Connected Mobility and Federated AI: How Can We Quantify and Minimize Energy Costs?

• AI
• Connectivity
• CSR
• Networks
• Smart city

Share on Twitter : Connected Mobility and Federated AI: How Can We Quantify and Minimize Energy Costs? (New window) Share on Linkedin : Connected Mobility and Federated AI: How Can We Quantify and Minimize Energy Costs? (New window)

Our Cities of Today and Tomorrow

Connectivity has become, and will continue to be, an essential component of our cities. In Barcelona, for example, connected mobility relies on a network of IoT sensors and real-time telemetry integrated into public transportation. The traffic management center uses these data streams with Artificial Intelligence (AI) to dynamically synchronize traffic lights and optimize bus frequencies. This interconnection between infrastructure and control centers enables a 20% reduction in congestion and improves punctuality through predictive analysis of traffic flows [1]. On the other side of the Atlantic, in Columbus, United States, the focus is on the “CASE” ecosystem (Connected, Autonomous, Shared, Electric) to streamline urban mobility. The city has deployed connected mobility hubs and the Pivot application, which aggregates real-time data from buses and ride-hailing services. In parallel, autonomous vehicles equipped with advanced sensors and connected to urban networks, such as Zoox robotaxis, are testing vehicle-to-environment communication to secure complex journeys. By combining the Internet of Things (IoT), connectivity, and AI, these metropolitan areas are transforming data into operational decisions to develop smart cities and a robust, sustainable road network.

One of the challenges posed by federated AI is that it relies heavily on clients’ computing and communication resources, which incur high energy costs that must be optimized.

Focus on Connected Vehicles

Smart cities rely on connected vehicles as essential components of their digital and communication ecosystem. Researchers in the Sustainable Mobility program at Orange Belfort are studying the specific case of AI use in a connected vehicle context, as well as the development of mechanisms to minimize energy costs. The topic is gaining importance with the emergence of connected equipment and the growth in vehicle connectivity rates, which could reach 95% by 2030 [2]. Connected and autonomous vehicles will transform mobility management and can generate up to 25 Gigabytes [3] of data per hour: on the one hand, they are equipped with a multitude of sensors for rich environmental perception; on the other hand, advances in telecommunication networks enable low-latency exchanges between vehicles and urban infrastructure, the network, or pedestrians—this is known as V2X (Vehicle-to-Everything) communication. These data feed AI models to, for example, predict traffic conditions, issue collision alerts, or promote eco-responsible driving. However, these services raise issues related to user data privacy. Federated AI, or Federated Learning (FL), is an innovative approach standardized by the 3rd Generation Partnership Project (3GPP) [4] that enables the construction of high-performance AI models without collecting users’ personal data [5]. This approach is based on the following principle: each user, called a client, trains a purely local model on their own data (without ever moving it), then a central server aggregates these models to obtain a global model enriched with knowledge from each client. This global model is used by all clients for inference and is free of personal data. Federated AI has the advantage of being applicable to a wide variety of models: deep neural networks (convolutional networks for Computer Vision, recurrent networks or Transformers for language, etc.), logistic regression models, as well as decision trees and random forests. This flexibility allows federated AI to adapt to many use cases on decentralized data.

Managing the Energy Consumption of Federated AI

One of the challenges posed by federated AI is that it relies heavily on clients’ computational and communication resources, with high energy costs that are crucial to optimize [6-8]. The literature on FL energy optimization addresses several aspects, notably strategies such as optimal client selection [9] or model weight compression [10], a technique that reduces the storage space required for the model without significantly diminishing its performance. Nevertheless, most studies consider a static environment that is therefore not subject to variations in mobile network state [11], with consequently energy cost estimations that are, on the one hand, not generalizable to the vehicular study case due to the rapid fluctuation of vehicle client states (particularly the state of connectivity to the mobile network), and on the other hand, based on theoretical approximations.

From Theory to Practice

The Belfort research team conducted a proof of concept on the feasibility of federated AI on the Orange network, as well as a quantification of the field energy cost of this service for the client. The test architecture relies on a V2X platform deployed at the Belfort site. An MQTT (Message Queuing Telemetry Transport) server acts as the federated learning server and sends an MQTT trigger, that is, an alert message, to the moving connected vehicle. The vehicle, which in this test plays the role of several clients, then performs an http request secured by basic access authentication to download the current global model and compute the local models with Python’s PyTorch framework. Finally, the vehicle transmits client models to the MQTT server for aggregation. The energy cost of computations is estimated on the hardware in real time by the Python library CodeCarbon [12], chosen for reasons of hardware compatibility and ease of use. CodeCarbon estimates the energy cost of Python code execution based on the power consumption of the computer’s computing components (CPU, GPU, and RAM). As for the cost of communications with the server, it is calculated from the measurement of network KPIs (Key Performance Indicators), particularly via the signal-to-noise ratio (SNR), which allows estimation of the vehicle’s transmission channel capacity at each instant. The maps below illustrate an example of a journey with, on the left, the vehicle speed, and on the right, the estimated communication energy per model transmission. The measurements allow observation of the impact of mobility on signal quality, but also identification of levers to optimize the energy cost of federated AI. In particular, the energy cost per packet transmission on the client side is on average 30% higher when the client performs a handover, that is, switches from one relay antenna to another during its movement.

Perspectives

To reduce this energy overhead, Orange Belfort researchers are exploring several avenues. First, a partial participation strategy where clients that are least likely to be subject to handovers are selected. A second avenue is a transmission queuing mechanism where the client waits for its radio conditions to stabilize before transmitting its local model. To be even more robust, this solution can be combined with asynchronous aggregation methods on the server side.