Research On Short-term Passenger Flow Prediction Of Intercity Railway Based On Feature Mining

The characteristics of intercity passenger flow present high quality,time concentration,complex spatio-temporal distribution,and obvious peaks under the background of city cluster integration.The sustainable development of city cluster not only needs the support of the construction of intercity railway system,but also requires accurate passenger flow forecasting to provide a basis for the design and adjustment of intercity railway transportation plan.In the design process of passenger transport product,an interactive relationship is presented between passenger flow and transportation plan.Specifically,the transportation plan is scheduled based on forecasting original passenger flow,which is generated by social,economic and cultural exchange needs.And in turn,the transportation plan influences travel choosing behavior,futher affects the attracting passenger flow.Therefore,it is necessary to do two works of passenger flow forecasting in the designing process of transportation plan: one is that the original passenger flow demand forecasting,which provides basic data support for designing transportation plan of intercity railway;the second is attracting passenger flow prediction of the transportation plan,that is the reference for optimization and adjustment of the train transportation plan and the ticket allocation strategy.For this reason,thesis has done the following work:(1)To improve efficiency of prediction model,a feature selection method based on LightGBM(Light gradient boosting machine)is proposed to select effective factors and lay the foundation for construction of prediction models.A feature engineering method is constructed to mine the characteristics of influencing factors,and a feature selector based on LightGBM is established to objectively choose factors which have a great impact on passenger flow.(2)Aiming at the problem of original passenger flow forecasting of intercity railways,an "end-to-end" prediction framework based on feature mining is constructed to capture the characteristics of passenger flow time series,spatial characteristics and external factors.Inspired by the idea of differencing,a novel unit named M-LSTM(Memory in memory of long short-term memory)is proposed based on the improvement of LSTM(Long short-term memory),which is used to extract high-order nonlinear features of passenger flow sequences.And the encoder based on vertical and horizontal time series is constructe to mine the temporal dependence and periodic characteristics of passenger flow.On the basis of mining OD(Origindestination)spatial characteristics,a fusion mechanism based on attention mechanism is established to extract and fuse external factors,OD spatial characteristics and dynamic correlation characteristics of time series.Finally,taking the Chengguan Intercity Railway as an example,the adaptability and effectiveness of the model are verified.(3)To realize attracting passenger flow prediction of intercity railway transportation plan,a gated attention neural network prediction method based on Transformer is established to forecast feedback passenger flow.A time-series dependency mining module based on M-LSTM is proposed to mine the time-series dependencies of suppressed passenger flow.A dynamic dependency mining module based on similarity and gating mechanism is designed to simulate passenger preference in time,date and weather and capture the "time-jumping" feature of passenger travel time selection.The experimental results show that the method can be effectively applied to the feedback passenger flow prediction of the intercity railway.