Context-Aware Long-Range Transportation Flow Prediction for Supporting Urban Mobility Informatics

Improving urban mobility efficiency is a challenging endeavor due to increasing populations and at-capacity transportation infrastructure systems. Transportation flow prediction is a critical component of urban mobility informatics to enable efficient utilization of transportation infrastructure systems for enhanced mobility efficiency. However, existing methods for transportation flow prediction are limited in enabling long-range flow prediction for proactive mobility improvement decision making. This limitation primarily stems from the challenges in effective and efficient modeling of transportation flow dynamics from the temporal, spatial, and contextual modeling perspectives. To address these critical challenges, this paper proposes a new context-aware long-range transportation flow prediction method, which consists of three components: (1) a proposed similarity-based shared dilated convolution method to enhance the efficiency and effectiveness of spatial flow dynamics modeling, (2) a proposed multi-context feature embedding method to extract, embed, and integrate multiple urban contextual features for effective contextual modeling, and (3) an attention transformer-based encoder-decoder to ensure the effectiveness of temporal dynamics modeling for long-range flow prediction. A set of model selection, hyperparameter tuning, and baseline comparison experiments were conducted to evaluate and benchmark the performance of the proposed method, using real-world transportation flow data in New York City. The experimental results show that, at the flow zone pair level, it achieved an hourly mean absolute percentage error of 2.59% and 0.95%, and an hourly mean absolute error of 0.14 and 0.02 for the 2019 and 2023 data, respectively.