MMCAformer: A New Approach to Traffic Prediction

Proactive traffic management requires accurate speed prediction. A new study introduces MMCAformer (Macro-Micro Cross-Attention Transformer), a model that combines macroscopic traffic flow data with microscopic data related to individual driving behavior, obtained from connected vehicles.

Model Details

MMCAformer uses self-attention mechanisms to analyze the intrinsic dependencies in macroscopic traffic data and cross-attention to capture the spatiotemporal interplay between macroscopic traffic status and microscopic driving behavior. The model was optimized with a Student-t negative log-likelihood loss function to provide point-wise speed predictions and estimate uncertainty.

Experimental Results

Tests on four Florida freeways demonstrated that MMCAformer outperforms baseline models. The introduction of microscopic driving behavior data improved prediction accuracy (RMSE, MAE, and MAPE reduced by 9.0%, 6.9%, and 10.2%, respectively) and reduced model prediction uncertainty (mean predictive intervals decreased by 10.1-24.0%).

Hard braking and acceleration frequencies emerged as the most influential features. The improvements are more pronounced under congested, low-speed traffic conditions.