Passenger flow distribution prediction method and device based on multi-view passenger flow change trend modeling

Abstract

The present invention relates to the intersection of deep learning and urban traffic planning technology, and specifically to a passenger flow distribution prediction method and device based on multi-view passenger flow change trend modeling. The method and device use nodes and edges to construct a dynamic complete graph from a multi-view perspective according to the current time period, and the nodes and edges respectively extract the characteristics of the travel destination and the destination station, and connect the original isolated travel destination nodes with the destination station nodes through the relationship between the nodes, and calculate the passenger flow distribution of the entire network in the next time period based on the constructed connection data, so that the nodes can perceive more relevant information in the traffic road network, thereby effectively improving the accuracy of the downstream destination station prediction task.

Description

Technical Field

[0001] The present invention belongs to the intersection of deep learning and urban traffic planning technology, and in particular, relates to a passenger flow distribution prediction method and device based on multi-view passenger flow change trend modeling. The invention can be used to predict the passenger flow from the departure point to the destination station at public transportation stations, the number of highway vehicles from the departure point to the destination station, the number of shared bicycles from the departure point to the destination station, etc. Background Art

[0002] Passenger flow distribution prediction is also called OD (Origin-Destination) matrix prediction, which aims to predict the number of trips from one area (station) to another area (station) in the future. Passenger flow distribution prediction is an important step in traffic demand prediction and a prerequisite for route planning, traffic decision-making, etc. Existing passenger flow distribution prediction models mainly include long-term passenger flow distribution prediction and short-term passenger flow distribution prediction. Long-term passenger flow distribution prediction mainly uses months or years as the time granularity. Based on survey data, according to the current traffic generation and attraction of each traffic area, as well as the planning information of each traffic area in the future, it predicts the distribution of long-term urban traffic in the future. Long-term passenger flow distribution prediction methods mainly include pure model-driven methods such as gravity model and growth rate model. These methods are based on ideal assumptions and are based on survey data. By fitting pre-built functions, they predict the future passenger flow distribution. For example, the gravity model considers the generation of the starting point and the attraction of the end point, as well as the distance between the two points, and uses the physical gravity model to model the passenger flow distribution relationship between the two stations. Since these methods consider limited factors and use simple correlation functions, they can only make rough estimates of passenger flow distribution and are not suitable for real-time traffic decision-making scenarios that require high real-time performance of passenger flow distribution.

[0003] Short-term passenger flow distribution prediction is used to predict the urban passenger flow distribution in the short term (e.g., 15 minutes, 30 minutes, 60 minutes, etc.) in the future. In recent years, with the advancement of traffic sensing technology, existing technologies can obtain real-time passenger flow distribution of some traffic subsystems (subways, taxis, etc.), which provides a data basis and new research ideas for the present invention to predict short-term passenger flow distribution. Since short-term passenger flow distribution involves many influencing factors, and there are problems such as high dimensionality, sparseness, and random uncertainty, passenger flow distribution prediction faces the following challenges.

[0004] There are many related factors and the temporal and spatial correlation is complex: From the perspective of factors affecting passenger flow distribution, the distribution of urban passenger flow involves two areas (the starting area and the destination area). From the perspective of the starting station, its future passenger flow distribution is not only related to the changing trend of the passenger flow generation volume at this station, but also to the distribution trend of the destination stations of the passenger flow departing from this station. From the perspective of the destination station, the distribution of passenger flow arriving at this station is not only related to the changing trend of the attraction volume of this station, but also to the distribution trend of the starting stations of the passenger flow arriving at this station. In addition, the distribution of urban passenger flow also presents complex dynamic changes with factors such as time and station distance. A major challenge is how to integrate these influencing factors to predict the future trend of passenger flow changes.

[0005] High dimensionality and sparse and uncertain information: From the perspective of passenger flow distribution, the number of OD pairs in the entire subway network grows quadratically with the number of stations. Since the total travel demand is scattered across different OD pairs, there are a large number of OD flows that are very small or even zero, and a large amount of sparse OD flows are caused by random and uncertain travel. The aggregation of these large amounts of random and uncertain traffic information further increases the difficulty of prediction.

