基于遗传算法优化深度神经网络的站点客流预测  

Passenger flow prediction for stations using genetic algorithmoptimized deep neural networks

作　　者：胡晓伟[1] 吴则洋 卢泓博 王健[1] HU Xiaowei;WU Zeyang;LU Hongbo;WANG Jian(School of Transportation Science and Engineering,Harbin Institute of Technology,Harbin 150090,China)

机构地区：[1]哈尔滨工业大学,交通科学与工程学院,哈尔滨150090

出　　处：《交通运输工程与信息学报》2025年第1期72-84,共13页Journal of Transportation Engineering and Information

基　　金：国家自然科学基金项目(52272332);黑龙江省自然科学基金项目(YQ2021E031);中央高校基本科研业务费专项资金项目(HIT.OCEF.2022026)。

摘　　要：【背景】客流预测是城市轨道交通运营和管理的重要组成,近年来结合多源数据和深度神经网络的精准客流预测受到越来越多的关注。【目标】提升轨道交通站点客流预测的精度,为运营管理提供有效支持。【方法】首先,搭建一种融合多特征的站点客流预测模型,该模型通过卷积神经网络(CNN)提取地铁客流的时空特征,并结合残差单元(ResNet)增强特征提取能力,构建特征传播矩阵挖掘站点间的空间特征,采用长短期记忆网络(LSTM)提取影响因子序列数据的时间特征,在特征融合过程中应用注意力机制突出关键特征。随后,引入遗传算法(GA)对模型进行优化,并采用多层感知器(MLP)修正模型的预测结果误差,提高模型的预测精度。【数据】杭州地铁站点刷卡数据及对应的气象数据、POI数据。【结果】优化ResNet-CNN-LSTM-Attention模型(IO-RCLA)的预测精度最高。相比于RCLA模型,IO-RCLA所有站点预测结果的MAE、RMSE、MAPE分别降低了17.09%、16.09%和8.91%,证明了方法在多站点客流预测中的高精度和有效性。[Background]Passenger flow forecasting is a critical component of urban rail transit operations and management.In recent years,precise passenger-flow predictions based on multi-source data and deep neural networks has garnered increasing attention.[Objective]Improving the accuracy of passenger flow forecast of rail transit station,and providing effective support for operation management.[Methods]A multi-feature passenger-flow prediction model is developed.The model extracts the spatial and temporal characteristics of subway passenger flow through Convolutional Neural Network(CNN),combines with Residual Neural Network(ResNet)to enhance the feature extraction ability,constructs the feature propagation matrix to mine the spatial characteristics between stations,and uses Long Short-Term Memory(LSTM)to extract the temporal characteristics of the impact factor sequence data.In the process of feature fusion,attention mechanism is applied to highlight key features.Subsequently,genetic algorithm(GA)is introduced to optimize the model,and Multilayer Perceptron(MLP)is used to correct the error of the model’s prediction results to improve the prediction accuracy of the model.[Data]Metro station card-swipe,weather,and POI data from Hangzhou.[Results]The optimized ResNet-CNN-LSTM-Attention model(IO-RCLA)achieved the highest prediction accuracy.Compared with the original RCLA model,the IO-RCLA model reduced the MAE,RMSE,and MAPE for all station prediction results by 17.09%,16.09%,and 8.91%,respectively,thus demonstrating its high precision and effectiveness in multi-station passenger flow forecasting.

关 键 词：城市交通 客流预测 组合神经网络 多源数据 遗传算法 

分 类 号：U293.13[交通运输工程—交通运输规划与管理]