Scholarlink

Efficient management and accurate prediction of urban traffic flow have become critical challenges in modern smart city governance, where origin–destination (OD) flow prediction plays a vital role in optimizing traffic resource allocation. Existing research indicates that the type distribution, density, and spatial distance of points of interest (POIs) surrounding stations are significant factors influencing passenger travel decisions. However, traditional prediction models, when characterizing the complex associations between POI features and passenger travel purposes, often overlook the dynamic interactive relationship between passenger types and travel intentions, resulting in predictions that inaccurately reflect actual travel behavior patterns. To address this issue, this chapter proposes a bus OD prediction framework based on multisource information fusion. The model comprises three key modules: a feature extraction module based on passenger types and station POIs, a time-series prediction module leveraging historical OD data, and an innovative conservative fusion mechanism. Its key innovations and advantages include the following: firstly, the fine-grained modeling of travel preferences for diverse passenger groups (e.g., commuters, students, and tourists) through a parameterized interaction matrix between passenger types and POIs; secondly, the employment of a multi-head self-attention mechanism for deep feature extraction from spatiotemporal data, enhancing the model’s capacity to capture temporal patterns and long-term dependencies; and, finally, a conservative fusion strategy dynamically allocates model weights and introduces residual connections, effectively integrating prior knowledge from POIs with statistical patterns from historical data, while also incorporating multidimensional external factors such as weather conditions and holiday information, thereby constructing a robust and highly interpretable comprehensive prediction framework. Experimental results demonstrate that the proposed model significantly outperforms traditional methods in multiday prediction tasks, exhibiting superior adaptability, particularly under complex weather conditions and on special dates.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multisource Information Fusion-Based Bus Origin–Destination Prediction Model

  • 章节

  • 发表日期：

    2026年

查看全文
  • Zeyu Wen,
  • Mingquan Shi,
  • Yi Li,
  • Wenyuan Wu

摘要

Efficient management and accurate prediction of urban traffic flow have become critical challenges in modern smart city governance, where origin–destination (OD) flow prediction plays a vital role in optimizing traffic resource allocation. Existing research indicates that the type distribution, density, and spatial distance of points of interest (POIs) surrounding stations are significant factors influencing passenger travel decisions. However, traditional prediction models, when characterizing the complex associations between POI features and passenger travel purposes, often overlook the dynamic interactive relationship between passenger types and travel intentions, resulting in predictions that inaccurately reflect actual travel behavior patterns. To address this issue, this chapter proposes a bus OD prediction framework based on multisource information fusion. The model comprises three key modules: a feature extraction module based on passenger types and station POIs, a time-series prediction module leveraging historical OD data, and an innovative conservative fusion mechanism. Its key innovations and advantages include the following: firstly, the fine-grained modeling of travel preferences for diverse passenger groups (e.g., commuters, students, and tourists) through a parameterized interaction matrix between passenger types and POIs; secondly, the employment of a multi-head self-attention mechanism for deep feature extraction from spatiotemporal data, enhancing the model’s capacity to capture temporal patterns and long-term dependencies; and, finally, a conservative fusion strategy dynamically allocates model weights and introduces residual connections, effectively integrating prior knowledge from POIs with statistical patterns from historical data, while also incorporating multidimensional external factors such as weather conditions and holiday information, thereby constructing a robust and highly interpretable comprehensive prediction framework. Experimental results demonstrate that the proposed model significantly outperforms traditional methods in multiday prediction tasks, exhibiting superior adaptability, particularly under complex weather conditions and on special dates.