Vehicles’ risky lane-changing (LC) maneuver has significant impact on road traffic safety. As an innovation compared with the posterior LC risk prediction methods proposed in previous studies, this study develops a pre-emptive LC risk level prediction (P-LRLP) method, which is able to estimate the crash risk level of an LC event in advance before the LC car completes the LC maneuver. The basic concept of this method is to apply a machine learning classifier to predict the LC risk level based on cars’ key space-series features at the beginning of the LC event. To boost the prediction performance, an innovative resampling method, namely ENN-SMOTE-Tomek Link (EST), and an advanced machine learning classifier, namely LightGBM, are proposed and employed in the development of the P-LRLP method. Meanwhile, an algorithm which can measure the stability of the selected key features in terms of the randomness and size of training samples is developed to evaluate the feature selection methods. A digitalized vehicles’ trajectory dataset, the Next Generation Simulation (NGSIM) is used for method validation. The validation results manifest that the EST can achieve satisfactory resampling performance while Random Forest (RF), as an embedded FS method, achieves remarkable performance on both stability of selected features and prediction of risk level. The results also show that the LC risk level can be most accurately predicted when the LC car moves to the position where the distance between the longitudinal center line of the LC car and the marking line separating the two lanes equals 1.5ft. As an innovative LC risk level prediction technique, the P-LRLP method could be integrated with advanced driver-assistance system (ADAS) and vehicle-to-vehicle (V2V) communication to remedy potential risky LC maneuver in the future.

车辆的危险变道（LC）操作对道路交通安全有重大影响。作为与先前研究中提出的后LC风险预测方法相比的一项创新，本研究开发了先发性LC风险水平预测（P-LRLP）方法，该方法能够提前估计LC事件的碰撞风险水平LC车完成LC机动。该方法的基本概念是应用机器学习分类器，根据LC事件开始时汽车的关键空间序列特征预测LC风险水平。为了提高预测性能，提出了一种创新的重采样方法，即ENN-SMOTE-Tomek Link（EST），以及一种先进的机器学习分类器，即LightGBM，并将其用于P-LRLP方法的开发。与此同时，开发了一种可以根据训练样本的随机性和大小来测量所选关键特征的稳定性的算法，以评估特征选择方法。数字化车辆的轨迹数据集，下一代仿真（NGSIM）用于方法验证。验证结果表明，EST可以实现令人满意的重采样性能，而作为嵌入式FS方法的随机森林（RF）在所选特征的稳定性和风险级别的预测上均具有出色的性能。结果还表明，当LC车移至LC车的纵向中心线与分隔两条车道的标记线之间的距离等于1.5 ft的位置时，可以最准确地预测LC风险等级。作为一种创新的LC风险水平预测技术，