GPU平行化之微觀車流模擬

Abstract

隨著城市發展與交通需求的增長，車流所造成的交通壅塞已成為全球城市的普遍現象。傳統的交通模擬方法在面對複雜的大規模道路路網時，難以提供高效且精確的結果，然而，近年來GPU平行運算技術的快速發展為解決這一問題提供了新的契機。為此，本研究提出了一種應用GPU平行運算的微觀車流模擬方法，以細胞自動機為基本架構，並結合Gipps跟車模型，旨在提高模擬的運算效率和精細度。利用CUDA平台及其平行運算技術，本研究設計了三個平行化運作的核函數：細胞參數更新、重繪網格和記錄車輛，完成模擬的主要運算工作。研究結果顯示，本研究所提出的離散化模擬可以有效於邏輯上和數值上與連續性Gipps跟車模型相近，並且能從模擬結果觀察到常見的巨觀現象。相較於CPU單核處理、CPU多核處理，本研究模擬架構以GPU運算能有最高26.32倍、5.63倍的模擬速度表現，且GPU的硬體規格得以反映在模擬速度上。最後，透過離散化空間的模擬架構，本研究克服了傳統車流模擬面對大量車輛計算的效能瓶頸，為車流模擬的應用提供更廣闊的前景。

With the development of urbanization and the growth of transportation demand, traffic congestion has become a common phenomenon in cities around the world. Traditional traffic simulation methods are difficult to provide efficient and accurate results when facing large-scale road networks. However, the rapid development of GPU parallel computing technology in recent years has provided a new opportunity to solve this problem. For this reason, this study proposes a microscopic traffic flow simulation method that applies GPU parallel computing, which uses cellular automata as the framework and combines the Gipps(1981) car-following model, aiming to improve the computational efficiency and precision of the simulation. Using the CUDA platform and its parallel computing technology, this study designs three kernel functions: cell parameter update, grid redraw, and vehicle recording, to complete the main computational work of the simulation parallelly. The research results show that the discretized simulation proposed in this study can be effectively logically and numerically close to the continuous Gipps car-following model, and common macroscopic phenomena can be observed from the simulation results. Compared with single-core CPU processing and multi-core CPU processing, the simulation of this study can achieve up to 26.32 times and 5.63 times the simulation speed performance with GPU computing, and the hardware specifications of the GPU can be reflected in the simulation speed. Finally, through the simulation framework of discretized space, this study overcomes the performance bottleneck of traditional traffic flow simulation facing a large number of vehicle, providing a broader prospect for the application of traffic simulation.