考虑稀疏特性的城市轨道交通短时OD时空预测方法

A spatio-temporal prediction method of short-term OD in urban rail transit with sparse data

准确快速地获取城市轨道短时OD需求对城轨交通管理者及时获取乘客出行需求变化、做出科学决策具有重要意义。由于OD数据存在高维度稀疏特性,导致短时OD预测存在预测精度不高和计算效率低的问题。为进一步提升预测准确性和时效性,考虑OD需求的时空特性和矩阵可分解特性,提出一种基于时空分解和动态模式分解的短时OD预测模型(STDMD)。该方法首先采用融合时间序列分解和离散小波变换的时空分解模块,将原始数据分解为多个时空分量,提取时空特征;同时,利用动态模式分解预测模块通过奇异值分解截断数据矩阵的特征值,对数据进行降维去噪,并集成各分量预测结果,实现城市轨道OD的快速、精准预测。为验证模型的有效性,采用北京地铁数据进行案例验证。研究结果表明:STDMD模型具有较高的预测精度和较短的预测时间,在预测精度上,比向量自回归模型、卷积长短期记忆网络和时间正则化矩阵分解模型分别提高了5.0%,15.3%和17.9%;在预测时间上,比向量自回归模型和卷积长短期记忆网络分别缩短了95.7%和37.6%。STDMD模型各模块均可有效提升模型的预测精度。STDMD模型在不同数据集上表现出较强的鲁棒性。STDMD模型可为稀疏条件下的OD分析预测提供新的思路和方法,具有研究意义与现实意义。

Accurate and rapid acquisition of short-term Origin-Destination(OD)demand is critical for urban rail transit managers to catch passenger travel demand changes in a timely manner and make scientific decisions.Due to the high dimensional and sparse characteristics of OD data,short-term OD prediction has issues with low prediction accuracy and slow calculation.To improve the prediction accuracy and timeliness,an OD predication model was proposed based on spatio-temporal decomposition with dynamic mode decomposition(STDMD)by considering the spatio-temporal characteristics and matrix decomposability of OD demand.First,the spatio-temporal decomposition module that incorporates time series decomposition and discrete wavelet transform was utilized to decompose the original data into several spatio-temporal components and capture the space-time aspects.Meanwhile,the eigenvalues of the data matrix were truncated by using singular value decomposition and the dynamic mode decomposition prediction module.The data dimensionality was reduced and denoised,and the prediction results of each component were integrated to realize the fast and accurate prediction of urban rail OD.To verify the validity of the model,the Beijing subway data was used to illustrate the effectiveness of the proposed model.The results show that,the STDMD has higher prediction accuracy and shorter prediction time that improves accuracy by 5.0%,15.3%and 17.9%than vector autoregression,convolution long and short memory networks,and time regularized matrix factorization.It also reduced prediction time by 95.7%and 37.6%as compared to vector autoregression and convolution long and short memory networks,respectively.Each module in STDMD model can effectively improve the prediction accuracy.The STDMD model has strong robustness on a variety of metro data sets.The suggested STDMD provides a new idea and method for OD prediction with sparse data,and it has both research and practical value.