Deep Learning–Based Spatial Translation of Traffic Prediction Using Newell’s Theory

Abstract

Spatiotemporal traffic flow forecasting using deep learning (DL) models typically relies on convolutional or graph-convolutional filters combined with recurrent neural networks to capture both spatial and temporal dependencies utilizing network-level traffic data. However, these models face limitations in their adaptability to new locations due to their reliance on specific detector configurations and localized traffic characteristics. Consequently, transferring these models to new configurations requires retraining the model with data from that configuration, which presents a significant challenge when data are unavailable. These models, unless retrained, yield predictions that inaccurately reflect the traffic dynamics of the new location, as both the model inputs and parameters remain unchanged. To address this challenge, we propose a novel feature transformation that is consistent with traffic flow theory in traffic prediction models. This transformation incorporates Newell’s estimators for uncongested and/or congested traffic flow states at target locations, enabling the models to capture broader dynamics beyond local characteristics. Empirical validation of our methodology is conducted using traffic data from distinct freeway sections. Our results highlight the enhanced performance of the proposed method in traffic flow forecasting along freeway sections. A key advantage of our framework lies in its ability to be transferred to locations where training data may be unavailable. This transferability is facilitated by accounting for spatial dependencies based on distances of the transfer location from the detector stations and various traffic parameters. However, Newell’s assumption of uniform traffic conditions throughout the segment may not hold, especially in the presence of entrance and exit ramps along freeway segments. This discrepancy poses a challenge in accurately applying Newell’s transformations in such scenarios, potentially affecting the model’s predictive capability. Moreover, a spatial sensitivity analysis could not be conducted due to data limitations, highlighting the necessity for further research utilizing simulated data to comprehensively explore this aspect.