Analysis of influencing factors of lane change prediction with data missing

The accurate behavior prediction of autonomous vehicles relied heavily on precise data supplied by various sensors. In spite of this, data missing due to a communication barrier or equipment failure is an unavoidable issue in the process of data acquisition, which significantly impacts lane-changing (LC) prediction. To solve this issue, the influence of data missing on lane change prediction was explored from the several aspects such as missing pattern, missing proportion and data interpolation. The different degrees of data missing in three patterns, namely block missing, temporally correlated missing and random missing, were examined. Missing values are repaired and LC behavior is predicted based on machine/deep learning models, and hyperparameters are optimized using genetic algorithms. By analyzing the factors that affect the model's generalization ability, the difference between the best data repair result and the best prediction result was further explored. The results show that: (1) the proposed methods can effectively solve the problems of data missing repair and prediction accuracy, especially the KNNImputer for repair and the Transformer for prediction; (2) temporally correlated missing is inferior to random missing for filling in errors, but superior for prediction performances. This indicates that the selection of repair methods should take into account not only the best data imputation, but also the best prediction performance; and (3) there are significant differences in the stability of prediction models under different data missing patterns and the CNN-BiLSTM-SA has achieved relatively good performance