[0006] Existing technologies extract travel patterns from historical passenger flows from travel destinations to destinations and use machine learning methods to predict passenger flows from travel destinations to destinations, including: (1) a linear regression method that predicts passenger flows from travel destinations to destinations from historical travel destination and destination data. (2) a time series model based on travel rules. (3) based on regional functional destination prediction, combined with POI features, extracting semantic features of each fine-grained location, constructing station features, and finally multiplying the station features with their own transpose to obtain the prediction target.

[0007] The existing implementation scheme that is most similar to the present invention is to use a graph convolutional neural network to extract node features and then input them into a time series module to obtain a destination site transfer vector. The destination site transfer vector, the constructed learnable transfer matrix, and the transpose of the destination site transfer vector are multiplied together to obtain the predicted travel destination site prediction result.

[0008] So far, there are some patents related to subway OD prediction. For example, based on the occurrence and attraction of each station, as well as the passenger flow distribution information of the past days, and using sequence learning models such as LSTM to establish the temporal correlation between traffic flows. Based on the characteristics of adjacent nodes being related to travel distribution, the high-dimensional sparse uncertainty problem caused by randomness is alleviated by aggregating the characteristics of surrounding neighbor nodes. However, due to the shortcomings of these methods such as insufficient considerations and one-sided modeling of spatiotemporal correlation, there is room for further improvement in subway OD prediction. The specific limitations are summarized as follows:

[0009] (1) Only some traffic dynamic information is taken into account. When predicting passenger flow distribution, only the dynamic inbound and outbound passenger flow of each subway station and the distribution of passenger flow in the past few days are considered, ignoring the real-time passenger flow distribution information, which is also very important for future passenger flow distribution.

[0010] (2) Only the spatiotemporal correlation of local traffic information is modeled. Most existing methods analyze the trend of future passenger flow distribution from the perspective of the starting station. However, the result of the passenger flow distribution of the entire network in the future is the result of the combined effect of the passenger flow changes and distribution trends of the starting and destination stations.

[0011] (3) In terms of sparsity uncertainty, although the sparse uncertainty caused by dynamic randomness can be alleviated to a certain extent by aggregating the passenger flow characteristics of adjacent stations, the relevance of surrounding stations to the target station will show different degrees of relevance depending on factors such as their own characteristics and the distance between stations. However, existing methods only measure the degree of relevance between stations based on some factors (such as physical distance or traffic similarity, etc.) and cannot accurately measure the complex relevance between stations. Summary of the invention

[0012] The embodiments of the present invention provide a method and device for predicting passenger flow distribution based on multi-view passenger flow change trend modeling, so as to at least solve the technical problem that the prediction accuracy of the existing passenger flow distribution prediction technology is not high.

[0013] According to an embodiment of the present invention, a passenger flow distribution prediction method based on multi-view passenger flow change trend modeling is provided, comprising the following steps:

[0014] Build a dynamic complete graph using nodes and edges from multiple viewpoints according to the current period;

[0015] The nodes and edges extract the features of the travel destination and the destination station respectively, and the original isolated travel destination nodes are connected with the destination station nodes through the relationship between the nodes;

[0016] The passenger flow distribution of the entire network in the next time period is estimated based on the constructed contact data.

[0017] Furthermore, the method comprises the following specific steps:

[0018] Construct a deep representation of the passenger flow change trend of each station as the starting station. Based on two types of real-time observed dynamic passenger flows related to the starting station and integrating the station proximity relationship, model the passenger flow change trend of the starting station at different angles from two views. Finally, the results of the two views are fused as the passenger flow trend representation of the starting station.

[0019] Construct a deep representation of the passenger flow change trend of each station as a destination station. Based on two types of real-time observed dynamic passenger flows related to the destination station and integrating the station proximity relationship, model the passenger flow change trend of the destination station from different angles from two views. Finally, the results of the two views are fused as the passenger flow trend representation of the destination station.

[0020] The transfer matrix from the starting site to the destination site is constructed based on the proximity relationship between sites, and the passenger flow distribution of the entire network in the next period is estimated based on the passenger flow change trends of the constructed starting and destination sites.

[0021] Furthermore, based on two types of real-time observed dynamic passenger flows related to the destination station, the station proximity relationship is integrated, including:

[0022] The vector of the destination station allocation is based on the inbound passenger flow in the recent period and the completed trips in the recent period, and the proximity of the stations and the influencing factors of time and weather are integrated.

[0023] Furthermore, based on two types of real-time observed dynamic passenger flows related to the destination station, the station proximity relationship is integrated, including:

[0024] The vector is allocated based on the outbound passenger flow in the recent period and the destination station that has completed the trip in the recent period, and the proximity relationship between stations and the influencing factors of time and weather are integrated.

