基於圖神經網路之城市交通狀態預測研究：以三種交通系統為例

Abstract

交通運輸已成為現今社會生活中不可或缺的一部分，各國皆致力於發展大眾運輸與共享式載具，以減少交通壅塞並降低運輸成本。準確預測交通狀態有助於優化運輸規劃，提高交通效率。

近年來，許多研究利用圖神經網路（Graph Neural Network, GNN）來預測交通狀態，GNN 能夠同時捕捉時間與空間上的關聯性，為交通狀態預測帶來重大進展。然而，由於大眾運輸與共享式載具站點之間的關係高度動態，如何有效定義空間關係仍是一大挑戰，但現有的 GNN 模型尚未充分考慮空間關係的多樣性。

本研究提出多圖時空轉換器（Multi Graph Spatial Temporal Transformer，MGSTT），能夠有效捕捉不同空間關係下的特徵，並透過將外部因素轉換為嵌入，以提升模型的預測能力。實驗方面，本研究不僅使用常被用於評估的高速公路交通資料集，還自行收集台灣的大眾運輸與共享式交通資料集進行測試，藉此評估模型在不同類型交通數據上的適用性與廣泛性。

Nowadays, transportation has become a part of daily life in modern society. Many countries develop Transportation has become an essential part of modern society. Many countries are actively developing public transportation and shared mobility solutions to reduce traffic congestion and transportation costs. Accurate traffic state prediction plays a crucial role in optimizing transportation planning and improving efficiency.

In recent years, Graph Neural Networks (GNNs) have been widely used for traffic state prediction due to their ability to capture both spatial and temporal dependencies, leading to significant advancements in this field. However, public transportation and shared mobility systems exhibit highly dynamic relationships between stations, making it challenging to define spatial relationships effectively. Existing GNN-based models have yet to fully consider the diversity of spatial relationships, which limits their predictive performance.

This study proposes the Multi Graph Spatial Temporal Transformer (MGSTT), which effectively captures features under various spatial relationships and enhances predictive performance by embedding external factors. In the experiments, in addition to using commonly adopted highway traffic datasets, this study also collects public transportation and shared mobility datasets from Taiwan to evaluate the model's adaptability and generalizability across different types of traffic data.