基于CEEMDAN-IPSO-LSTM的城市轨道交通短时客流预测方法研究  

Short-term passenger flow prediction method of urban rail transit based on CEEMDAN-IPSO-LSTM

作　　者：曾璐 李紫诺[1] 杨杰 许心越 ZENG Lu;LI Zinuo;YANG Jie;XU Xinyue(School of Electrical Engineering and Automation,Jiangxi University of Science and Technology,Ganzhou 341000,China;State Key Lab of Rail Traffic Control and Safety,Beijing Jiaotong University,Beijing 100044,China;Ganjiang Innovation Academy,Chinese Academy of Sciences,Ganzhou 341000,China)

机构地区：[1]江西理工大学电气工程与自动化学院,江西赣州341000 [2]北京交通大学轨道交通控制与安全国家重点实验室,北京100044 [3]中国科学院赣江创新研究院,江西赣州341000

出　　处：《铁道科学与工程学报》2023年第9期3273-3286,共14页Journal of Railway Science and Engineering

基　　金：国家自然科学基金资助项目(62063009);轨道交通控制与安全国家重点实验室(北京交通大学)开放课题基金(RCS2020K005);江西省教育厅科学技术研究项目(GJJ200825);中国科学院赣江创新研究院科研项目(E255J001);江西理工大学高层次人才科研启动项目(205200100428)。

摘　　要：消除客流数据随机噪声和确定神经网络超参数是城市轨道交通短时客流预测组合模型需要解决的关键问题。基于弱化客流数据噪声的自适应噪声完全集成经验模式分解算法(CEEMDAN)将客流时序数据分解为若干个频率和复杂度均不同的固有模态函数分量和剩余分量后,利用引入自适应策略的改进粒子群算法(IPSO)动态求解长短期记忆神经网络(LSTM)超参数的最优值,构建CEEMDAN-IPSO-LSTM组合模型预测城市轨道交通短时客流量。以广州地铁杨箕站自动售检票系统采集的历史进(出)站客流数据为例进行实验,研究结果表明:IPSO算法较PSO算法在基准测试函数Sphere,Sum Squars,Sum of Different Power,Rosenbrock,Rastigrin,Ackley,Griewank和Penalized上的最小值、最大值、平均值和标准差均更接近最佳优化值,CEEMDAN-IPSO-LSTM模型较LSTM模型、CEEMDAN-LSTM模型、CEEMDAN-PSO-LSTM模型的全月全日进(出)站的预测误差评价指标SD,RMSE,MAE和MAPE分别降低了12~40人次(13~35人次)、13~44人次(12~35人次)、6~37人次(12~31人次)和5.08%~46.89%(6.5%~35.1%),R和R2分别提高了0.07%~2.32%(0.86%~3.63%)和0.13%~2.19%(0.67%~1.67%),同时在工作日不同时段和非工作日全日的预测性能均达到最优效果。IPSO算法的收敛速度和参数寻优精度均优于PSO算法,且CEEMDAN-IPSO-LSTM模型可应用于城市轨道交通短时客流量的精确预测,同时可为设计规划线网路线、缓解交通压力、提高乘客出行服务质量等提供基础数据支撑。Eliminating the random noise of passenger flow data and determining the hyper parameters of neural network were the key problems that needs to be solved in the short-term passenger flow prediction combination model of urban rail transit.The passenger flow data were decomposed into intrinsic mode functions and a residual sequence with different frequency and complexity by complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)algorithm,which can reduce the interference of passenger flow data noise on the prediction model.The improved particle swarm optimization(IPSO)algorithm with adaptive strategy was used to dynamically solve the optimal value of the hyper parameters in long-short term memory neural network(LSTM).Then,a CEEMDAN-IPSO-LSTM model was established to predict the short-term passenger flow of urban rail transit.The historical inbound and outbound passenger flow data of Yangji station at Guangzhou Metro was taken as an example.The experimental results showed that the value of minimum,maximum,mean and standard deviation of IPSO in benchmark functions(including Sphere,Sum Squars,Sum of Different Power,Rosenbrock,Rastigrin,Ackley,Griewank and Penalized)were closer to the optimal value than those of PSO.The prediction error evaluation indicators(including SD,RMSE,MAE,and MAPE)decreased by 12~40/person(13~35/person),13~44/person(12~35/person),6~37/person(12~31/person)and 5.08%~46.89%(6.5%~35.1%)respectively,R and R2 increased by 0.07%~2.32%(0.86%~3.63%)and 0.13%~2.19%(0.67%~1.67%)respectively.The proposed model can also achieve favorable prediction results during weekdays and weekends.The research results showed that the convergence speed and optimization accuracy of IPSO are better than those of PSO,and CEEMDAN-IPSO-LSTM can be applied to the accurate prediction of short-term passenger flow of urban rail transit,provide basic data support for designing and planning network routes,relieving traffic pressure and improving passenger travel service quality,etc.

关 键 词：城市轨道交通 短时客流预测 自适应噪声完全集成经验模式分解算法 改进粒子群算法 长短期记忆神经网络 组合模型 CEEMDAN-IPSO-LSTM 

分 类 号：U293.1[交通运输工程—交通运输规划与管理]