A public transport flow prediction method combining urban interest points and spatio-temporal causality

Abstract

The application provides a public traffic flow prediction method combining city interest points and space-time causal relationship, and belongs to the technical field of traffic flow prediction.The method of the application firstly performs periodic preprocessing on public traffic flow data;secondly, information extraction is performed on city interest point geographic coordinate data and classification of station groups is completed;finally, a deep learning model based on space-time causal relationship is used to make prediction on public traffic station flow in a future time window, wherein different station groups will use different trainable parameters.The application solves the problem that the traditional traffic flow prediction cannot effectively cope with the complex space-time causal relationship of city traffic, and further integrates external factors of interest points into the construction and training stage of the deep learning model, is suitable for city public traffic flow prediction, and has good prediction accuracy and small error.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of traffic flow prediction, and relates to a public transport flow prediction method combining city interest points and space-time causal relationship. BACKGROUND

[0002] Public transport tools such as buses and subways are important components of urban transportation, and will bring many problems in urban planning, traffic control and public safety. The fluctuation of public transport passenger flow brings diversified operation challenges, especially the huge contrast between passenger travel demand and bus supply in a short period of time, which leads to waste of traffic resources or crowd congestion. Flow prediction is an important part of the transportation system, which can be used to fine-tune travel behavior, reduce passenger congestion and improve the service quality of the transportation system. The prediction result can be used to support operation planning, station passenger flow regulation planning and other transportation system management. The station can be fully prepared before the peak time which is difficult to grasp, or the departure time of the bus or subway can be dynamically adjusted, so as to realize higher operation efficiency and lower energy consumption through accurate prediction of passenger flow.

[0003] The core of public transport flow prediction is to solve the space-time causal problem affected by various factors in the city. Unlike traditional numerical sequence prediction, inferring the traffic flow under a specific space-time state involves various factors, including spatial causal relationship, temporal causal relationship and external factors: (1) Spatial causal relationship. The emphasis on spatial causal relationship in the application scenario studied in the application involves adjacent relationship in non-Euclidean domain. The public transport route is fixed for a long time, and a station is directly affected by other stations on the same line. Two adjacent stations may have no traffic interaction because there is no operating route covering the two stations. Similarly, the flow of distant stations on the same line may also be homogeneous.(2) Temporal causal relationship. The flow of a bus station is affected by the time interval. Due to the regular time arrangement of passengers and the fixed traffic line, the subjective randomness is lower compared with personal driving. The passenger flow change maintains a stable trend in a short period of time, and the peak time of bus or subway station is similar on most days.(3) External factors. This refers to those factors that are not within the public transport system but will affect people's travel plans to some extent. In a complex modern city, numerous unpredictable influencing factors such as impulsive travel bring great uncertainty to the prediction task. From this point of view, modeling of some important external factors helps to perform the prediction task.

[0004] Currently, there are many studies on urban public transport flow prediction. Initially, people often give predictions based on real experience without sufficient mathematical argument. Soon, early researchers began to use mathematical statistics methods to predict future time series data. However, pure mathematical methods do not consider the spatial causal relationship between public transport stations. Today, public transport infrastructure is highly information-based, and the passenger flow involved in the prediction task can be monitored in a timely manner, and big data can be mined for deep patterns. Accordingly, deep learning-based methods are widely studied for traffic prediction. Some researchers use recurrent neural networks and their variants to capture time series information, but all spatial causal relationships are lost. Obviously, it is far from enough to apply time series methods to spatio-temporal traffic prediction. Researchers have begun to build joint prediction models to consider both spatial and temporal information. However, these methods cannot take advantage of interest points to process structured flow data in public transport without fully considering the spatio-temporal causal relationship.

[0005] In order to solve the above problems, there is currently a need for a traffic prediction method that adapts to the complex spatio-temporal environment of the city. SUMMARY

[0006] The problem to be solved by the present application is that in the urban public transport system, relying solely on experience, mathematical statistics, and sequence prediction cannot effectively capture the passenger flow variation in the complex spatio-temporal environment. And the existing flow prediction method based on machine learning or deep learning does not consider the influencing factors of the flow outside the public transport system. In order to solve the above problems, the present application provides a public transport flow prediction method combining urban interest points and spatio-temporal causal relationship.

