Delay Aware Multi-Scale Gated Time Network for Non-Stationary Car-Following Prediction

Accurate car-following (CF) behavior prediction is crucial for autonomous driving systems, especially under non-stationary traffic conditions with high nonlinearity and uncertainty. Existing data-driven CF models often use attention mechanisms but inadequately model driver reaction times due to hard-coded causal masks and fail to capture cross-scale dynamics due to single-scale modeling limitations. To address these issues, we propose the Delay aware Multi-Scale Gated Temporal Network (DMGTimeNet). First, the Delay Aware Causal Attention (DACA) module is introduced. This module encodes the driver’s physiological reaction delay as learnable parameters within a causal mask, ensuring the prediction adheres to strict temporal causality. Second, a Hierarchical Gated Fusion (HGF) module is designed. It constructs parallel feature sequences across multiple time scales and adaptively integrates the multi-scale outputs, effectively handling the coupling between instantaneous maneuvers and long-term traffic variations. Finally, the Mirage Search Optimization (MSO) framework is implemented for global hyperparameter optimization. Validation on the HighD naturalistic driving dataset demonstrates that DMGTimeNet significantly outperforms existing baseline models in both prediction accuracy and overall model performance metrics. The results verify the effectiveness of the proposed causal attention, multi-scale fusion mechanisms, and optimization strategy in handling non-stationary CF behaviors, supporting the development of safer and more reliable prediction modules for autonomous driving.