[0025] Furthermore, the results of the two views are fused using the same passenger flow change trend learner.

[0026] Furthermore, the passenger flow change trend learner is used to:

[0027] Firstly, the passenger flow dynamic information of each station in the subway system in the recent period, the station proximity matrix, and other global factors affecting the dynamic change trend of passenger flow are organized into a graph sequence that supports the learning of complex spatiotemporal relationships.

[0028] Then, the extended GAT graph attention neural network is used to model the dynamic uncertainty of passenger flow changes at each station in the recent observation period by aggregating passenger flow change information at nearby stations.

[0029] Then, the threshold recurrent unit model is used to model the overall changing trend of passenger flow over time, and finally a deep representation of the passenger flow changing trend of each station is obtained.

[0030] Furthermore, the passenger flow information of neighboring nodes is aggregated to alleviate the uncertainty of the model's learning of passenger flow change trends caused by the randomness of local observation data.

[0031] Furthermore, KDD technology is used from a global perspective to model the transformation relationship between the origin and destination sites through a network-wide shared transformation matrix.

[0032] According to another embodiment of the present invention, a passenger flow distribution prediction device based on multi-view passenger flow change trend modeling is provided, comprising:

[0033] The starting station dynamic feature learning module is used to construct a dynamic complete graph using nodes and edges from multiple view perspectives according to the current period;

[0034] The destination station dynamic feature learning module is used to extract the travel destination features and destination station features of nodes and edges respectively, and to establish a connection between the original isolated travel destination nodes and destination station nodes through the relationship between nodes;

[0035] The starting station-destination station transfer module is used to estimate the passenger flow distribution of the entire network in the next time period based on the constructed contact data.

[0036] Furthermore, the device specifically comprises:

[0037] The starting station dynamic feature learning module is used to construct a deep representation of the passenger flow change trend of each station as the starting station. Based on two types of real-time observed dynamic passenger flows related to the starting station and integrating the station proximity relationship, the passenger flow change trend of the starting station at different angles is modeled from two views. Finally, the results of the two views are integrated as the passenger flow trend representation of the starting station.

[0038] The destination station dynamic feature learning module is used to construct a deep representation of the passenger flow change trend of each station as a destination station. Based on two types of real-time observed dynamic passenger flows related to the destination station and integrating the station proximity relationship, the passenger flow change trend of the destination station at different angles is modeled from two views. Finally, the results of the two views are integrated as the passenger flow trend representation of the destination station.

[0039] The starting station-destination station transfer module is used to build a transfer matrix from the starting station to the destination station based on the proximity relationship between the stations, and to calculate the passenger flow distribution of the entire network in the next period based on the passenger flow change trends of the constructed starting station and destination station.

[0040] The passenger flow distribution prediction method and device based on multi-view passenger flow change trend modeling in the embodiment of the present invention use nodes and edges to construct a dynamic complete graph from a multi-view perspective according to the current time period, and the nodes and edges respectively extract travel destination characteristics and destination station characteristics, and establish a connection between the original isolated travel destination nodes and the destination station nodes through the relationship between the nodes, and infer the passenger flow distribution of the entire network in the next time period based on the established connection data, so that the nodes can perceive more relevant information in the transportation network, thereby effectively improving the accuracy of the downstream destination station prediction task. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0042] Figure 1 It is a model overall framework diagram of the passenger flow distribution prediction device based on multi-view passenger flow change trend modeling of the present invention;

[0043] Figure 2 This is the specific operation process of the passenger flow change trend learner of the present invention. DETAILED DESCRIPTION

[0044] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0045] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0046] Compared with the previous travel destination prediction method, the present invention proposes to use nodes and edges to construct a dynamic complete graph from a multi-view perspective. The nodes and edges extract the travel destination characteristics and the destination spatiotemporal characteristics respectively. After extracting these two characteristics, the original isolated travel destination nodes are connected through the relationship between nodes, so that the nodes can perceive more relevant information in the traffic network, thereby effectively improving the accuracy of the downstream destination prediction task. It is expected that the evaluation index MAE will be improved by 5% compared with the previous method.