[0007] The public transport flow prediction method combining urban interest points and spatio-temporal causal relationship provided by the present application comprises the following steps:

[0008] Step 1) Data preprocessing. Obtain the passenger in-out flow of each time period at the subway and bus stations in the city. In order to capture the periodicity of public transport flow, the previous time slices of the prediction target time window and the same time window in the past few days are spliced to obtain a sequence containing the flow information of the same time period in the near future.

[0009] Step 2) The area function points around the public transport stations are designed as interest points, an interest point extractor is designed to establish the connection between the area function points and the public transport stations, and a station attribute table is calculated.

[0010] The interest point extractor obtains a list of all interest points in the city, traverses the list of interest points, and calculates a site attribute table. Each row of the site attribute table records the attribute value of each site for each type of interest point. The attribute value is calculated as follows: the initial attribute value of each site is 0. For each interest point, find the sites within a certain range of straight-line distance, and add the influence weight of the interest point to the attribute value of the site. After traversal, normalize the values in the site attribute table by row. The influence weights of different types of interest points are set in advance.

[0011] The interest point extractor divides all public transportation sites into several site groups, and sites in the same site group are close to the same type of interest point. Let q be the number of interest point types. From the site attribute table, there are q+1 types of site groups, the i-th type of site group represents sites that are severely affected by the i-th type of interest point, 1≤i≤q, and the q+1-th type of site group represents sites that are not affected by any interest point; q is a positive integer. Thus, a reasonable task classification is made for public transportation flow prediction from the perspective of city interest points.

[0012] Step 3) A deep learning model is established based on spatio-temporal causality for predicting public transportation site flow.

[0013] The model includes P stacked spatio-temporal causal graph convolution units, and the output of each unit will be used as the input of the next unit. The spliced sequence obtained in step 1 is input into the first spatio-temporal causal graph convolution unit, and the output of the last spatio-temporal causal graph convolution unit is calculated by two one-dimensional convolutions in series to generate the flow of each public transportation site in the prediction time window; P>2.

[0014] The spatio-temporal causal graph convolution unit includes a time series causal convolution layer and a spatial causal graph convolution layer; the time series causal convolution layer is composed of two parallel time series causal convolutions; the input data is first processed by two parallel time series causal convolutions, and the outputs of the two time series causal convolutions are fused by Hadamard product and then input into the spatial causal graph convolution; q+1 sets of model parameters of the time series causal convolution layer are set, corresponding to q+1 types of site groups, and the output weights of the site corresponding to the first q sets of models are determined according to the site attribute table. The output weight of the q+1 set of models is set to 1 for the q+1 type of site and to 0 for other types of sites; the input data is processed by the time series causal convolution layer with q+1 sets of model parameters, and the output data corresponding to each site is weighted and summed according to the output weight.

[0015] The model is trained, and the trained model is used to predict the flow of public transportation sites.

[0016] In the step 1, the preprocessing splices the (m1+1) time windows of the m1 time slices of the target time window of the day and the same (m1+1) time windows in the recent m2 days to obtain the passenger flow in and out of all stations in the same time period in the recent period.

[0017] In the step 3, the inflation convolution technology is adopted in the time sequence causal convolution layer, and the two time sequence causal convolutions adopt the sigmoid activation function and the tanh hyperbolic tangent activation function respectively.

[0018] In the step 3, in the spatial causal graph convolution layer, the station features are extracted according to the station adjacency matrix A; the station adjacency matrix A is calculated using the historical passenger flow between public transport stations, and the number of public transport stations is N, A is an N*N matrix, and the value of the [i,j] position in the matrix represents the historical total flow of the station i to the station j.

[0019] Compared with the prior art, the advantages and positive effects of the present application are that:

[0020] (1) The method fully considers the complex space-time characteristics of public transport flow, and cross-captures the time causality and spatial causality in the data in the deep learning model; the present application fully considers the characteristics of public transport in the real operation, places the spatial causality in the non-Euclidean domain, and uses the adaptive graph convolutional neural network to learn the structured information; on the time causality, the present application fully utilizes the periodicity of the time sequence data, and establishes the time causal convolution to model the periodic change rule of the public transport flow on different dates.

