle (V2V) charging: It is based on wireless power transfer technique which has high power transfer efficiency with a long transmitting distance. This technology can achieve charging between moving EVs. In our work, we plan to use several charging technologies to meet requirements for EVs with a low charging state. We will design a FL-based framework to assists the driver in finding a suitable power source based on the current circumstances of the environment, e.g. energy requirement, distance to destination and to possible CSs, pricing, waiting time, ways of charging such as (mobile CS, plug-in CS, V2V charging, or inductive charging), and other factors. Our FL-based framework will be assisted by V2X communication, SoC detection, multimetric routing protocol, and edge computing. Figure (1) demonstrates the scenario of decentralized smart charging coordination for EVs based on FL. This charging coordination can offer solutions for many problems, such as: 1) Offering solutions for the issue of having a limited number of CSs. 2) Solving the problem of increasing numbers of EVs in a plug-in EV network. 3) Reducing charging delay, and minimizing overall EV energy costs. 4) Avoiding the grid congestion, especially in the evenings, that results from simultaneously charging too many EVs. 5) Forecasting of EVs energy requirements. 6) Exchanging data among EVs or with the central server while ensuring EV’s privacy preservation. 7) Generating monetary benefits for EV owners. 8) Finding the cheapest charging options by giving an EV driver the optimal choice for charging. Fig. 1. Decentralized smart charging coordination framework for EVs based on federated learning (FL) algorithms. 9) Encouraging consumers to purchase EVs. V. ONGOING WORK The main contributions that we have planned in this Ph.D. thesis, can be summarized as follows: 1) Employing mobile EVs as infrastructures for communication and computation to achieve a better utilization of edge resources. 2) Privacy of each EV can be preserved by using FL. 3) Efficient collaboration amongst EVs with FL could reach the level of collaborative intelligence. Such intelligent collaboration can contribute to remarkably minimize the waiting time of charging. 4) Due to the potentially high mobility and sparse/dense environments, a multimetric routing protocol with FL can be employed to address the issue of providing an efficient forwarding scheme in V2X communications. 5) SoC, which indicates the level of charge available in each EV, is used to predict the residual energy of EVs based on the designed FL framework. 6) As for charging time available, determine the best power source either via wireless or plug-in through: •Considering the charging cost and duration when selecting either the best (e.g. the nearest) CSs or those EVs having high level of power to share, among other strategies. •Using an efficient multimetric routing protocol to improve the data transmission needed in the charging service regarding the chosen power source. 7) Estimating travel time to save energy according to the charging available time and the EV’s destination, and boost the energy economy of EVs in charging processes. 8) Based on our FL-based charging framework, reducing load on CSs while keeping the grid stable can be achieved. VI. IMPLEMENTATION AND SIMULATION Because of the complexity and the high deployment costs of vehicular applications, it is recommended that This work is licensed under a Creative Commons 4.0 International License (CC BY-NC-ND 4.0) 214 Al-Zuhairi, Kannan, Aguilar, 2021. proposals are extensively tested and evaluated in possible ways before being put into practice. TensorFlow [15] is an open-source FL library that offers an extensive range of algorithms. It also works in conjunction with Keras and Python. The construction of ML models and the design of parameters can be evaluated by computing metrics such as accuracy or confusion matrix. Through simulations we will represent the whole system trustworthy as close to reality as possible, and we will assess the performance of the proposed system. To simulate a vehicular network in urban scenarios, OMNeT++, SUMO, VEINS, and Artery-C are used. •OMNeT++: It is a modular, object oriented and discrete event simulator based on C++ [16]. •Simulation of Urban Mobility (SUMO) tool: It is a highly portable and open-source software to simulate the movement of the vehicles [17]. •Vehicles in networks simulations (VEINS) tool: It is a framework for vehicular network simulation. It facilitates the bidirectional interaction among SUMO and OMNeT++ simulators [18]. •Artery-C framework: Artery-C [19] is a simulation framework for the performance evaluation of Cellular V2X protocols and V2X applications. It is an extension of the simulation framework SimuLTE [20], developed under the OMNeT++ platform. VII. CONCLUSIONS In vehicular environments, it is not always feasible to transfer massive amounts of data for processing to a remote central server due to limiting factors, such as unstable and limited wireless connectivity, unacceptable latency, and network bandwidth. Moreover, the interests on data availability