Energy Consumption Estimation for Electric Vehicles via Driving Style-Aware Optimization

Driven by the dual-carbon strategy, accurate personalized energy consumption estimation of battery electric vehicles (BEVs) is a core demand of intelligent transportation. Differentiated long-term driving styles lead to 20–30% energy consumption differences under identical road environments, which cannot be captured by traditional prediction models. Current mainstream physical and data-driven modeling methods suffer prominent technical limitations: they only extract instantaneous driving features without mining stable latent driving preferences and adopt fixed-parameter transfer strategies that fail to eliminate distribution bias for long-tail users with few historical trips. To address these bottlenecks, this paper first establishes standardized mathematical definitions for road networks, trajectories, and trip datasets, quantifies seven energy-sensitive driving events via statistical thresholds, and builds a multi-dimensional driving feature system. On this basis, a four-module Driving-Style-Aware Adaptive Transfer (DST-AdaTrans) learning framework is designed. The framework leverages CNN-Transformer to denoise trajectory sequences and extract steady driving preferences, utilizes multi-head attention to infer scenario-matched personalized driving behaviors, introduces dynamic weight fusion to balance scene and individual features, and realizes cross-user knowledge sharing through preference-aware dynamic fusion and cross-user knowledge transfer. Comparative experiments on four real traffic datasets demonstrate that the proposed method reaches an overall determination coefficient R2 > 0.95 with average MAPE only 2.25%. For sparse long-tail drivers, the prediction error is reduced by more than 28% compared with optimal baselines. Ablation studies further verify the collaborative optimization effect of each module, offering complete theoretical and algorithmic support for accurate personalized EV energy prediction.