[0047] The present invention predicts the passenger flow from a travel destination to a destination station based on the historical passenger flow and time from the travel destination to the destination station. The present invention takes the subway system scenario as an example, and predicts the passenger flow from the travel destination to the destination station in Q time periods starting from the current time period when the passenger flow from the travel destination to the destination station in the previous P time periods of the current time period is known. The method can also be used for passenger flow prediction from the travel destination to the destination station in other similar scenarios, such as predicting the number of passengers from the travel destination to the destination station in a specific time period, and the number of vehicles from the travel destination to the destination station.

[0048] The technical solution of the present invention is introduced by taking a subway as an example.

[0049] (1) Data introduction

[0050] The present invention uses the smart card transaction data set collected by the automatic fare collection (AFC) system. The automatic fare collection system records the information of each passenger's every ride on public transportation. Each transaction record contains four fields, namely: ID, SID, Time, and Type. Among them, ID is the unique identifier of the smart transportation card, which can be used to uniquely identify the passenger; SID is the identifier of the subway station or bus station; Time is the transaction (card swiping) time, and Type is the transaction type (entry and exit).

[0051] (2) Basic Concept Definition

[0052] Subway system: The subway system consists of multiple physical stations, represented by S = (s1, s2, ..., s N ), and the average time matrix C∈R between sites N×N . C i,j Indicates site i With sites j The time spent between.

[0053] A trip: A passenger's trip tr refers to a logical riding process, that is, the whole process from entering the station to leaving the station, including tr.s o , tr.s d , tr.t o , tr.t d The four fields represent the starting station, the destination station, the starting time and the exit time. The AFC system can collect all the smart card transaction records of each passenger based on the ID information, sort these records by time, and combine the trip type Type into two consecutive trip records of entry and exit, so as to get each trip of the passenger.

[0054] Time period: The present invention divides the time covered by the analyzed data into multiple time periods at fixed intervals, which are marked as: T = {t1, t2, ..., t |T|}.

[0055] OD matrix: The present invention uses OD t ∈R N×N represents the passenger flow distribution information of the entire network in time period t, namely, the OD matrix, which is also the basic unit of data to be predicted in the present invention. t ) i,j Indicates that in time period t from s i Passengers entering the station, from s j The passenger flow out of the station, that is The operation I(.) means that if the condition in the brackets is true, it is 1, otherwise it is 0.

[0056] The following are four types of passenger flow data that can be observed in real time in time period t∈T based on the collected transaction data. Among them, inbound passenger flow and outbound passenger flow involve passenger flow information of a single station, and destination station allocation passenger flow and starting station allocation passenger flow involve travel distribution information between stations.

[0057] Inbound passenger flow: This invention uses non-negative values Indicates that in time period t from site s i The passenger flow entering the station is referred to as the inbound passenger flow. Represents the inbound passenger flow of all stations in the entire network during time period t.

[0058] Outbound passenger flow: This invention uses non-negative values Indicates that in time period t from site s i The passenger flow out of the station is referred to as the outbound passenger flow. Represents the outbound passenger flow of all stations in the entire network during time period t.

[0059] Destination station distribution of passenger flow: The present invention uses Indicates that before time period t, from s i The destination station distribution vector of all trips that enter the station and complete in time period t (exit the station in time period t). The jth element is calculated as The present invention uses an N×N dimensional matrix It represents the destination station allocation matrix of the starting station of the entire network of trips completed in time period t. The i-th row represents the starting station s i The destination site allocation vector.

[0060] Starting station allocation of passenger flow: The present invention uses Indicates that in time period t from site s i The starting station distribution vector of the outgoing passenger flow. The jth element is calculated as The present invention uses represents the starting station allocation matrix for trips that arrive at all destination stations in time period t, where the i-th row represents station s i The starting site allocation vector.

[0061] Given all physical stations S of the subway network, the average time C between stations, and the passenger flow information observed in the first P (P>0) time periods, including the inbound passenger flow sequence X In ={In t-p+1 ,…,In t-1 ,In t}, outbound passenger flow sequence X Out ={Out t-p+1 ,…,Out t-1 ,Out t}, the starting station allocates passenger flow sequence X Do ={Do t-p+1 ,…,Do t-1 ,Do t}, the destination station allocates passenger flow sequence X Od ={Od t-p+1 ,…,Od t-1 ,Od t}, predict the passenger flow distribution matrix for the next time period: that is, formally described as: {S,C,X In ,X Out ,X Do ,X Od}→OD t+1 .

