Residual dynamic spatio-temporal graph attention network for enhanced urban traffic origin/destination demand calibration

Purpose The purpose of this paper is to introduce the residual dynamic spatio-temporal graph attention network (RDSTGAT), a novel deep learning architecture designed to address traffic origin-destination (OD) flow calibration. The study aims to accurately estimate the temporal evolution of OD demand matrices using observed sensor data, addressing the limitations of traditional statistical and mathematical models and the lack of datasets for deep learning in this domain. Design/methodology/approach The research employs a microscopic simulation environment (SUMO) as a digital twin to generate synthetic, empirically shaped datasets for training and validation, thereby overcoming the scarcity of real-world ground-truth OD data. The RDSTGAT architecture improves upon the dynamic spatio-temporal graph attention network (DSTGAT) by encapsulating residual logic within modular ResidualGATv2Conv blocks (integrating GATv2, skip connections and normalization) and replacing GRU layers with long short-term memory units to better capture long-range temporal dependencies. The model is trained using a stochastic demand generator to recover the underlying OD flows that generate observed sensor counts. The RDSTGAT model needs to be trained for each urban scenario, as it captures the structural internal capacity relationships of the city. Findings RDSTGAT offers high efficiency in deciphering the spatio-temporal patterns in urban traffic flows and stands out by dynamically assessing the relevance of distinct road network segments and their temporal correlations through advanced graph attention mechanisms. RDSTGAT also reduces the impact of data noise, elevating the accuracy of traffic predictions. Experimental results from different urban scenarios and data-acquisition sensors demonstrated high robustness, with minimal error increase (less than 12%) even when up to 70% of traffic sensors were deactivated. Different prediction models and architectures were analyzed, and the baseline DSTGAT model was outperformed by 4.6%. Research limitations/implications The study currently serves as an algorithmic validation within a controlled digital twin environment driven by synthetic demand profiles. While the demand is empirically shaped using official traffic statistics, external validation against real-world sensor networks and independent OD estimates is reserved for future research phases. In order to capture the internal structural capacity relationships of a city, the RDSTGAT model must be trained separately for each urban scenario. Practical implications The proposed model enables urban planners and authorities to anticipate traffic demand patterns dynamically, which is essential for optimizing traffic signaling, managing lane preferences and designing efficient low-emission zones. The system is designed to function as an offline or near-real-time module within traffic management systems, ingesting data from existing sensor infrastructure. Social implications By enabling more accurate traffic calibration and prediction, this research contributes to the development of sustainable, livable cities by mitigating congestion and its associated environmental impacts. Additionally, accurate flow prediction enhances road safety and facilitates more efficient emergency responses by ensuring the availability of effective evacuation routes. Originality/value This work is original in its specific application of deep learning techniques to the traffic flow calibration problem (OD matrix estimation). Data limitations have been a significant limitation in previous works. The paper presents a reproducible methodology for building calibration models using microscopic simulation and introduces architectural innovations that provide more flexible, robust building blocks for deeper architectures. It provides a fundamental basis for designing urban digital twins.