Abstract:  Vehicle driving speed characteristics in extralong freeway tunnels are analyzed using radar networking technologybased realtime traffic parameter acquisition method. The highprecision timeseries speed data of all samples from the entire tunnel are extracted. The tunnel is then divided into three sections: entrance, driving, and exit, and a shortterm prediction model for detecting vehicle driving speed in extralong freeway tunnels is constructed using a timeseries Transformer framework. The main findings of the study are as follows: (1) The lowest average vehicle driving speed is discretely distributed at the entrance section. (2) The statistical value of 85% quantile vehicle driving speed in each lane shows a decreasing trend along the same road section. (3) The driving speed of a passenger car at different sections exhibits a decreasing law, and there is a considerable difference between the entrance and exit sections. (4) Although the freight cars slow down at the entrance section, the actual driving speed of passenger cars is higher than the maximum speed limit of the main lane. (5) The maximum accuracy of the speed prediction model is 97.82%, and the average absolute error is 1.67 km/h. These results demonstrate that dark adaptation has a significant impact on the driver′s speed control behavior. The speed limit of the main lane is not fully applicable on the extralong tunnel environment, and thus, guidance must be provided for different vehicles and lanes in advance to avoid accidents. The timeseries Transformer model established in this paper is suitable for shortterm speed prediction of all lanes in extralong freeway tunnels.

Key words: extralong freeway tunnel, vehicle speed prediction, Transformer framework, prediction model