[0062] The overall architecture of the model proposed in this invention is as follows Figure 1 Specifically, it includes three modules, namely, the starting station dynamic feature learning module, the destination station dynamic feature learning module and the starting station-destination station transfer module.

[0063] The starting station dynamic feature learning module is used to learn the deep representation of the passenger flow change trend of each station as the starting station. It is based on two types of real-time observed dynamic passenger flows related to the starting station, namely the inbound passenger flow in the recent period and the destination station allocation vector of the trip completed in the recent period, and integrates the station proximity relationship and other influencing factors (such as time, weather, etc.), and models the passenger flow change trend of the starting station at different angles from two views. Finally, the results of the two views are fused as the passenger flow trend representation of the starting station. The two views use the same neural network structure (called the passenger flow change trend learner in the present invention).

[0064] The destination station dynamic feature learning module is used to learn the passenger flow change trend of each station as the destination station. Its idea is similar to that of the starting station dynamic feature learning module. It uses the passenger flow change trend learner with the same structure, based on two types of observed dynamic passenger flows related to the destination station, including the outbound passenger flow in the recent period and the station allocation vector of the starting point of the completed trip, and combines the station proximity relationship and other influencing factors to model the passenger flow change trend of the destination station from two views.

[0065] The origin-destination transfer module learns the transfer matrix from the origin to the destination by combining the proximity relationship between the sites, and combines the passenger flow change trends of the origin and destination learned through the above modules to estimate the passenger flow distribution of the entire network in the next period.

[0066] The present invention builds a passenger flow change trend learner to learn the high-level representation of passenger flow changes at each station as the starting station and destination station of a trip. The core of this is how to model the temporal trend and random uncertainty of passenger flow changes based on the observed passenger flow dynamic information under multiple views, station proximity relationship information and other factors. The following first introduces the passenger flow change trend learner, and then introduces the station passenger flow change trend representation of multiple view fusion.

[0067] The specific operation process of the passenger flow change trend learner is as follows: Figure 2 As shown in the figure. First, the passenger flow dynamic information of each station in the subway system in the recent period, the station proximity matrix, and other global factors that affect the dynamic change trend of passenger flow are organized into a graph sequence that supports the learning of complex spatiotemporal relationships. Then, the dynamic uncertainty of passenger flow changes at each station in the recent observation time period is modeled by aggregating the passenger flow change information of adjacent stations based on the extended GAT graph attention neural network. Then, the threshold recurrent unit model is used to model the overall change trend of passenger flow over time, and finally a deep representation of the passenger flow change trend of each station is obtained. The following introduces the execution process of the passenger flow change trend learner through three aspects: traffic flow graph sequence construction, passenger flow change uncertainty modeling based on the GAT multi-head graph attention mechanism, and GRU sequence temporal correlation learning.

[0068] 1. Construction of passenger flow dynamic sequence diagram

[0069] For any time period t' in the first P (P>0) time period, the present invention constructs a dynamic passenger flow graph Where V is the node set (site set) of the graph, A∈R N×N Indicates the proximity relationship or edge features between nodes. It represents the passenger flow characteristics of each node in time period t', also known as node characteristics. The following introduces these two types of characteristics respectively.

[0070] Characteristics of nodes: represents the passenger flow information of each station at time t', where the i-th row represents the passenger flow characteristics of station i. Figure 1 There are four sub-modules with different passenger flow change trends in the , namely, the incoming passenger flow change trend module, the incoming passenger flow destination station allocation trend module, the outgoing passenger flow change trend module and the outgoing passenger flow starting station allocation trend module, and the characteristics corresponding to station i are Its characteristic dimension D f They are 1, N, 1, and N respectively.

[0071] Edge characteristics: E∈R N×N Represents the proximity relationship between stations, which is calculated based on the inverse of the time spent between two stations. E∈R N×N It is static, that is, the constructed passenger flow map is shared.

[0072] Based on the above analysis, Figure 1 For each submodule in the present invention, a dynamic traffic state timing diagram of the most recent P time periods can be obtained.

[0073] 2. Site Embedding Representation Based on GAT Graph Attention Mechanism

[0074] The passenger flow changes at each station are sparse and random uncertain. In view of the similarity of passenger flow change trends at adjacent nodes, the present invention alleviates the uncertainty learning of the model on passenger flow change trends caused by the randomness of local observation data by aggregating passenger flow information of neighboring nodes.

