Accurate analysis and prediction of low-temperature energy consumption in pure electric vehicles can provide a reliable reference for energy optimization strategies, thereby alleviating range anxiety. Here, we propose a data-driven energy consumption analysis and prediction approach for real-world electric vehicles in cold conditions. Specifically, the dataset was divided into multiple kinematic segments by the fixed-step intercept method, and principal component analysis was applied on segment parameters, showing the average speed and acceleration time had the greatest impact on energy consumption at −7 °C. Then, a Bayesian optimized XGBoost model, with the two factors above as input, was constructed to predict the cumulative driving and total energy consumption. This method was validated with two different types of pure electric vehicles under different dynamic driving cycles. The results demonstrated that the model could predict low-temperature energy consumption accurately, with all mean relative errors less than 3%.