An Ecological Assessment Method for Lane-Changing Overtaking Behavior Based on FL-XGBoost

Lane-changing overtaking is a continuous and complex process that has a significant impact on energy consumption. Traditional eco-driving research primarily focuses on common factors, such as overall acceleration frequency, while overlooking the heterogeneous effects on driving behaviors across different periods of driving process. This study develops an improved extreme gradient boosting model based on the local loss function (FL-XGBoost model) to account for temporal heterogeneity and analyze the stage-dependent impacts of driving behaviors on energy consumption. Considering the dynamic characteristics of lane-changing overtaking, the entire process is divided into four phases, and corresponding stage-specific driving behavior datasets are constructed. To achieve dimensionality reduction and extract key information from the feature space, a hybrid feature selection strategy integrating random forest (RF) and ant colony optimization (ACO) is adopted. Furthermore, aiming at the model bias caused by class imbalance in the dataset, the focal loss function is introduced as the optimization objective in place of the traditional cross-entropy loss, thereby enhancing the robustness and generalization performance of the model. Results show that the proposed FL-XGBoost model outperforms other baseline models such as support vector machine (SVM). Compared with the unmodified XGBoost model, the FL-XGBoost model achieves improvement of 3% in accuracy and 5.1% in the F1-score. To further reveal the causal relationships between influencing factors and energy consumption, SHAP (Shapley Additive Explanations) is adopted for model interpretability analysis. The results indicate that the proportion of acceleration duration during the two lateral lane-changing phases exerts the most significant impact on the eco-driving performance of the entire lane-changing overtaking process. Nonlinear coupling effects exist among eco-driving features across multiple phases of the process.