摘要
准确预测短时客流对城市轨道交通管理者组织客流、有效分配运力资源具有重要意义。构建一种融合注意力机制和时空图卷积门控递归单元的轨道交通短时客流预测模型(STGGA)。基于旅行时间及OD量构建邻接矩阵,采用图卷积神经网络(GCN)捕获客流空间关系。同时,将注意力机制融入门控递归单元(GRU),提取客流时间相关性。进一步引入外部因素,采用GRU进行特征提取,捕捉外部因素对客流的影响。选取北京地铁客流数据进行案例分析。研究结果表明:与自回归移动平均(ARIMA)、支持向量回归(SVR)以及GRU相比,提出的STGGA在总体预测、单站预测效果方面最优,其精度分别至少提高了22.3%,19.3%与8.0%;加入的外部因素能有效提高STGGA预测性能,使其均方根误差至少降低3.4%;引入的注意力机制能识别客流相关输入时间步,增强模型解释性,有效降低STGGA的均方根误差达16.4%;与基于地理连接关系的模型(STGGA_GC)相比,基于旅行时间与OD量的模型(STGGA_TT和STGGA_OD)在均方根误差方面分别降低了35.5%和24.1%;对不同时段预测效果进行分析:与STGGA_OD相比,STGGA_TT在晚高峰展现出了明显的预测优势。所提出的STGGA能够实现轨道交通短时客流的高精度预测,为管理者分析、控制客流提供一定的数据支撑。
Accurate short-term passenger flow forecast is of great significance for metro managers to organize passenger flow and effectively allocate capacity resources. A spatial-temporal graph convolution gated recurrent unit with attention mechanism model(STGGA) for short-term passenger flow prediction of rail transit was constructed. Based on travel time and OD volume, the adjacency matrix was constructed and the graph convolution network(GCN) was used to capture the spatial relationship of passenger flow. At the same time,attention mechanism was integrated into gated recurrent unit(GRU) to extract the time correlation of passenger flow. Furthermore, external factors were introduced and the GRU was used to capture the impact of external factors on passenger flow. The case study of Beijing Metro passenger flow data showed that compared with autoregressive moving average(ARIMA), support vector regression(SVR) and GRU, the proposed STGGA is the best in overall prediction and single station prediction, and its accuracy increases by at least 22.3%, 19.3%and 8.0%, respectively. The external factors can effectively improve the prediction performance of STGGA and reduce root mean square error by at least 3.4%. The introduced attention mechanism can identify the input time steps related to passenger flow, enhance the interpretability of the model, and effectively reduce the root mean square error of STGGA by 16.4%. Compared with the model based on geographical connection(STGGA_GC),the root mean square error of the model based on travel time and OD volume(STGGA_TT, STGGA_OD) is reduced by 35.5% and 24.1%, respectively. For the prediction effect in different periods, STGGA_TT shows prediction advantages in evening peak as compared to STGGA_OD. The proposed STGGA can realize the highprecision prediction of short-term passenger flow of rail transit, and provide some data support for managers to analyze and control passenger flow.
作者
王雪琴
许心越
伍元凯
刘军
WANG Xueqin;XU Xinyue;WU Yuankai;LIU Jun(State Key Lab of Rail Traffic Control and Safety,Beijing Jiaotong University,Beijing 100044,China;McGill University and IVADO,Montreal,Quebec H3A 0G4,Canada)
出处
《铁道科学与工程学报》
EI
CAS
CSCD
北大核心
2022年第12期3557-3568,共12页
Journal of Railway Science and Engineering
基金
国家自然科学基金资助项目(71871012)
北京市自然科学基金资助项目(9212014)。
关键词
城市轨道交通
注意力机制
图卷积网络
门控递归单元
短时客流预测
urban rail transit
attention mechanism
graph convolution network
gated recurrent unit
short term passenger flow forecast
