Short-term Passenger Flow Prediction At Public Transit Stations Based On Combined Deep Learning Model

In the urban public transport system of China,as an independent unit,bus stations are the foundation of public transportation organization.The number and the variation of short-term passenger flows can reflect the actual demands of passengers and the instability of passengers under objective external factors.How to use scientific methods to predict short-term passenger flow at public transit stations and obtain accurate travel rules and characteristics of passengers,in order to help public transport companies enterprises scientifically manage and facilitate passengers’ travel has become a hot issue.This article is based on HN City’s bus IC card data,vehicle GPS data and bus line data to obtain bus station passenger flow data.And then combining weather data,POI data and station passenger flow data,it puts forward the basis of how to construct the characteristics of shorttime passenger flow data and the selection of forecast objects in short-time passenger flow prediction after analyzing the influence of time factors,weather factors and site land factors on station passenger flow.Bi-directional Long Short-Term Memory Network(Bi LSTM)was used as the basis for the construction of the model.Particle Swarm Optimization(PSO)algorithm was introduced to optimize the problem about the selection of parameters such as the number of hidden layer neurons in the basic model,the learning rate,and the number of iterations to build the PSO-Bi LSTM model.Complete Ensemble Empirical Mode Decomposition With Adaptive Noise(CEEMDAN)was introduced to decompose the short-time passenger flow data as for the problem of instability characteristics of short-time passenger flow data of the site,which can establish and the CEEMDAN-PSO-Bi LSTM composite model as the short-time passenger flow prediction model.Three error evaluation methods(RMSE,MSE and MAE)are used to construct an evaluation index system to estimate the prediction effect of different models.The results of the study found that when three stations with relatively concentrated passenger flow were used as the prediction object,a detailed analysis of the prediction results of Hualian Building Station shows that the Bi LSTM model had improved prediction accuracy compared with the LSTM model,and RMSE,MAE,and MSE decreased by 1.23～2.58%,0.86～3.45%,and 0.00～4.17%,respectively;the RMSE,MAE,and MSE of the CEEMDANPSO-Bi LSTM combined model are reduced by 6.62～18.01%,0.90～11.30%,13.04～30.77%,respectively;compared with the Bi LSTM base model,the RMSE,MAE,and MSE of the PSOBi LSTM optimized model are respectively reduced 2.08～10.81%,0.90～2.86%,4.76～18.18%;compared with CEEMDAN-Bi LSTM decomposition model,RMSE,MAE,MSE decreased by0.70～6.38%,0.90～2.78%,0.00～10.00%,respectively.When sites with different land use properties are taken as prediction objects,the overall evaluation index range of the combined model prediction is RMSE[7.0%,16.8%]、MSE[0.5%,2.8%] and MAE[3.3%,13.0%],which are higher than the predictions of other models accuracy.Judging from the interval comparison of the average value of the short-term passenger flow prediction results of stations with different land use properties,the average range of the evaluation indicators of the combined model is RMSE[8.5%,15.9%] 、MSE[0.8%,2.5%] and MAE[4.4%,11.8%],the average range of has different degrees of decline compared with other models.The research shows that: The prediction accuracy of the prediction model shows that the combined model has relatively higher prediction accuracy than LSTM,GRU and Bi LSTM models,indicating the combined model has outstanding effectiveness.The point of view of the applicability of the predictive performance of the predictive model shows that the combined model has a relatively low average range of evaluation indicators for the prediction results of sites with different land properties compared with LSTM,Bi LSTM and PSO-Bi LSTM model,which demonstrates the combined model has universal applicability.