Search

CN-117112931-B - Urban rail transit passenger flow prediction method based on multi-mode information fusion

CN117112931BCN 117112931 BCN117112931 BCN 117112931BCN-117112931-B

Abstract

The invention discloses a city track traffic passenger flow prediction method based on multi-mode information fusion, which comprises the steps of firstly obtaining feature vectors of city interest points, then constructing a track traffic station directed graph based on the track traffic station directed graph and the position relation between the track traffic station and the interest points, extracting feature vectors of the track traffic station, crawling social media texts of each time period, extracting public travel interest feature vectors from the social media texts, then fusing the public travel interest feature vectors, the track traffic station feature vectors and historical passenger flow feature vectors based on a multi-head cross attention model, and then predicting the passenger flow feature vectors of the next time period by adopting an LTSM network. According to the method, urban interest point features and social media text features are integrated to predict urban rail transit passenger flows, and prediction accuracy is improved.

Inventors

  • ZHANG DINGKAI
  • HE JIFENG
  • WANG XIAOLING
  • JIN BO

Assignees

  • 同济大学

Dates

Publication Date
20260505
Application Date
20230913

Claims (4)

  1. 1. The urban rail transit passenger flow prediction method based on multi-mode information fusion is characterized by comprising the following steps of: The method comprises the steps of S1, obtaining city interest point data, wherein the city interest point data comprises names, categories, coordinates and areas of interest points, respectively encoding the names and the categories of the interest points to obtain name vectors and category vectors, normalizing the coordinate vectors and the areas of the interest points, and then splicing the name vectors, the category vectors, the normalized coordinate vectors and the normalized areas of the interest points to obtain feature vectors of the interest points, wherein the number of the interest points in a city is recorded as D, the feature vectors of the D-th interest point are recorded as S d , d=1, 2, and D; s2, acquiring a characteristic vector of a rail transit station, which comprises the following steps: S2.1, acquiring position data of each rail transit Station n , wherein the position data comprise Station numbers, geographic coordinates, belonging line codes, previous Station distances and next Station distances, and n=1, 2; then, a track traffic undirected graph is built according to the position data of each Station, wherein each Station is taken as a node, the sides between stations are determined according to the connection relation between the stations, and the weight of the sides between adjacent stations is the running time of the train between the two adjacent stations; S2.2, for each interest point POI d , calculating the geographic position distance between the interest point POI d and each rail transit station, screening out K 1 geographic nearest neighbors with the nearest distance, setting the value of K 1 according to actual conditions, and acquiring the public from each geographic nearest neighbor station The length of travel time required to point of interest POI d (n k ,d),n k represents the original sequence number of the kth geographically nearest neighbor site, k=1, 2,., K 1 ; for each other Station n′ ,n′＝1,2,…,N,n′≠n k , calculating the track traffic undirected graph and each geographical nearest neighbor Station The weight of the upper edge of each path and the travel time length of the path, if a certain station is adjacent to a certain geographic nearest station No path exists between the two, the travel time length is infinity, and then the station is Screening out the minimum value in all the travel time lengths, and marking the serial number of the corresponding geographic nearest neighbor station as The corresponding travel time length is recorded as Calculating and obtaining the journey time length from the Station n′ to the point of interest POI d Correlation r (n ', d) =1/time (n', d); S2.3, for each POI d , screening the front K 2 rail transit stations with the minimum journey time length, wherein the front K 2 rail transit stations are used as the journey nearest neighbor stations of the POI d , and the value of K 2 is set according to actual conditions; S2.4, for each track traffic Station n , acquiring a point of interest set omega n taking Station n as a travel nearest neighbor Station, splicing feature vectors of all points of interest in the point of interest set omega n , aligning the spliced feature vectors to a preset dimension gamma 1 through orthogonal mapping to obtain a feature vector S n , forming a relevance vector by the relevance of all points of interest in the point of interest set omega n and Station n , and aligning the relevance vector to a preset dimension gamma 2 through orthogonal mapping to obtain a feature vector R n , thereby forming a feature vector G n ＝[Line n ,Num n ,S n ,R n of Station n , wherein Line n represents the Line number of Station n of track traffic, and Num n represents the Station number of Station n of track traffic Station; s3, extracting public travel city interest point features from texts of social media, wherein the method comprises the following steps of: s3.1, setting keywords and time period duration according to actual needs, crawling a plurality of groups of text information from social media, wherein each group of text information comprises text information of T continuous time periods, the text information comprises user release content, comments and labels, and preprocessing each text information by adopting a preset preprocessing method to obtain text data; S3.2, extracting the topics from each text data collected in the step S3.1, screening out the text data of the traffic class topics, and deleting the text data of the rest topics; S3.3, integrating the text data into one combined text data according to time periods for each group of text data subjected to the traffic class theme screening, and forming a text data sequence from the combined text data of T time periods; s3.4, mining travel interest content features from each piece of merged text data, analyzing emotion features, and then constructing a travel