Proxy Reconstruction Pre-training for Ramp Flow Prediction at Highway Interchanges

Researchers propose STDAE, a spatio-temporal deep learning framework that reconstructs missing ramp flow data at highway interchanges using mainline traffic information. The model matches the performance of systems with actual ramp data, addressing a critical infrastructure gap where real-time ramp detectors are unavailable.

Highway interchanges represent critical chokepoints in transportation networks, yet most lack real-time ramp flow sensors, creating significant blind spots in traffic prediction systems. This research tackles a genuine infrastructure problem by developing a proxy reconstruction approach—using available mainline data to infer missing ramp measurements. The STDAE framework employs a two-stage architecture where decoupled spatial and temporal autoencoders separately extract different feature dimensions, then combine insights during prediction.

The innovation lies in cross-modal reconstruction pretraining, forcing the model to understand fundamental spatio-temporal relationships before making predictions. This approach reflects broader trends in machine learning toward self-supervised and transfer learning methods that reduce reliance on complete labeled datasets. Transportation agencies worldwide face similar detector scarcity issues, making this methodology broadly applicable across jurisdictions.

For traffic management agencies and smart city developers, this work offers practical value. The plug-and-play architecture integrates with existing forecasting pipelines like GWNet without requiring infrastructure overhauls or expensive sensor deployments. Testing across three real-world datasets demonstrates consistent outperformance against thirteen baseline methods, suggesting robustness across different interchange configurations.

The research signals emerging opportunities in infrastructure optimization through AI-driven inference. As cities continue deploying smart transportation systems, methods that maximize value from existing sensor networks—rather than requiring complete sensor overhauls—gain strategic importance. Future development could extend these techniques to other transportation datasets and potentially other infrastructure domains with incomplete sensor coverage.

• →STDAE reconstructs missing ramp flow data using only mainline measurements, eliminating need for expensive ramp detectors at highway interchanges
• →Decoupled spatial-temporal architecture separately processes heterogeneous features for improved prediction accuracy
• →The framework achieves performance comparable to models with actual historical ramp data across three real-world datasets
• →Cross-modal pretraining forces the model to capture intrinsic spatio-temporal relationships before downstream forecasting tasks
• →Plug-and-play design integrates with existing traffic prediction systems, enabling rapid deployment without infrastructure changes