基于经验模态分解与长短时记忆神经网络的短时地铁客流预测模型

为降低样本噪声对客流预测模型的干扰, 结合深度学习理论, 提出了一种基于经验模态分解与长短时记忆神经网络的短时地铁客流预测模型; 将预测过程分为3个阶段, 第1阶段预处理原始地铁刷卡数据, 构建进(出)站客流时间序列, 运用经验模态分解法将时间序列转化为一系列本征模函数及残差, 第2阶段利用偏自相关函数确定长短时记忆神经网络的输入变量, 第3阶段基于深度学习库Keras, 完成长短时记忆神经网络的搭建、训练及预测; 以上海地铁2号线人民广场站客流数据验证了模型的有效性。计算结果表明: 与代表性的预测模型(差分自回归移动平均模型、支持向量机、经验模态分解与反向传播神经网络、长短时记忆神经网络)相比, 经验模态分解与长短时记忆神经网络预测模型分别将工作日高峰、平峰、全日的进(出)站客流预测精度分别至少提升了2.1%(2.5%)、2.7%(3.5%)、2.7%(3.4%), 将非工作日全日的进(出)站客流预测精度至少提升了3.3%(3.5%), 说明经验模态分解与长短时记忆神经网络的组合是一种预测短时地铁客流的有效模型; 当预测步长由5 min逐渐增加至30 min时, 工作日高峰、平峰和全日进(出)站客流的平均绝对百分比预测误差分别由14.8%(13.9%)、16.8%(17.4%)和16.6%(17.0%)逐渐降低至7.0%(6.2%)、8.3%(7.5%)和8.1%(7.4%), 说明该方法预测误差与预测步长呈负相关。

Short-term metro passenger flow prediction based on EMD-LSTM

To weaken the interference of sample noise to the prediction model of passenger flow, a short-term metro passenger flow prediction model was proposed based on the deep-learning theory, empirical mode decomposition(EMD) and long short-term memory neural network(LSTMNN). The prediction process was divided into three stages.In the first stage, the raw automatic fare collection(AFC) data were preprocessed, and the tap-in(tap-out) passenger flow time series were constructed and decomposed into a series of intrinsic mode functions(IMFs) and a residues by the EMD. In the second stage, the input variables of the LSTMNN were determined by the partial autocorrelation function(PACF). In the third stage, the LSTMNN was developed, trained and predicted through the deep learning library Keras. A case study of Shanghai People's Square Station on metro line 2 was conducted to validate the model performance. Calculation result shows that, compared to the representative prediction models(differential autoregressive integrated moving average model, support vector machine, empirical mode decomposition and back propagation neural network, and LSTMNN), the EMD-LSTM prediction model increases the weekdays' tap-in(tap-out) passenger flow prediction accuracy of peak hour, off-peak hour, and full-day by at least 2.1%(2.5%), 2.7%(3.5%), and 2.7%(3.4%), respectively, and also increases the weekends' tap-in(tap-out) passenger flow prediction accuracy of full-day by at least 3.3%(3.5%). Thus, the EMD-LSTM is effective to predict the short-term metro passenger flow. When the forecasting step gradually increases from 5 minutes to 30 minutes, the weekdays' tap-in(tap-out)average absolute percentage prediction errors of peak hour, off-peak hour, and full-day gradually decreases from 14.8%(13.9%), 16.8%(17.4%), and 16.6%(17.0%) to 7.0%(6.2%), 8.3%(7.5%), and 8.1%(7.4%), respectively. Therefore, the forecasting error of EMD-LSTM is negatively correlated with the forecasting step length.