[0021] (2) The method uses the interest points in the city to guide the parameter learning of the deep neural network for the public transport stations near different interest points, effectively introduces the external factors of the public transport system into the prediction task, and improves the prediction accuracy. According to the actual test verification, the method of the present application obtains excellent prediction results, and the prediction curve and the actual in and out of the station passenger flow curve are basically consistent.

[0022] (3) The method solves the problem that the traditional traffic flow prediction cannot effectively cope with the complex space-time causal relationship of urban traffic, and also integrates the external factors of the traffic system, interest points, into the construction and training stage of the deep learning model, which is suitable for urban public transport flow prediction, has good prediction accuracy and small error. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is the overall implementation diagram of the public transport flow prediction method combining the city interest points and the space-time causal relationship of the present application;

[0024] Figure 2 is a time sequence causal convolution schematic diagram.

[0025] Figure 3 is a schematic diagram of the test results of the present application on a certain urban public transportation data set. DETAILED DESCRIPTION

[0026] The technical solutions of the present application will be described below in conjunction with the drawings and examples.

[0027] The present application provides a public transportation flow prediction method combining urban interest points and spatio-temporal causal relationships. First, the public transportation flow data is periodically preprocessed. Second, the urban interest point geographic coordinate data is information extracted and the classification of the station group is completed. Finally, a deep learning model based on spatio-temporal causal relationships is used to make predictions for future time windows, where different station groups will use different trainable parameters. The application of the present application solves the problem that traditional traffic flow prediction cannot effectively deal with complex spatio-temporal causal relationships in urban traffic. It also integrates external factors of interest points into the construction and training stage of the deep learning model. The present application is suitable for urban public transportation flow prediction and has good prediction accuracy and small error.

[0028] The public transportation flow prediction realized by the embodiment of the present application has an overall process as shown in Figure 1 , which includes the following three steps, which will be described below.

[0029] Step 1) Data preprocessing.

[0030] In the present application, the public transportation system including buses and subways involves fixed stations and fixed routes. From the physical world to the non-Euclidean space, the station is regarded as a node in the directed graph, and the directed graph is represented as where the vertex set represents the public transportation station; the edge set represents the traffic route; and the adjacency matrix A∈R N×N is an N×N real number field matrix representing the correlation between N stations. The present application uses the historical passenger flow between public transportation stations to calculate the adjacency matrix A. Please note that due to the asymmetry of the flow between public transportation stations, an asymmetric adjacency matrix will be generated. In particular, the diagonal line of the adjacency matrix is 0. R represents the real number set. N is a positive integer representing the number of stations.

[0031] There are two kinds of passenger flow concerned in the present application, which are the in / out flow of the station and the transition flow between stations. The in / out flow of the station is the input data and the prediction target, denoted as X. The transition flow between stations can be represented by a tuple tr={s ori ,t ori ,s des ,t des}, which describes the passenger at t oriEntering station s during the time period ori and in the time period t des Complete to s des The journey. Without involving the input sequence and prediction target, the inter-station transition flow represents the correlation between stations, which is a suitable source for generating the adjacency matrix A. Let the graph at time step t... There is a passenger flow matrix X (t) ∈R N Passenger flow matrix X (t) Record the traffic flow at each bus stop at time step t, with each line representing one stop. Given a graph... Given its historical S-step passenger flow matrix and S-time steps, this invention needs to learn a matrix that can predict the passenger flow X in the next time step. (t+1) The function f is represented as follows:

[0032]

[0033] In order to capture the periodic patterns in public transportation traffic, this invention splices together the first few time slices of the predicted target time window with the same time windows in recent days to obtain a sequence containing traffic information for the same period in recent times.

