Abstract
Electric vehicles (EVs) are central to low-carbon urban mobility, but range anxiety persists. In real fleet operations, vehicles are rarely discharged to low State-of-Charge (SOC), so the remaining driving range (RDR) labels are incomplete, hindering accurate RDR prediction and analysis of operating conditions. This paper proposes a label completion framework that reconstructs low SOC mileage and a hybrid mileage-factor-oriented residual regressor (MF-CMR) to learn mileage factors under SOC imbalance. Applied to one year of data from eight EVs in Guangzhou, China, the method achieves a mean absolute error of 0.88 and a coefficient of determination of 0.64, yielding completed trip-level RDR labels whose distribution centers around 241.73 km. Using the completed labels, a two-way analysis of variance (ANOVA) with ambient temperature and driving style as factors shows that temperature is the dominant determinant of RDR, while driving style exerts a secondary but substantial effect, with a significant interaction. Together, the label completion framework and the quantified impacts of temperature and driving style enable more reliable RDR estimation from fleet logs, offering a quantitative basis for dispatching policies, charging margins, and eco-driving guidance in EV fleet services involving long distance trips or low SOC deep discharge scenarios.