interest feature matrix, wherein the method comprises the following steps of: S3.4.1 extracting word vectors f (w i )＝[t i,1 t i,2 … t i,H , i=1, 2,) of a plurality of central words w i in each merged text data by using a CBoW model for each merged text data sequence, wherein M represents the number of central words in the text data, and H represents the dimension of the word vectors; S3.4.2 carrying out emotion feature analysis on the word vector f (w i ) of each central word in each text data sequence to obtain emotion polarity e i , wherein when the emotion feature analysis result is positive, emotion polarity e i =1, when the emotion feature analysis result is neutral, emotion polarity e i =0, and when the emotion feature analysis result is negative, emotion polarity e i = -1; S3.4.3 combining word vectors and emotion polarities of each central word in each text data sequence to obtain a feature vector F i ＝[t i,1 t i,2 … t i,H e i , and constructing each feature vector F i as a row vector to obtain a travel interest feature matrix F with the size of M× (H+1); S3.5, constructing an interest point feature matrix S with the size of D multiplied by J according to the interest point feature vector S d as a row vector, wherein J represents the dimension of the interest point feature vector S d , and then transforming the travel interest feature matrix F' of each piece of merged text data into a vector space of the interest point feature matrix S by adopting orthogonal mapping to obtain a transformed travel interest feature matrix F; S3.6, calculating the pearson correlation coefficient PCC (F, S) of the travel interest feature matrix F and the interest point feature matrix S for each merged text data in each text data sequence; s3.7, for each merged text data in each text data sequence, expanding the travel interest feature matrix F into a vector Expanding the interest point feature matrix S into vectors Then splice to obtain the public travel interest feature vector of each merged text data Further obtaining a public travel interest feature sequence corresponding to the text data sequence; S4, setting a rail transit station to be predicted according to actual needs, counting passenger flow data of the station in a time period corresponding to each combined text data and a time period next to the text data sequence for each text data sequence, and extracting to obtain passenger flow feature vectors; s5: constructing a passenger flow prediction model comprising a multi-head cross attention model and an LSTM network, wherein: The multi-head cross attention model is used for fusing station feature vectors of a rail transit station to be predicted, public travel interest feature vectors and passenger flow feature vectors of corresponding time periods to obtain fusion feature vectors, and then inputting a fusion feature vector sequence formed by T fusion feature vectors obtained by a text data sequence into an LSTM network; The LSTM network is used for extracting features from the received fusion feature vector sequence and predicting to obtain a passenger flow feature vector of the next time period; The public travel interest feature vectors of the combined text data in each text data sequence obtained in the step S3 are combined with the station feature vector of the rail transit station to be predicted and the passenger flow feature vector of the corresponding time period to be used as input, the passenger flow feature vector of the station of the next time period of the text data sequence is used as a label, and the passenger flow prediction model is trained to obtain a trained passenger flow prediction model; And S6, crawling text information of social media of the last T time periods for the rail transit station to be predicted, extracting a passenger flow feature vector sequence according to the same method in the step S3, and then combining the station feature vector of the rail transit station to be predicted and the passenger flow feature vector of the corresponding time period, and inputting the passenger flow feature vector into the passenger flow prediction model trained in the step S5 to obtain the predicted passenger flow feature vector of the next time period.
  2. 2. The urban rail transit passenger flow prediction method according to claim 1, wherein the specific method for extracting the theme in the step S3.2 is as follows: S3.2.1, obtaining word frequency vectors corresponding to each piece of text data obtained in the step S3.1 based on a word bag model, and then forming a word frequency matrix by taking the word frequency vectors of the piece of text data as row vectors; s3.2.2 training a word bag model by using an LDA model, and then generating a topic-word distribution probability matrix and a document-topic distribution probability matrix from the trained word bag model; s3.2.3 extracting the topic of the text data by using the generated topic-word distribution probability matrix and the document-topic distribution probability matrix.
  3. 3. The urban rail transit passenger flow prediction method according to claim 1, wherein the emotion feature analysis in step S3.4.2 is performed by using word frequency-inverse document frequency, and the specific calculation formula is: TF-IDF(f(w i ),d)＝TF(f(w i ),d)×IDF(f(w i ),d) Where TF (f (w i ), d) represents the frequency of occurrence of the word vector f (w i ) in the text data d, IDF (f (w i ), d) represents the inverse document frequency of the word vector f (w i ) in the text data d.
  4. 4. The urban rail transit passenger flow prediction method according to claim 1, wherein the passenger flow characteristics in the step S4 include three types of characteristics, wherein the time characteristics include hours, weeks, seasons, and years, the in-out characteristics include in-coming passenger volume and out-coming passenger volume, and the characteristic statistics in-coming passenger flow statistics and out-coming passenger flow statistics characteristics include average, maximum, minimum, and variance, respectively, passenger flow characteristic vector= [ hours, weeks, seasons, years, average of in-coming passenger volume, maximum in-coming passenger volume, minimum in-coming passenger volume, variance of in-coming passenger volume, average of out-coming passenger volume, maximum in-coming passenger volume, minimum in-coming passenger volume, and variance of in-coming passenger volume ].

