An Explainable Federated Learning Framework for Privacy-Aware Modeling of Fuel Consumption and CO2 Emissions in CAVs

초록

Connected and autonomous vehicles (CAVs) raise critical challenges related to energy efficiency, on-road CO2 emission reduction, and data privacy in intelligent transportation systems. This study addresses a data-driven, privacy-aware, and explainable modeling problem for fuel consumption and CO2 emission prediction in CAV environments and does not investigate catalyst materials, catalytic conversion mechanisms, or reaction engineering. Conventional centralized learning paradigms are constrained by large-scale data aggregation, privacy risks, and limited transparency. To address these limitations, we propose an explainable federated learning (XFL) framework for privacy-aware distributed modeling of fuel consumption and CO2 emissions. In the proposed setting, a benchmark vehicular dataset is partitioned into multiple client-specific subsets to simulate decentralized CAV data ownership under controlled and reproducible conditions. Local models are trained on these partitions, while only model parameters or updates are communicated to a central aggregator. In addition, explainable artificial intelligence (XAI) techniques, specifically LIME and SHAP, are integrated to provide transparent interpretation of the relationships between driving behavior, vehicle characteristics, environmental factors, and emission outcomes. From an applied perspective, such explainable and privacy-aware emission modeling can support system-level analysis of CO2 mitigation strategies without requiring centralization of sensitive vehicular data. The proposed federated model selection (FMS) strategy achieved strong predictive performance (R-2 = 0.9172, RMSE = 14.94), outperforming benchmark models under the adopted experimental setting. These results demonstrate the feasibility and effectiveness of the proposed framework as an accurate, privacy-aware, and interpretable proof-of-concept solution for fuel consumption and CO2 emission modeling in a simulated distributed CAV environment.

키워드