Inference of Transit Alighting From Automatic Boarding Count Data: A Double DQN Clustering

The availability of both boarding and alighting counts is crucial for transit planning, as it allows for the estimation of Origin-Destination patterns that reflect demand. However, automatic data collection methods often provide only boarding counts. This paper investigates the use of transit symmetry to address the lack of alighting counts by analyzing cluster-level symmetry, which captures the dynamics of trip exchanges between nearby stops to achieve balanced ingoing and outgoing flows. An optimization problem is formulated to maximize symmetry across clusters, evaluating the effectiveness of clustering based on stop location using k-means. A novel deep reinforcement learning algorithm utilizing the Double Deep Q-Network framework is proposed to identify optimal clusters, demonstrating an 8% improvement in symmetry optimization over k-means. Notably, k-means clusters still achieve an average symmetry level of 62% when accounting for geographical proximity and walking distance limits. Through the proposed clusters, transit planners can better understand passenger flow dynamics and predict alighting counts with incomplete data.