Range-risk estimation for electric vehicles in low-coverage mountain road segments: A digital-twin study with prequential beacon-based Federated Learning

List of Authors

  1. Yeray Mezquita ,  
  2. Jesús Emmanuel Vidal Cuevas ,  
  3. Diego Valdeolmillos ,  
  4. Albano Carrera
Article Authors & roles Authors Peer review record Peer review Metrics
Cite

Abstract

Electric vehicles operating on isolated mountain roads may receive over-optimistic range estimates precisely where connectivity and charging alternatives are scarce. This paper proposes a beacon-centric Vehicle-to-Everything and Federated Learning architecture for range-risk estimation in low-coverage mountain-road segments. Entry and exit beacons maintain a local segment-specific model, broadcast it to approaching vehicles, and update it from traversal evidence contributed by vehicles after leaving the segment, without requiring raw telemetry to be centralised. The approach is evaluated in a digital twin of a 10.6 km fragment of the Puerto de la Quesera mountain pass using 12 simulated traversals from two electric-vehicle models under dry, wet and snow/low-temperature conditions. This study presents a proof of concept: its scale demonstrates the mechanism but not population-level accuracy. Road grade is derived from vehicle pitch; against the digital-twin mesh geometry, the average resulting error is 0.36% of route energy and the worst error is 0.74%. The evaluation is route-level, matching the operational decision made before entering the segment. Generic estimators systematically underestimate the energy required to complete the section: the nominal-consumption estimator produces a mean absolute error of 0.777 kWh, while a conventional non-local physics estimator produces 1.975 kWh. A segment-specific residual model reduces the leave-one-route-out error to 0.220 kWh. More importantly, the proposed alert-oriented evaluation shows that false-safe decisions are reduced from 14.58% for the non-local estimator and 8.33% for the nominal estimator to 1.04% for the segment-specific residual model. In the prequential beacon simulation, after eight previous vehicles, FedProx reduces the false-safe rate to 3.38%. Cross-vehicle generalisation remains the main limitation: the leave-one-vehicle-out error rises to 1.568 kWh, and both vehicles share one energy model with different parameters. These results support the feasibility of local, privacy-preserving range-risk intelligence for challenging road segments.

Corresponding author
Albano Carrera
Grant information

Project name “A catalyst for EuropeaN ClOUd Services in the era of data spaces, high-performance and edge computing (NOUS)”, Grant Agreement Number 101135927. Funded by the European Union.

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Copyright
© 2026 Yeray Mezquita (Co-Author); Jesús Emmanuel Vidal Cuevas (Co-Author); Diego Valdeolmillos (Co-Author); Albano Carrera (Corresponding Author, Co-Author) Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

First version published
Latest version published
Revised

Amendment

The revised version 2 addresses the reviewers’ comments. No new experiments were conducted, and no previously reported results were changed. All newly added quantitative results were derived from the twelve traversal logs previously published in Version 1.

Text. The proof-of-concept scope, the origin of the road-grade signal, and the limitations of cross-vehicle generalisation are now specified in the abstract. In the Introduction and Section 2.6, homologated (WLTP) range is distinguished from dynamic on-board range estimation. In Section 3.3.2, the pitch-derived road grade is quantitatively validated against the digital-twin mesh geometry, yielding a mean absolute difference of 0.039 degrees and route-energy differences of 0.36% on average and 0.74% at worst. The unfiltered nature of the pitch signal is clarified, and its residual relative to longitudinal acceleration is decomposed. In Section 3.4.1, “Coulomb counting” is corrected to “energy-based integration”. The shared longitudinal model structure used for both vehicles is also highlighted, clarifying that the results evaluate the methodology rather than production-vehicle behaviour. This limitation is reiterated in Sections 4.1 and 5. Section 3.5 describes the aggregation of federated results. Section 4.5 specifies the number of independent observations underlying the alert decisions and explains the non-local estimator’s zero risk recall. In Section 3.3.1, the tool name is corrected to mapperV3.

Tables, figures, and references. Two tables have been added: a comparison of centralised and federated models using common metrics (Section 4.4, Table 9), and a sensitivity analysis of alert results with respect to the reserve threshold (Section 4.5, Table 12). A FedAvg false-safe column has also been added to Table 11.

Data and code. In Section 6, the descriptions of the archived record, logged variables, and their units have been corrected. The software repository remains unchanged.

Comments

Authors & roles

Yeray Mezquita

Deep Tech Lab, AIR Institute, Salamanca, Castilla y León, 37003, Spain

Roles
Co-Author, Conceptualization, Data Curation, Formal Analysis, Investigation, Methodology, Software, Validation, Writing – Original Draft Preparation
0000-0003-4813-4043

Jesús Emmanuel Vidal Cuevas

Universidad Politecnica de Pachuca, 43830 Zempoala, Hidalgo, Mexico

Roles
Co-Author, Conceptualization, Data Curation, Formal Analysis, Investigation, Methodology, Software, Visualization, Writing – Original Draft Preparation
0000-0001-5881-6676

Diego Valdeolmillos

Deep Tech Lab, AIR Institute, Salamanca, Castilla y León, 37003, Spain

Roles
Co-Author, Conceptualization, Funding Acquisition, Investigation, Project Administration, Resources, Supervision, Visualization, Writing – Review & Editing
0000-0001-8424-5395

Albano Carrera

Deep Tech Lab, AIR Institute, Salamanca, Castilla y León, 37003, Spain

Roles
Corresponding Author, Co-Author, Conceptualization, Funding Acquisition, Methodology, Resources, Supervision, Validation, Writing – Original Draft Preparation, Writing – Review & Editing

Downloads over the last 12 months:

Cited by

via Crossref

View citing articles