Research On Short-Term Passenger Flow Forecast Of Urban Rail Transit Based On Deep Learning And Multi-Feature Fusion

The increasing demand of urban rail transportation has brought more challenges to the operation and management of urban rail transportation.In the background of intelligent transportation,in order to ensure the safe,efficient and orderly operation of urban rail transit system,it has become an inevitable trend to use new technologies such as big data and artificial intelligence for intelligent operation and management of urban rail transit.With the rapid development of urban intelligent transportation systems,real-time traffic control and guidance has become a research hotspot in the field of transportation.The key factor to achieve better traffic control and guidance is the accuracy and reliability of passenger flow forecast.Short-term passenger flow forecast is the core of intelligent transportation system.Therefore,this paper researches the short-term forecast method for the outbound and inbound passenger flow of urban rail transit stations,and the main work and results are as follows.(1)Analysis of outbound and inbound passenger flow characteristicsBased on metro AFC data,specific analysis of the outbound and inbound passenger flow characteristics of urban rail transit stations.First,the daily passenger flow distribution characteristics of outbound and inbound passenger flows are analyzed,and the differences in the daily flow distribution characteristics of workdays,weekends and holidays are compared in detail.Secondly,the distribution of outbound and inbound passenger traffic is analyzed in detail over a one-month time scale for three different time types: workdays,weekends and holidays.Third,a detailed analysis is made on the unique operational characteristics of urban rail transit outbound passenger flow.Fourth,the differences in passenger flow between two different types of urban rail stations with high and low traffic volumes are analyzed in detail.(2)Construction of a hybrid forecast architecture with multi-feature fusionBased on the analysis of the outbound and inbound passenger flow characteristics of urban rail transit stations,the coding and quantification method of passenger flow characteristics of hybrid forecast architecture is proposed,and the multi-feature fusion strategy based on Embedding and encoder network is clarified.This paper innovatively integrates Bagging and Transfer learning into a deep learning model for passenger flow forecast.The proposed hybrid forecast architecture is divided into two main components.First is the base predictor with multi-feature fusion using deep learning.The predictor contains a historical traffic processing module and several scalable traffic feature processing modules.It is possible to fuse multiple passenger flow features for passenger flow forecast.We design the neural network structure of each module according to its data characteristics.The above modules are fused and learned through MLP.Second,the group forecast based on Bagging strategy and training optimization using Transfer learning.The Bagging strategy is used to train a number of base predictors for group forecast.Meanwhile,to solve the problem of missing local information in the random sub-training set under the Bagging strategy,the base predictor training process is improved using Transfer learning to enhance the forecast performance of the hybrid forecast architecture.(3)Evaluating the performance of hybrid forecast architecturesValidate the effectiveness of the hybrid forecast architecture and evaluate its forecast performance based on Chengdu Metro AFC data.In this study,three widely used passenger flow forecast models,ARIMA,SVR and LSTM,were selected as benchmark models to validate the forecast performance of the hybrid forecast architecture proposed in this study under three different time types,that is,workdays,weekends and holidays.In addition,the hybrid forecast architecture was evaluated for the forecast performance of two different types of stations,high and low volume stations.Experimental results indicate that the proposed multi-feature hybrid forecast architecture in this study achieves the best forecast performance in all the above-mentioned forecast scenarios.In the experiments,the impact of different passenger flow characteristics on the forecast performance of the model is quantified based on the error evaluation metrics.Finally,the performance improvement of Transfer learning on the base predictor is verified by quantitative analysis of the training process.