We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
DGA-NCDE: Dual-Graph Attention Neural Controlled Differential Equation for Accurate Urban Rail Passenger Flow Prediction.
- Authors
Lin Bai; Hong Dai; Shuang Wang
- Abstract
Urban rail transit systems constitute a crucial component of modern transportation engineering, where accurate predictions of station-level passenger flow are essential for optimizing the efficiency of rail transit operations. Traditional spatio-temporal models, such as graph-based models and recurrent neural networks, typically use static approaches. These methods often require high-dimensional parameter spaces to capture the complex dynamics of urban rail systems, resulting in high computational costs and low model efficiency. Moreover, it is difficult for traditional graph learning models to dynamically adjust their focus when capturing changes and subtle patterns, finally failing to fully utilize heterogeneous information between nodes. As a result, these limitations prevent models from adapting to the evolving characteristics of urban rail passenger flows and identifying fine-grained patterns with precision. To address these challenges, this study presents the DualGraph Attention-Neural Controlled Differential Equation (DGA-NCDE) model, which combines local geographical associations and global semantic associations through a novel Dual-Graph Attention (DGA) Module. The DGA Module is seamlessly integrated into the Neural Controlled Differential Equation (NCDE) framework, enabling the model to dynamically adapt to urban rail passenger flow patterns. To validate the effectiveness of the DGA-NCDE model, comprehensive experiments were conducted on two largescale public datasets, HZMetro and SHMetro. When compared to the best baseline models, DGA-NCDE reduces MAPE by 3.44% on HZMetro and by 2.27% on SHMetro, while using just 1.12% of the parameters. In addition, the DGA-NCDE model significantly cuts the training time by approximately 51.31%, and the inference time by about 87.23%, demonstrating the significant improvement in both accuracy and efficiency of the DGA-NCDE model.
- Subjects
URBAN transit systems; DIFFERENTIAL equations; RECURRENT neural networks; TRANSPORTATION engineering; PASSENGERS; URBANIZATION
- Publication
Engineering Letters, 2024, Vol 32, Issue 7, p1486
- ISSN
1816-093X
- Publication type
Article