[0034] In this embodiment of the invention, the input sequence is set to be a sequence containing 8 time windows, collecting data from the prediction time point to 7 days prior, with each time window containing 8 time steps. For example... Figure 1 As shown, when predicting traffic flow at 8 PM on October 21st, the public transportation traffic data obtained includes 316 stations, covering 8 days from October 14th to the predicted time on October 21st. Each day corresponds to a time window, for a total of 8 time windows. The traffic data has 4 collection points per hour, and each time window contains 8 time steps, meaning data from 8 collection points within a two-hour period is obtained. 8 PM on October 21st falls within the 8th collection point of the 8th time window. Therefore, traffic data for each station within the same 8 time windows starting from October 14th is obtained, specifically the traffic data from 6 PM to 8 PM on October 14th, 6 PM to 8 PM on October 15th, and so on, up to 6 PM to 7:45 PM on October 21st. This data is then stitched together as shown in the image. Figure 1 The passenger flow matrix shown is 316*63. This preprocessing ensures that the temporal causal convolutional layers (TC layers) of the deep learning model can obtain information about the time periodicity of the data.

[0035] Step 2) Construct an interest point extractor to obtain a site feature matrix or site attribute table.

[0036] In the application scenario of the present application, the point of interest is the most concerned factor. The point of interest can be a shop, a school, a hospital, etc., which represents the urban functional attribute of a block. Meanwhile, the passenger trip is greatly affected by the point of interest. For example, for a bus station around a primary school, the passenger flow change trend will be related to the school class time. The financial street and the hospital of the urban functional area also have similar influences, and it is important to make full use of the point of interest in the prediction model. This reflects that the bus stations at different geographical positions can have different passenger flow change patterns, and the point of interest extractor of the present application will construct the station feature matrix and the station attribute table based on the distance from the bus station to the surrounding points of interest.

[0037] The method of the present application takes the regional functional points around the public transportation station as the point of interest, such as schools, hospitals, etc., which contain the potential external factors for passengers to go to or leave the area. Therefore, it is necessary to study the points of interest around the public transportation system. The present application designs a point of interest extractor to establish the connection between the points of interest and the public transportation station, and also considers the regional functions around the public transportation system in the flow prediction. According to the straight-line distance from the different types of points of interest to the public transportation station, the point of interest extractor divides all the public transportation stations into several station groups, and the stations in the same station group are close to the same type of point of interest. Through the point of interest extractor, all the stations in a station group are affected by the same external point of interest, while different station groups mean that they are affected by different external points of interest, so that the method of the present application classifies the tasks of the public transportation flow prediction from the perspective of urban points of interest.

[0038] In the embodiment of the present application, considering the distribution of urban functions in the real world, three types of geographical coordinate information of the point of interest are selected, which are primary and secondary schools P edu , tertiary grade A hospitals P hos and financial and insurance agencies P fin . The method of the present application calculates the Euclidean distance from each point of interest to the bus station, and assumes that only the points of interest within a distance of 500 meters will affect the passenger flow of the bus station. From this, the number of various urban points of interest in the station area can be calculated, and then a station feature matrix with a shape of N x 3 is generated, which represents the number weight of three types of urban points of interest around N stations. The weight is the ratio of the number of this type of point of interest to the total number of points of interest around the station. In addition, in order to facilitate the flexible configuration of the model by the city manager, the influence weight of the three types of points of interest can also be manually set, and the corresponding weights are represented as ω edu , ω hos , ω fin . The greater the weight, the greater the influence of this type of point of interest on the public transportation flow.

[0039] The specific steps of constructing the station attribute table are as follows:

[0040] Firstly, the geographic longitude and latitude of all stations in the public transportation system under study and the geographic longitude and latitude of all points of interest in the urban area need to be obtained, and a station attribute table initially set as a zero matrix is constructed, with a size of NxC, wherein each row in the table corresponds to the attributes of a station, and C represents the number of characteristics of the station with respect to the points of interest, since three types of points of interest are selected in the embodiment of the application, C = 3. The three characteristics of the station attribute table in the embodiment of the application correspond to primary and secondary schools, third-grade A-level hospitals and financial and insurance institutions in turn.

