A Temporal Convolutional Network-Based Microscopic Car-Following Model

Accurately modeling car-following decisions is fundamental for understanding traffic flow dynamics and enabling the safe operation of connected and automated vehicles (CAVs) in mixed traffic environments. While theorydriven models offer interpretability, they often struggle to capture the complex, nonlinear stochasticity of human driving behavior. Conversely, prevalent data-driven models, such as those based on Long Short-Term Memory (LSTM) networks, still face challenges in achieving robust long-term prediction accuracy and maintaining low collision rates. To address these limitations, this paper proposes a novel microscopic carfollowing model based on the Temporal Convolutional Network (TCN). The proposed architecture leverages dilated causal convolutions and residual connections to effectively capture both long- and short-term temporal dependencies in vehicle trajectory data, ensuring strict causality in prediction. The model is trained and validated using the HighD trajectory data. A comprehensive Bayesian optimization framework is employed to systematically tune the model's hyperparameters, enhancing its predictive performance and generalization capability. Simulation results demonstrate that the optimized TCN model significantly outperforms several state-of-the-art benchmarks, including the Intelligent Driver Model (IDM), LSTM, and Transformer, in terms of mean squared error on spacing prediction. Furthermore, the proposed model achieves a substantial reduction in collision rate compared to other deep learning approaches, highlighting its superior balance between prediction fidelity and safety in microscopic traffic simulation.