基于门控循环神经网络的客运站客流预测

Passenger Flow Prediction of Passenger Station Based on Gated Recurrent Unit Neural Network

分析客运站历史客流数据并进行客流量预测,可以帮助客运人员掌握车站客运调度情况,应对客流突发事件等,对旅客运输组织的优化极为重要。以北京北站全年实际发送客流量为样本数据集,采集客流数的频率为30 min,将全年数据集以8:2比例划分为训练集和测试集。首先对数据集做预处理,通过分析数据的周期性和波动性特点,采用Z-Score方法进行标准化处理。利用循环神经网络(RNN)、长短期记忆网络(LSTM)、门控循环神经网络(GRU)对标准化后的数据集进行训练,并对最后1天的客流量进行预测。通过对3种模型的预测结果进行波形观察与均方根误差(RMSE)比对,基于GRU模型的客流量预测具有更好的波峰响应与更低的误差,更接近原始波形。

Analyzing historical data to predict passenger flow can help the staff to master passenger scheduling situations and deal with passenger flow emergencies,which is vital to optimizing passenger transportation organization.This paper took the annual actual passenger flow of Beijingbei Railway Station as the sample data set with the number of passengers collected every 30 min.The data set was divided into training set and test set with the ratio of 8:2.Firstly,the data set was pre-processed and standardized by the Z-Score method after analyzing the periodicity and fluctuation of the data.Then the standardized data set was trained by recurrent neural network(RNN),long-short term memory(LSTM)and gated recurrent unit(GRU)neural network,and the passenger flow on the last day was predicted.The waveform observation and root mean square error(RMSE)of the three models indicate that the GRU model has a better peak response and a lower error with its waveform closer to the original.