Description

Urban rail transit passenger flow prediction method based on multi-mode information fusion Technical Field The invention belongs to the technical field of urban rail transit passenger flows, and particularly relates to a method for predicting urban rail transit passenger flows based on multi-mode information fusion. Background Traffic flow prediction is an important problem in the field of transportation, and can help traffic management departments to better plan and schedule traffic resources, improve traffic efficiency and improve the service level of traffic systems. Along with the urgent demands of rapid promotion of economical strength and urban development, mass urban rail transit construction is carried out in a plurality of cities, and accurate prediction of passenger flow of urban rail transit becomes an important problem of rail transit management. The prediction can help traffic managers predict future passenger flow, so that the transport capacity is adjusted, train shifts are optimized, the traveling demands of passengers are met, meanwhile, the basis can be provided for future track traffic line and station layout, and the service level of a traffic system is continuously improved. The existing mainstream rail traffic flow prediction method is mostly based on historical traffic flow data, and modeling and prediction are carried out by combining other relevant space-time data such as weather, time, holidays and the like through time sequence analysis, machine learning and deep learning methods. However, the accuracy and robustness of the time series analysis method are low, and the adaptability to nonlinear and non-stationary data is poor. The machine learning and deep learning methods need a large amount of historical data for training, have high requirements on the quality and quantity of data, and meanwhile need feature engineering aiming at different modes and different scenes, so that how to perform cross-mode and cross-domain transfer learning related to rail transit is not effectively solved. Along with the diversification of information acquisition channels in the traffic field, a multi-mode information base composed of data, texts, images, videos and the like is formed, and the multi-mode information is used for predicting traffic passenger flow together, so that the accuracy of prediction can be improved, and the robustness of prediction can be enhanced. The existing multi-modal information related to urban rail transit comprises public travel card swiping information, passenger position information, distribution information of lines and stations, vehicle running information and the like, and urban environment information, social media information and the like closely related to the rail transit. Due to the different sources of information, some of the information is text-type, some of the information is digital-type, and some of the information is graphic or image-type, so that multi-mode cross-domain mixed information is formed, and the multi-mode cross-domain information describes the traffic trip intention, attitude and behavior of people from different angles, different methods and different dimensions. However, how to integrate and analyze the information, accurately and reliably predict urban rail transit passenger flow, and further research is needed. Disclosure of Invention The invention aims to overcome the defects of the prior art, provides a urban rail transit passenger flow prediction method based on multi-mode information fusion, and integrates urban interest point characteristics and social media text characteristics to predict urban rail transit passenger flow, so that prediction accuracy is improved. In order to achieve the aim of the invention, the urban rail transit passenger flow prediction method based on multi-mode information fusion comprises the following steps: The method comprises the steps of S1, obtaining city interest point data, wherein the city interest point data comprises names, categories, coordinates and areas of interest points, respectively encoding the names and the categories of the interest points to obtain name vectors and category vectors, normalizing the coordinate vectors and the areas of the interest points, and then splicing the name vectors, the category vectors, the normalized coordinate vectors and the normalized areas of the interest points to obtain feature vectors of the interest points, wherein the number of the interest points in a city is recorded as D, the feature vectors of the D-th interest point are recorded as S d, d=1, 2, and D; s2, acquiring a characteristic vector of a rail transit station, which comprises the following steps: S2.1, acquiring position data of each rail transit Station n, wherein the position data comprise Station numbers, geographic coordinates, belonging line codes, previous Station distances and next Station distances, and n=1, 2; then, a track traffic undirected graph is built according to the position