【摘要】 为了更加合理地进行地铁车辆调度和制定人员配置方案，并在满足人们出行需求的基础上实现资源利用最大化，对地铁客流量进行准确地短时预测是非常必要的，同时客流预测对地铁站厅空调系统的运行调节也具有重要作用。文章通过对郑州某地铁车站2014年6—7月的进站小时客流量数据进行统计分析，构建季节性差分自回归滑动平均(SARIMA)模型、非线性自回归神经网络(NAR)模型和长短期记忆网络(LSTM)模型，用统计数据进行模型训练并实施预测。通过在工作日客流预测中，发现LSTM模型在MAE、RMSE和R~2上均优于其他模型，拟合系数R~2达到0.981 4,MAE为55.84,均方根误差为88.56;在非工作日客流预测中，LSTM模型同样表现出最好的效果，R~2达到0.981 7;SARIMA模型精度接近LSTM模型。这说明在对具有明显周期性数据预测时，无论是经典的时间序列方法还是先进的深度学习方法预测结果都很好，传统的神经网络因为无法捕捉周期性所以预测效果较差，预测精度相对较低。更多还原

【Abstract】 In order to dispatch the metro vehicles and establish the staffing scheme more reasonably, and maximize the use of resources on the basis of satisfying the people’s travel needs, it’s very necessary to make accurate short-term predictions of the metro passenger flow, meanwhile, the passenger flow prediction is also important for adjusting the operation of the air conditioning system in the metro station hall. This article conducts a statistical analysis of the passenger flow data entered per hour during June and July of 2014 at a metro station in Zhengzhou, constructs a seasonal autoregressive integrated moving average(SARIMA) model, a non-linear autoregressive(NAR) neural network model, and a long and short-term memory network(LSTM) model, trains the models with statistical data, and generates predictions. From the passenger flow predictions of working days, it turns out that the LSTM model is better than the other models in MAE, RMSE and R~2, with the fit coefficients R~2=0.981 4, MAE=55.84 and RMSE=88.56. In the passenger flow prediction of non-working days, the LSTM model also gives the best effect, with R~2=0.981 7; and the accuracy of the SARIMA model is close to that of the LSTM model. It shows that in the prediction of data with obvious periodicity, the prediction result is good both with the classical time series method and with the advanced deep learning approach, while the conventional neural network has a worse effect of prediction and a lower accuracy of prediction because it is incapable of catching the periodicity.更多还原