基于LSTM深度神经网络的高速铁路短期客流预测研究

Short term passenger flow prediction of high speed railway based on LSTM deep neural network

为验证长短期记忆网络(long short-term memory,LSTM)模型对于高速铁路短期客流预测的有效性,本文以京广高速铁路2010年1月至2015年12月车站发送客流实绩数据为基础,分析了车站发送客流的特征和变化规律,针对客流时间序列变化特征和LSTM模型数据输入要求,完成了客流时间序列的标准化处理和重构,构建了基于LSTM的高速铁路客流预测模型,利用多层网格搜索完成了模型的精细化调参,分析了模型参数对于预测效果的影响,并将LSTM模型与其它客流预测模型的预测效果进行了比较.结果表明:相较于对比模型,LSTM客流模型预精度较高,郴州西站,衡阳东站,韶关站单步预测平均绝对百分比误差(mean absolute percentage error,MAPE)值分别为7.36%,7.33%,8.03%;模型中隐含层数,神经元个数,输入步长对客流预测精度影响较大,适当增加模型隐含层数及神经元数量可以提高模型收敛速度及预测效果;受客流周期性影响,模型输入步长为7时各车站客流预测误差最小.

To demonstrate the effectiveness of the long short-term memory(LSTM) deep neural network model on short-term passenger flow prediction of high-speed railways,the paper presents the characteristics of the departing passenger flow in different stations based on the real-record passenger flow data of Beijing-Guangzhou high speed railway,from January,2010 to December,2015.The passenger dataset is standardized and framed for the LSTM model,considering the expectation input format of LSTM layers and the characteristics of the data.LSTM model is fitted with tuning and regulating all the parameters necessary in the model.Then the fitted LSTM model is applied to forecast the short-term departing passenger flow of Beijing-Guangzhou high speed railway.The influence of important parameters in the LSTM model on the prediction accuracy is analyzed,and the comparison with other representative passenger flow forecast models is conducted.The results show that the LSTM model can achieve a better performance compared to other models.One-step forward passenger flow prediction errors valued by mean absolute percentage error(MAPE) are 7.36%,7.33%,8.03% respectively for Chenzhou west station,Hengyang east station and Shaoguan station.The parameters in the LSTM model such as the number of network layers,the number of neurons and the input format of the reshaped passenger dataset have a great influence on the prediction accuracy.The convergence speed and prediction accuracy can be improved as the number of network layers and neurons increase reasonably while LSTM model can performance better as the input size of passenger flow is 7,affected by the periodic characteristics of passenger flow.