[0041] Then, the list of points of interest is traversed, and for each point of interest such as primary and secondary schools, third-grade A-level hospitals and financial and insurance institutions, public transportation stations within a set range r from the point of interest are found, and the corresponding point of interest position of the found station is added to the corresponding weight value ω edu / ω hos / ω fin r can be set according to experience or experiment, for example, set to 500 meters. The weight value can be set in advance according to the actual situation.

[0042] Finally, the station attribute table matrix is normalized by row, and the influence of the three different points of interest on each station can be obtained. If the station attribute value is 0, it means that the station is not affected by any selected point of interest, and belongs to other types of stations. Otherwise, if the attribute of a station is [0.1, 0.2, 0.7], it means that the station is most affected by the financial and insurance institutions, and the weight value is 0.7. The point of interest extractor determines the quantitative influence of different types of points of interest on the bus station by calculating the feature matrix of each station and the station attribute table.

[0043] According to the station attribute table, all public transportation stations can be divided into four types of station groups, and the stations in the same type of station group are close to the same type of point of interest. In the embodiment of the application, the four types of station groups are: stations that are severely affected by primary and secondary schools, third-grade A-level hospitals and financial and insurance institutions, and stations that are not affected by any selected point of interest. All stations in a station group are more affected by the same point of interest, while different station groups mean that they are more affected by different points of interest. Through the station attribute table, a reasonable task classification is performed for the public transportation flow prediction from the perspective of urban points of interest, which will be described in step 3.

[0044] Step 3) Establish a deep learning model based on spatiotemporal causal relationship for predicting public transportation station flow.

[0045] The deep learning model designed in the application is stacked by multiple spatiotemporal causal graph convolution units. For example, Figure 1As shown, the pre-processed data with a shape of 316*64 output by step 1 is first input into the first spatio-temporal causal graph convolution unit, and then the output of each spatio-temporal causal graph convolution unit will be used as the input of the next spatio-temporal causal graph convolution unit. In each spatio-temporal causal graph convolution unit, the data will first pass through two parallel temporal causal convolutions (step 3.1), which respectively adopt a sigmoid activation function and a tanh hyperbolic tangent activation function, the former of which plays a gating role on the latter to achieve better nonlinear fitting effect, and the outputs of the two temporal causal convolutions are input into the spatial causal graph convolution (step 3.2) after being fused by Hadamard product.

[0046] In other deep learning-based traffic prediction, the convolutional network for extracting temporal information uses the same model parameters for all traffic nodes, which assumes that all traffic nodes share the same traffic variation pattern. However, as emphasized in step 2 of the present application, due to the uneven distribution of urban interest points, bus stops distributed in different areas are also likely to have different passenger flow variation patterns. Therefore, the present application innovatively uses four different sets of model parameters for the spatio-temporal causal convolution to correspond to the traffic pattern under the influence of three different interest points. Specifically, four sets of model parameters of the temporal causal convolution layer are set to correspond to four types of station groups, and the fourth type of station is a station not affected by any selected interest point. According to the station attribute table, the output weight of the station corresponding to the first three sets of models is directly determined. If the station is the fourth type of station, the output weight of the station to the fourth set of models is set to 1, otherwise, the output weight of the station to the fourth set of models is set to 0. The calculation results of the input data under the configuration of the four sets of model parameters are weighted and summed according to the output weight, and the output data obtained is input into the spatial causal graph convolution layer.

[0047] For example, if the corresponding attribute of a station on the station attribute table is [0.1, 0.2, 0.7], the station has undergone calculation under the configuration of the four sets of model parameters in the deep spatio-temporal causal convolution, and the output results of the temporal causal convolution network of the four sets of model parameters corresponding to the station are weighted and summed according to the weight ratio of [0.1, 0.2, 0.7, 0]. If the attribute of a station on the station attribute table is [0, 0, 0], the output results of the four sets of temporal causal convolution network corresponding to the station are weighted and summed according to the weight ratio of [0, 0, 0, 1].

[0048] The construction of the deep learning model based on spatio-temporal causal relationship of the present application will be described in detail below.

