Informer-FDR: A short-term vehicle speed prediction model in car-following scenario based on traffic environment

Drivers’ car-following behaviors on urban roads are influenced by various factors, including pedestrians, cyclists, adjacent vehicles, and roadside parking. However, few models consider these factors’ influence on drivers’ speed selections during car-following, limiting the human-like driving capability of advanced driver assistance systems (ADAS). This paper proposes a vehicle speed prediction model in car-following scenario that considers the influences of the traffic environment. The vehicle speed is predicted using Informer-FDR (Informer with fusion features, dilated causal convolution, and residual connection), which adopts an improved encoder-decoder structure based on the Informer model. Fusing features of traffic environment characteristics and vehicle dynamics parameters enables the dynamic interaction characteristics between drivers and the traffic environment and potential traffic conflicts to be effectively reflected, which enhances the model’s understanding of the complex driving environment. Moreover, the high computational complexity is reduced by using the ProbSparse self-attention mechanism, which will help to address the difficulty of applying Transformer class models to on-board platforms. Totally 3,980 car-following cases were extracted from naturalistic driving data (NDD), vehicle dynamics parameters and traffic environment characteristics in the car-following scenarios were obtained through target detection and ranging algorithm. The optimal feature set was mined using the combined feature selection method. The dilated causal convolution and average pooling layer are introduced to expand the receptive field of the model, enhance global feature extraction, and ensure the causality of temporal predictions. Furthermore, the residual connection was added to the encoder, realizing the direct deep transfer of cross-layer information. Verifications on the test set show that Informer-FDR has the lowest MAE (0.583), MSE (2.942), RMSE (1.715), and the highest speed prediction accuracy (97.76%), spacing gap accuracy (94.27%), acceleration accuracy (95.35%), which outperforms other baseline models in terms of prediction performance. The ablation study confirms the importance of the improved distilling layer module, residual connection module, and fusion features for predictive performance improvement. Additionally, the road-type experiment reveals performance differences of the model on different road types, emphasizing the importance of incorporating traffic environment on urban road.