Research On The Application Of Passenger Demand Forecast Driven By Large Trajectory Data

With the development of the national economy,floating vehicles have gradually become an important part of the transportation system,which provides residents with a convenient and quick way of travel.However,passenger distribution has high randomness and flexibility at different times and in different regions,so it is difficult for drivers to predict the distribution of the passenger.As a result,the imbalance between the supply and demand of floating vehicles has been caused,which was reflected in many aspects of our life.For example,although the number of floating vehicles in the city can meet the requirements of the city scale for the floating vehicles ownership,residents still need to wait a long time to get a car,while the no-load rate of floating vehicle remains high.This phenomenon not only wastes the user's valuable time,but also increases the cost of operating floating vehicle and reduces the incomes of drivers.In order to solve above problems,this thesis designs and implements a passenger distribution forecasting algorithm to guide and resolve the contradiction between vehicle supply and demand,which can improve the operating efficiency of floating vehicles.Firstly,based on the Spark framework,this thesis design and implement a parallel algorithm for passenger distribution estimation that can improve the efficiency of obtaining the passenger distribution information.This algorithm reduces the time of acquiring the passenger distribution information by applying the Spark parallel framework,and estimates total passenger demands in a certain period of time through the two steps of gridding and demand estimation.Based on information extracted from GPS data,the research analyzes respectively the time-varying characteristics of passenger distribution under different conditions.which provides a basis for predicting passenger distribution.Secondly,because non-homogeneous Poisson model has disadvantages in predicting passenger distribution,a novel model named optimized non-homogeneous Poisson model(ONHPM)is proposed in this thesis.ONHPM filters historical data by using correlation coefficient method,which removes the data that has a large different from the passengers' distribution during the target period.In addition,the smoothing index method is used to weight the historical data according to the degree of correlation between the data,which achieves the smooth transition between data in different day.Based on ONHPM and Kalman filter model(KFM),a combined named KalmanONPHM is further established.Kalman-ONPHM that takes the advantage of ONPHM and KFM takes into account the real-time passenger distribution and weather conditions to achieve an accurate prediction of passenger distribution.Finally,based on the GPS data of Zhengzhou city,the performance of KFONPHM is analyzed from several aspects and the effectiveness of KF-ONPHM and other predicting model is furtherly compared.Experimental results show that KFONPHM has advantages in symmetric mean percentage error,maximum relative error and maximum absolute error.