摘要
地铁系统客流量预测在地铁管理中起着至关重要的作用。由于地铁系统运营策略和市场动态的变化,客流量的时间模式会动态变化,因此利用短期客流数据进行客流量预测更为高效和准确。研究提出一种基于短期训练数据的多条地铁线路客流量预测模型——卷积双线性泊松回归模型,结合潜在因子模型与传统回归模型,采用随机变分贝叶斯法求解优化问题,混合更新模型参数。通过北京地区的GPS信号数据对所提出模型的预测性能进行评估,评估实验结果显示,卷积双线性泊松回归模型采用短期观察数据,相比单一的双线性泊松回归模型和对每个分段分别运行双线性泊松回归模型具有显著优势。此外还揭示集体预测模型相比单独分段模型更不易过拟合。通过不断更新训练数据,模型参数得以实时调整,从而可提供更准确的客流量预测。
Passenger flow prediction in the metro system plays a crucial role in metro management.Due to changes in metro system operation strategies and market dynamics,the time mode of passenger flow will change dynamically,so it is more efficient and accurate to use short-term passenger flow data for passenger flow prediction.This study proposes a passenger flow prediction model for multiple metro lines based on short-term training data-the convolutional bilinear Poisson regression model.It combines the latent factor model with the traditional regression model,and uses the stochastic variational Bayesian method to solve the optimization problem,thus realizing hybrid update of model parameters.This paper evaluates the prediction performance of the proposed model using GPS signal data from the region of Beijing.The evaluation experiment results indicate that the convolutional bilinear Poisson regression model uses short-term observation data,which has significant advantages over the single bilinear Poisson regression model and the approach of running separate bilinear Poisson regression model for each segment.In addition,it reveals that collective prediction models are less prone to overfitting compared to individual segmented models.By continuously updating the training data,the model parameters can be adjusted in real time,providing more accurate passenger flow predictions.
作者
窦道飞
DOU Daofei(Signal & Communication Research Institute,China Academy of Railway Sciences Corporation Limited,Beijing 100081,China)
出处
《中国铁路》
北大核心
2025年第5期125-132,共8页
China Railway
基金
中国铁道科学研究院集团有限公司科研开发基金项目(2023YJ113)。
关键词
地铁
客流预测
卷积双线性泊松回归模型
潜在因子
变分贝叶斯法
metro
passenger flow prediction
convolutional bilinear Poisson regression model
latent factor
variational Bayesian method
