Prescriptive analytics for freeway traffic state estimation by multi-source data fusion

Abstract

In the context of freeway traffic state estimation, this study introduces prescriptive analytics—also known as “predict-then-optimize”—for integrating data from Electronic Toll Collection (ETC) systems and traffic sensors. Traditional single-method data fusion techniques are constrained by inherent limitations. For instance, optimization-based methods are generally predicated on prior assumptions that may induce systematic biases, whereas machine learning approaches are frequently criticized for their lack of interpretability and their inability to elucidate underlying traffic mechanisms. To address these limitations, a novel “retrieval and matching” algorithm is proposed that integrates machine learning with optimization. First, the concept of the “state gene” is introduced to encapsulate traffic structural knowledge representing frequently occurring traffic patterns. In the retrieval phase, a heterogeneous graph conventional network is employed to predict potential state genes for a given scenario. In the matching phase, the predicted state genes are utilized to minimize the discrepancy with the current traffic state. This integration not only enhances the interpretability of the estimation process but also endows the optimization component with reverse inference capability through the incorporation of machine learning. Validation using real-world data from the G92 Freeway in Zhejiang, China, demonstrates high accuracy, yielding Mean Absolute Percentage Errors (MAPE) of 1.12 − 1.65 % during peak periods and 1.28 − 1.67 % during off-peak periods.