摘要
高速公路出入口的短时交通流量预测是智能交通系统优化与交通规划的重要依据。为对高速公路出入口具有时空异质性的短时流量进行预测,以甘肃省高速公路为研究对象,基于ETC及人工收费数据,结合时空相关性特征,构建一种融合时空聚类与混合模型的短时流量预测方法,旨在提升交通管理效率与资源分配合理性。首先基于甘肃省181个高速公路出入口的历史数据采用结合地理坐标与客流量的加权K均值聚类(K-means Clustering)算法,将出入口划分为四个特征迥异的类别。然后,分别采用改进参数的自回归积分移动平均模型(Autoregressive Integrated Moving Average,ARIMA)与神经网络模型进行预测,其参数设定分别通过赤池信息量准则(Akaike Information Criterion,AIC)以及网格搜索进行参数寻优。最终,基于各模型预测误差的方差,以近似6∶4的权重对两模型的预测结果进行加权融合,形成混合预测模型。实验仿真结果表明:优化后的ARIMA模型在测试集上的平均绝对百分比误差(Mean Absolute Percentage Error,MAPE)为11.91%,神经网络模型MAPE为15.70%,通过加权融合模型结果的MAPE降至10.70%。研究证实:通过时空聚类的混合模型能有效处理交通流的时空异质性,基于方差倒数加权的混合模型兼具ARIMA模型的线性拟合能力与神经网络的非线性捕捉优势,可为高速公路动态管理提供更可靠的技术支持。
Short-term traffic flow forecasting at highway entrances and exits was an important foundation for optimizing intelligent transportation systems and urban transportation planning.To address the short-term traffic flow forecasting considering spatiotemporal heterogeneity at highway entrances and exits,this study took the highways in Gansu Province as the research object.Based on ETC and manual toll data,combined with spatiotemporal correlation characteristics,a short-term traffic flow prediction method integrating spatiotemporal clustering and a hybrid model was developed,aiming to improve traffic management efficiency and resource allocation rationality.First,historical data from 181 highway entrances and exits in Gansu Province were analyzed using a weighted K-means clustering algorithm incorporating geographic coordinates and passenger flow volumes,dividing the entrances and exits into four distinct categories.Then,an improved parameter Autoregressive Integrated Moving Average(ARIMA)model and a neural network model were employed,with their parameters optimized using the Akaike Information Criterion(AIC)and grid search,respectively.Finally,based on the variance of the prediction errors from each model,the prediction results of the two models were weighted and fused at an approximate 6∶4 ratio to form a hybrid prediction model.Experimental simulation results show that the optimized ARIMA model achieves a Mean Absolute Percentage Error(MAPE)of 11.91% on the test set,while the neural network model achieves a MAPE of 15.70%.The MAPE decreases to 10.70% after weighting and fusing the model results.This study confirms that the“clustering-hybrid modeling-weighted fusion”technical pathway effectively addresses the spatiotemporal heterogeneity of traffic flow.The hybrid model,weighted by the inverse of variance,combines the linear fitting capability of the ARIMA model with the non-linear capture advantage of the neural network,and can provide more reliable technical support for dynamic management of highways.
作者
陈龙
谷远利
CHEN Long;GU Yuan-li(Big Data Laboratory,Beijing Jiaotong University,Beijing 100044,China)
出处
《广州航海学院学报》
2025年第3期44-51,共8页
Journal of Guangzhou Maritime University
基金
北京市自然基金项目(2015A030313819)
北京市科技计划项目(Z121100000312101)。
