Short-Term driving speed prediction under consecutive Variable speed Limits: An interpretable deep learning approach...

Existing research on driver behavior under Variable Speed Limits (VSLs) primarily relies on simulations and loop detector-based cross-sectional traffic data, with limited studies using real-world microscopic vehicle trajectory data. This study proposes an interpretable deep learning framework for short-term driving speed prediction under consecutive VSLs control. Using wide-area trajectory data from a 2.2 km segment of the Shanxi Wuyu Freeway with two successive VSL signs, driver behavior was quantitatively analyzed. A Convolutional Neural Network - Bidirectional Long Short-Term Memory (CNN-BiLSTM) model enhanced with a Multi-View Spatio-Temporal Attention Mechanism (MSTAM) was developed to predict short-term speeds and assess the influence of spatiotemporal features on driver responses. Results show that heavy vehicles consistently decelerate under all VSL strategies, while light vehicles display minimal adjustment at 100 km/h and 80 km/h limits but respond more significantly at 60 km/h, with greater inter-driver variability. Drivers in the left lane respond more promptly and decisively than those in the right lane, with shorter response distances and greater speed reductions under all VSL conditions. Additionally, the second VSL sign generally exhibited superior regulatory effectiveness compared to the first VSL sign. Compared to the baseline CNN-BiLSTM, the proposed MSTAM model reduces MAE by 17.2 % and RMSE by 23.5 %. The MSTAM model further captures cognitively consistent spatiotemporal attention patterns, focusing on regulatory zones near VSL signs, sustaining elevated attention in the left lane, and selectively recalling past behaviors to simulate adaptive driver responses. These findings offer a scientific foundation for enhanced VSL deployment and lane-specific speed control strategies.