A data-driven recalibration framework for spatially transferring the vehicle ownership module of an agent-based integrated urban model

Abstract

Traditional transferability approaches can update only a single model at a time and are not designed for large systems. This constrains the spatial transferability of Integrated Urban Models, which comprise multiple micro-models. Recalibrating each micro-model independently with existing techniques would require detailed disaggregate data, undermining the practical objective of enhancing transferability. This study, therefore,  proposes a recalibration framework that leverages aggregate data to update all micro-models within a complex modeling system. The framework combines a Steady-State Genetic Algorithm with a Random Forest surrogate model to identify optimal parameter sets. It was applied to the vehicle ownership module of the STELARS model, which represents vehicle transaction and type choice through six micro-models. The module, originally estimated for the Okanagan region, was transferred to the Greater Vancouver Area using only aggregate data sources. The recalibrated model reproduced observed household vehicle ownership distributions within about four percentage points, and 71.6% of the recalibrated parameters remained within 75% of their original estimates, indicating substantial preservation of the underlying behavioral structure. Parameter deviations reflected consistent behavioral differences between the regions rather than model misspecification. The proposed framework offers a scalable and resource-efficient approach to enhancing the spatial transferability of complex urban models.