[0049] Step 3.1) Implementation of the Temporal Causal Convolution. The present invention uses gating technique to guarantee the nonlinearity of the convolution operation to achieve good fitting ability for complex data, and also uses dilated convolution technique to achieve a larger input information receptive field. As a variant of convolutional neural network, the temporal causal convolution will be used to capture hierarchical temporal correlation, which has obvious advantages in time series modeling, such as good parallelism, flexible receptive field, stable gradient and low memory consumption in training.

[0050] For a classic convolutional neural network, if its convolution kernel size is 2, each convolutional layer will transform the length of the time series from T to T-1, so T-1 layers of convolution are needed to obtain the final single-step prediction result. This deep learning network is inefficient in training, and the present invention uses dilated convolution calculation, which makes the sequence length transformed by each convolutional layer shrink at an exponential rate. As shown in Figure 2 , the present invention sets the dilation coefficient d to 2, so the receptive field of the temporal causal convolution from the second layer will be 2 times that of the previous layer. Obviously, this exponentially growing receptive field has a requirement for the length of the input sequence, for example, the length of the time series for single-step prediction needs to be several orders of magnitude of the dilation. In the case where the length of the input sequence T does not meet the requirement, the present invention will pad 0 at the beginning of the sequence to ensure that the sequence can be evenly divided. For example, the input sequence of length 48 is padded to 64 for single-step prediction. From a mathematical expression, the receptive field size rft t is calculated as follows:

[0051] rf t = (k-1) x d n -1

[0052] where k is the convolution kernel size, d represents the dilation coefficient, and n is the number of hidden layers. In addition, a gating mechanism is used to enhance the nonlinear fitting ability. The output h of the temporal causal convolution layer (TC layer) can be represented as:

[0053] h = φ sig (θ1★x+b)⊙φ tanh (θ2★x+c)

[0054] where x is the input time series, i.e. the preprocessed passenger flow data in the first spatio-temporal causal graph convolution unit, φ sig is a sigmoid function, φ tanh is a hyperbolic tangent function, θ1★and θ2★are two extended temporal causal convolutions, b, c are learning parameters, and ⊙ is Hadamard product (element-wise multiplication).

[0055] As Figure 2As shown, ignoring other algorithmic parts, the entire temporal-causal design pyramid is used to implement huge corresponding receptive fields in several layers. Due to exponential explosion, not all nodes of hidden layers contribute to the final result, while the prediction point can still accept full information from the input layer.

[0056] Step 3.2) Implementation of spatial-causal graph convolution. The spatial-causal graph convolution layer, abbreviated as SC layer, is arranged between two TC layers to extract the features of the station according to the adjacency matrix. In the present application, the SC layer is implemented by a graph convolution network, which is an emerging field in traffic prediction.

[0057] As described above, the graph of the public transportation system is defined as where A is the adjacency matrix describing the edge weight between nodes. In other graph convolution-based traffic prediction methods, the Euclidean distance is often used as the adjacency matrix A of the traffic node, but buses cannot travel between two stations on different routes, so the traditional mapping method does not conform to the actual operation of public transportation. The connection between public transportation stations is based on the topological relationship of public transportation routes rather than Euclidean distance, and the adjacency matrix calculation method used in the present application effectively avoids this defect. As described in step 1 data preprocessing, the present application uses the historical passenger flow between public transportation stations to calculate the N x N adjacency matrix A. First, the historical passenger departure and arrival volume between N bus stations is counted, and the value at position [i, j] in the A matrix represents the historical total flow from station i to station j, so this adjacency graph only contains the relationship between stations on the same bus line. In addition, the present application uses an adaptive graph convolution network and an adaptive adjacency matrix change in training. The SC layer is represented as follows:

[0058]

[0059]

[0060]

[0061]

[0062] where, A ∈ R N×N is the adjacency matrix generated in step 1 data preprocessing, and rowsum represents the sum of the matrix by row. f , P b respectively normalize the adjacency matrix A by row and by column; the superscript T represents the transpose. adp As an adaptive graph, the trainable parameter vector E1 ∈ R N×1 , E2 ∈ R N×1is generated. K represents the number of hops (K-hop) of graph convolution, representing the depth of graph convolution to aggregate station information, which is set to 2 by default in the application. X is in R N×D represents an input signal, and Z is in R N represents an output, and W k1 represents an output, and W k2 represents an output, and W k3 is in R D×M represents the graph layer parameter when the number of attribute features on the station is D and the number of output attribute features is M. represents a variable adjacency matrix generated when the number of hops of graph convolution is k, which will be continuously adjusted and optimized in the neural network training process. and and are matrices obtained by row normalization and column normalization of the adjacency matrix A when the number of hops of graph convolution is k.