have triggered the discussion of data ownership, and the concern on data privacy and confidentiality especially when sharing such data with central servers. In this research work we will design a decentralized smart charging coordination framework for EVs based on FL taking advantage of the distributed power of edge computing. Our goal is to encourage collaboration amongst participating EVs in training a global model locally while ensuring privacy preservation for each EV, and then sending it to a central server to obtain a global model. The aim is to optimize charging decisions, offer efficient EV charging service for urban areas when EVs are in a critical SoC, avoid grid overload, and acquire drivers’ comfort and satisfaction. As a future work, an incentive mechanism will be used to motivate EVs with their private data to participate in FL training in order to improve the global FL model accuracy. Eventually, a driver can also maximize their own utility through contribution to either the training process or charging another EV having low level SoC in wireless way. VIII. ACKNOWLEDGEMENTS This work was supported by the Spanish Government under research project “Enhancing Communication Protocols with Machine Learning while Protecting Sensitive Data (COMPROMISE)” PID2020-113795RBC31/AEI/10.13039/501100011033. REFERENCES [1] A. Ghosh, “Possibilities and challenges for the inclusion of the electric vehicle (ev) to reduce the carbon footprint in the transport sector: A review,” Energies, vol. 13, no. 10, 2020. [Online]. Available: https://www.mdpi.com/1996-1073/13/10/2602 [2] Global EV outlook 2019 - scaling-up the transition to electric mobility, 2019. [3] S. Zhang, W. Dou, Y. Zhang, W. Hao, Z. Chen, and Y. Liu, “A vehicle-environment cooperative control based velocity profile prediction method and case study in energy management of plug-in hybrid electric vehicles,” IEEE Access, vol. 7, pp. 75 965–75 975, 2019. [4] W. Xu, H. Zhou, N. Cheng, F. Lyu, W. Shi, J. Chen, and X. Shen, “Internet of vehicles in big data era,” IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 1, pp. 19–35, 2018. [5] V. Mugunthan, A. Peraire-Bueno, and l. Kagal, “Privacyfl: A simulator for privacy-preserving and secure federated learning,” in CIKM ’20: The 29th ACM International Conference on Information and Knowledge Management, 2020. [6] S. F. Lameh, W. Noble, Y. Amannejad, and A. Afshar, “Analysis of federated learning as a distributed solution for learning on edge devices,” in 2020 International Conference on Intelligent Data Science Technologies and Applications (IDSTA), 2020, pp. 66–74. [7] M. S. Sheikh, J. Liang, and W. Wang, “Security and privacy in vehicular ad hoc network and vehicle cloud computing: A survey,” Wireless Communications and Mobile Computing, vol. 2020, no. 5129620, 2020. [8] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, and S. Chen, “Vehicular fog computing: A viewpoint of vehicles as the infrastructures,” IEEE Transactions on Vehicular Technology, vol. 65, no. 6, pp. 3860– 3873, 2016. [9] Z. Du, C. Wu, T. Yoshinaga, K.-L. A. Yau, Y. Ji, and J. Li, “Federated learning for vehicular internet of things: Recent advances and open issues,” IEEE Open Journal of the Computer Society, vol. 1, pp. 45–61, 2020. [10] M. Boban, A. Kousaridas, K. Manolakis, J. Eichinger, and W. Xu, “Connected roads of the future: Use cases, requirements, and design considerations for vehicle-to-everything communications,” IEEE Vehicular Technology Magazine, vol. 13, no. 3, pp. 110–123, 2018. [11] M. Adil, J. Ali, Q. T. H. Ta, M. Attique, and T.-S. Chung, “A reliable sensor network infrastructure for electric vehicles to enable dynamic wireless charging based on machine learning technique,” IEEE Access, vol. 8, pp. 187 933–187 947, 2020. [12] Leticia Lemus C´ ardenas, Ahmad M. Mezher, Pablo A. Barbecho Bautista, M´ onica Aguilar Igartua, “A probability-based multimetric routing protocol for vehicular ad hoc networks in urban scenarios,” IEEE Access, vol. 7, pp. 178 020–178 032, 2019. [13] J. Posner, L. Tseng, M. Aloqaily, and Y. Jararweh, “Federated learning in vehicular networks: Opportunities and solutions,” IEEE Network, vol. 35, no. 2, pp. 152–159, 2021. [14] X. Mou, R. Zhao, and D. T. Gladwin, “Vehicle to vehicle charging (V2V) bases on wireless power transfer technology,” in IECON 2018 - 44th Annual Conference of the IEEE Industrial Electronics Society, 2018, pp. 4862–4867. [15] “TensorFlow,” https://www.tensorflow.org/federated/federated learning. [16] “OMNeT++,” https://omnetpp.org/intro, accessed: 2020-10-06. [17] “SUMO,” https://www.eclipse.org/sumo/. [18] “Veins,” http://veins.car2x.org/. [19] A. Hegde and A. Festag, “Artery-C: An OMNeT++ based discrete event simulation framework for cellular V2X,” in MSWiM ’20: 23rd International ACM Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems, 2020. [20] A. Virdis, G. Nardini, and G. Stea, “Modeling unicast device-todevice communications with simuLTE,” in 2016 1st International Workshop on Linkand System Level Simulations (IWSLS), 2016, pp. 1–6. This work is licensed under a Creative Commons 4.0 International License (CC BY-NC-ND 4.0) 215