Reliable Reconstruction of Missing Vehicle-Speed Measurements from Multivariate New Energy Vehicle Operational Time Series for Sustainable Intelligent Mobility

Résumé fourni par la source

High-frequency operational records from new energy vehicles (NEVs) are increasingly used to support data-driven sustainable mobility applications, including condition monitoring, energy management, and battery-state estimation. In practice, these records are often incomplete because of sensor faults, communication dropouts, and rapidly changing operating environments, which can distort downstream analyses and reduce the reliability of vehicle-state assessment. This study proposes MDCformer, a multi-period nonstationary modeling framework for reconstructing missing vehicle-speed measurements from multivariate NEV operational time series. MDCformer integrates timestamp-derived temporal descriptors, convolution-enhanced self-attention, and de-stationary attention modulation. The temporal descriptors provide explicit calendar context, the convolutional attention module strengthens local signal consistency before global dependency modeling, and the de-stationary module reintroduces time-varying statistical cues that may be suppressed by normalization. Because the battery electric vehicle (BEV) and fuel cell vehicle (FCV) datasets used in this study cover approximately 18 days and 2.6 days, respectively, the empirical evidence mainly supports daily and short-horizon periodic cues; longer-cycle descriptors are retained as extensible components for longer fleet-level records. Experiments on two real-world NEV datasets show that MDCformer consistently outperforms representative deep-learning baselines under missing rates from 10% to 50%. At a 10% missing rate, compared with the vanilla Transformer baseline, MDCformer reduces root mean square error (RMSE) and mean absolute error (MAE) by 11.35% and 19.00% on the BEV dataset and by 41.41% and 54.83% on the FCV dataset, respectively. Additional scenario-specific tests and downstream state-of-charge prediction further indicate that the reconstructed data preserve more useful temporal structure for sustainable intelligent-transportation analytics. These findings demonstrate the potential of reliable data reconstruction for improving the robustness of intelligent vehicle monitoring and supporting data-driven sustainable transportation applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.