[0063] If the attribute features on the station include boarding flow and alighting flow, then D = 2, and the station attribute features can be expanded if other factors such as weather are further considered. The output attribute feature M is determined according to the result to be predicted, for example, if only the boarding flow is to be predicted, then M = 1, and if the boarding flow and the alighting flow are to be predicted, then M = 2.

[0064] From the specific physical meaning, and represent the departure flow relationship graph from each bus station to other stations and the arrival flow relationship graph from other stations to each bus station, so that the graph convolution aggregates the station feature information from the two edge relationship angles of departure flow and arrival flow. And As a matrix generated entirely by trainable parameters, it can be continuously adjusted and optimized in the neural network training process, and plays a role in adaptively adjusting the output of the graph convolution layer.

[0065] From the spatial perspective of urban public transportation, the graph convolution smoothes the passenger flow of a station by aggregating and transforming its neighborhood flow to realize the propagation of vertex information, which is concentrated on the stations of the same route in public transportation.

[0066] The embodiment of the application linearly connects six spatio-temporal causal graph convolution units, each unit consisting of two time causal convolutions using different activation functions and one spatial causal graph convolution. The output vectors of the two time causal convolutions are Hadamard multiplied and then input into the spatial causal graph convolution. The output of each spatio-temporal causal graph convolution unit will be used as the input of the next unit. The output of the last unit will generate the final prediction result, i.e. the flow of each public transportation station in the prediction time window, after two one-dimensional convolution calculations in series.

[0067] From the above, it is illustrated that the public transport flow prediction method combining city interest points and space-time causality provided by the application is to take the boarding and alighting flow of each time period of the subway bus station in the city as the input, combine the city interest point information near the station, and use the deep learning method to make a prediction on the boarding and alighting flow in the future time period. Appropriate data preprocessing and interest point extractor are arranged to process the original passenger flow data and interest point position data, and a deep learning model based on space-time causality is used to perform the final flow prediction.

[0068] As Figure 3 shown, the method of the application is actually tested on 316 public transport stations of a public transport data set of a certain city in a certain urban area, and the prediction curve and the actual boarding and alighting passenger flow curve basically coincide.

[0069] In addition to the technical features described in the specification, they are known to those skilled in the art. The application omits the description of known components and known technologies to avoid redundancy and unnecessary limitation of the application. The embodiments described in the above embodiments do not represent all embodiments consistent with the present application. Various modifications or variations made by those skilled in the art on the basis of the technical solutions of the application without creative labor are still within the protection scope of the application.

