基于SARIMA模型的广珠城际铁路客流量预测

为实现铁路车站发送客流量的短期预测，研究预测步长对短期客流预测效果的影响，分析了广珠城际铁路车站发送客流的特征和变化规律，结合客流特征及季节性差分自回归滑动平均模型（seasonal autoregressive integrated moving average，SARIMA）的适用性，构建了SARIMA客流预测模型，利用Python软件中的Statsmodels模块完成了SARIMA客流模型的精细化调参，以广州南站、小榄站的发送客流量为例验证了模型的有效性. 结果表明，SARIMA预测模型可以较好地适用于不同数量等级的客流预测，其预测精度随预测步长的增加而降低. 预测步长为1时，广州南站、小榄站、珠海站客流预测平均绝对百分比误差（mean absolute percentage error，MAPE）值分别为3.97%，5.83%，5.43%；预测步长增加为2时，各车站客流预测误差显著增加，广州南站、小榄站、珠海站客流预测误差MAPE值分别为5.31%，6.79%，7.62%；预测步长大于2时，预测误差基本保持稳定. 将SARIMA模型预测效果与随机森林（random forest， RF）、支持向量机（support vector machine， SVM）、梯度提升算法（gradient boosting， GB）、K最近邻算法（K-nearest neighbor， KNN）模型或方法的预测效果进行对比，预测步长为1时，SARIMA模型预测效果略优于其余4种模型，5种预测模型预测精度差距较小；预测步长大于1时，RF、SVM、GB、KNN模型预测误差随预测步长显著增加，预测误差为SARIMA模型的数倍. SARIMA模型在客流时间序列的多步预测方面具有较大的优势. 

Passenger Flow Prediction for Guangzhou-Zhuhai Intercity Railway Based on SARIMA Model

To achieve the short-term prediction on the railway passenger flow and analyze the influence of prediction step on prediction accuracy, firstly, the characteristics and variation of passenger flow for Guangzhou-Zhuhai intercity railway were analyzed. Then, considering the passenger flow characteristics, a prediction model based on the seasonal autoregressive integrated moving average (SARIMA) was built with the Statsmodels module in Python. Next, the model performance was validated on different prediction steps. The conclusion shows that when the prediction step is 1, the mean absolute percentage error (MAPE) for Guangzhou South station, Xiaolan station and Zhuhai station is 3.97%, 5.83%, and 5.43%, respectively; when the prediction step increases to 2, the MAPE shows an increase trend, which is 5.31%, 6.79%, and 7.62% for Guangzhou South station, Xiaolan station and Zhuhai station, respectively; when the prediction step exceeds 2, the MAPE is stable. In addition, comparative results with other passenger flow prediction methods, i.e., random forest (RF), support vector machine (SVM), gradient boosting (GB), and K-nearest neighbor (KNN) demonstrate that when the prediction step is 1, the SARIMA model performs slightly better; when the prediction step exceeds 2, the MAPE of RF, SVM, GB, and KNN increases dramatically, amounting several times that of the SARIMA model. Finally, the experiment results show that the SARIMA model can achieve a better performance than other models in terms of the multi-step prediction for passenger flow time series.