Claims

1. A public transportation flow prediction method combining urban points of interest and spatio-temporal causality, characterized in that, The application relates to a public transport station passenger flow prediction method based on spatio-temporal causal graph convolution network. The method comprises the following steps: Step 1, obtaining passenger flow in and out of a public transport station in each time period in a city, pre-processing, and obtaining a splicing sequence containing passenger flow in and out of the station in the same time period in the recent period; Step 2, taking the regional function points around the station as interest points, using an interest point extractor to establish the connection between the regional function points and the station, and calculating a station attribute table; The interest point extractor obtains a list of all interest points in the city, traverses the list of interest points, and calculates the station attribute table; each row of the station attribute table records the attribute value of each station for each interest point, and the attribute value is calculated in the following manner: the attribute value of each station is initially 0, for each interest point, the attribute value of the station is added to the influence weight of the interest point by searching for the station within a set range in terms of straight-line distance; after the traversal is completed, the values in the station attribute table are normalized by row; wherein the influence weights of different interest points are set in advance; Let q be the number of interest point categories, q+1 categories of stations are analyzed from the station attribute table, the i-th category of stations represents stations that are seriously affected by the i-th category of interest points, 1<=i<=q, and the q+1-th category of stations represents stations that are not affected by any interest point; q is a positive integer; Step 3, establishing a deep learning model based on spatio-temporal causal relationship for predicting the passenger flow of the public transport station; The model comprises P stacked spatio-temporal causal graph convolution units, the output of each unit will be used as the input of the next unit; the splicing sequence obtained in step 1 is input into the first spatio-temporal causal graph convolution unit, and the output of the last spatio-temporal causal graph convolution unit is calculated through two one-dimensional convolutions in series to generate the passenger flow of each public transport station in the prediction time window; P>2; The spatio-temporal causal graph convolution unit comprises a time-series causal convolution layer and a spatial causal graph convolution layer; the time-series causal convolution layer is composed of two parallel time-series causal convolutions; the input data are first processed through the two parallel time-series causal convolutions, the outputs of the two time-series causal convolutions are fused through Hadamard product and then input into the spatial causal graph convolution; q+1 sets of model parameters of the time-series causal convolution layer are set, corresponding to the q+1 categories of stations, and the output weights of the stations corresponding to the first q sets of models are determined according to the station attribute table; the output weight of the q+1 set of models is set as 1 for the q+1 category of stations and 0 for other categories of stations; the input data are processed through the time-series causal convolution layer with q+1 sets of model parameters, and the output data corresponding to each station are weighted and summed according to the output weights; 2. The method of claim 1, wherein, The model is trained, and the trained model is used for predicting the passenger flow of the public transport station. In step 1, the pre-processing splices the first m1 time slices on the day of the prediction target time window and the same (m1+1) time windows in the recent m2 days to obtain the passenger flow sequence of all stations in the same time period in the recent period; wherein m1 and m2 are positive integers.

3. The method of claim 1, wherein, In the step 2, the interest point extractor further constructs a site feature matrix, for each site, the number of interest points around the site and the proportion of each interest point are counted, and the proportion is taken as the influence weight of the interest point on the site, and the site feature matrix records the influence weight of each interest point on each site.

4. The method of claim 1, wherein, In the step 3, in the time causal convolution layer, two time causal convolutions respectively adopt sigmoid activation functions and tanh hyperbolic tangent activation functions.

5. The method of claim 1, wherein, In the step 3, the time causal convolution layer is implemented in the following manner: the time causal convolution adopts dilated convolution, and the output h of the time causal convolution layer is as follows: h = φ sig (θ1 * x + b) ⊙ φ tanh (θ2 * x + c) where φ sig is a sigmoid function, φ tanh is a hyperbolic tangent function, θ1*and θ2*are two dilated temporal causal convolutions, b, c are learning parameters, and ⊙ is the Hadamard product.

6. The method of claim 1, wherein, In the step 3, the spatial causal graph convolution layer extracts site features according to a site adjacency matrix A; the site adjacency matrix A is calculated using historical passenger flow between public transportation sites, assuming that the number of public transportation sites is N, A is an N*N matrix, and the value of the [i,j] position in the matrix represents the total historical flow of the site i to the site j.

7. The method of claim 6, wherein, In the step 3, the spatial causal graph convolution layer extracts site features according to a site adjacency matrix A, as follows: First, the adjacency matrix A is normalized by row and by column, respectively, to obtain matrices P f and P b ; Then, a variable adjacency matrix A is generated using trainable parameter vectors E1 and E2 adp As follows: Wherein, E1 and E2 are both N*1 vectors, and N is the number of public transportation sites. Finally, the output Z of the spatial causal graph convolution layer is as follows: wherein X is input data, K represents the number of hops of graph convolution, W k1 ,W k2 ,W k3 are trainable parameters; represents a variable adjacency matrix generated when the number of graph convolution hops is k, which will be continuously adjusted and optimized in the neural network training process; and are matrices obtained by normalizing the adjacency matrix A by row and column respectively when the number of graph convolution hops is k.

Images