[0075] Traffic status diagram for each observation period t' Each node v in i , the present invention aggregates all neighbor nodes N through a graph convolutional neural network i = {v j |E ij >0} passenger flow characteristics. In order to gather information about neighbor nodes, unlike the existing initial model GCN, which treats all neighbors of a node equally, the present invention adopts a graph attention mechanism to assign different attention scores to each neighbor. The core of graph attention is how to collect feature representations of neighbor nodes. The present invention extends the graph attention mechanism GAT (which only works on graphs with node features) to work on graphs with node features and edge features, and calculates the v of each neighbor node by fusing node features and edge features. j ∈N i v i importance.

[0076] Specifically, the characteristics of node i and node j are and The edge feature is The present invention first converts the input features into high-dimensional features through a learnable linear transformation to obtain sufficient expression capabilities, including On this basis, the present invention performs self-attention w ij =a(z i ||z j ||z i,j ) calculates the attention coefficient, indicating the importance of the feature of node j to node i. Where || represents the connection operation, and the a(.) operation is implemented using a single-layer forward neural network and a Leaky ReLU nonlinear activation function. In order to make the coefficients easy to compare on different nodes, the present invention uses the softmax function to regularize the sparse attention of all neighboring nodes, that is,

[0077]

[0078] Finally, the present invention uses the normalization coefficient to calculate the aggregated neighbor passenger flow features. After Attention, the passenger flow feature vector of node j is:

[0079] In order to stabilize the learning process of self-attention, the present invention uses a multi-head attention mechanism to obtain the traffic feature representation of the node. If there are K attentions, it is necessary to concatenate the vectors generated by the K attentions, that is,

[0080]

[0081] Based on the above process, for each time period t', the present invention obtains the feature representation of all nodes in the graph after aggregation For the dynamic traffic status time series of the first P time periods The present invention can obtain the time series represented by node embedding

[0082] 3. GRU-based passenger flow temporal correlation learning

[0083] In view of the fact that the passenger flow change trend has the characteristics of recent period correlation and periodic correlation, the present invention firstly Learning implicit in passenger flow sequence Then, the short-term time correlation of passenger flow is modeled together with the periodically related time attributes of the predicted time period t+1, such as the day of the week, the time period of the day, holidays, weekdays or weekends.

[0084] To this end, firstly based on To learn the correlation of time series, the present invention uses a GRU model with a small number of parameters and a low risk of overfitting, wherein the GRU only uses two gate switches to achieve a result close to that of the LSTM.

[0085] r t′ =σ(W r ·[h t′-1 ,X t′ ]+b r )

[0086] u t′ =σ(W u ·[h t′-1 ,X t′ ]+b u )

[0087]

[0088]

[0089] Where σ() is the sigmoid activation function, ⊙ represents element-wise multiplication, and r t is the reset gate and u t is the update gate, and the weight vector W and the bias vector b are the parameters that need to be learned. Based on the GRU model, the present invention obtains the passenger flow change trend representation h at the most recent time t t , to simplify the period, it is expressed as

[0090] On this basis, combined with the processed discretized time attribute f of the forecast target period t+1 t+1 , after a layer of full connection operation, the passenger flow trend representation r under a specific view is obtained t =FC(f t+1 ,h t ).

[0091] 4. Multi-view fusion passenger flow trend representation of the starting station (destination station)

[0092] In the starting station dynamic feature learning module, based on the passenger flow change trend learner, the change trend of the inbound passenger flow at each station can be obtained. Changing trend of inbound passenger flow destination station allocation The concatenation of these two trends is passed through the fully connected layer to obtain a dimension of N×D o The passenger flow change trend matrix of all stations as the starting station The i-th row Indicates site i As a representation of the starting site.

[0093] Similarly, in the destination station dynamic feature learning module, based on the passenger flow change trend learner, the change trend representation of the passenger flow of each station as the destination station can be obtained. Changing trend of inbound passenger flow destination station allocation The concatenation of these two trends is passed through the fully connected layer to obtain a dimension of N×D d The passenger flow change trend of each station in the whole network as the destination station is expressed Where D t The i-th row of Indicates that site i is used as the destination site.

[0094] 5. Learning about passenger flow conversion relationships

[0095] Based on the above method, the present invention learns the passenger flow change trend representation of each station as the starting station and the destination station. t and D t In order to obtain the OD matrix of the next time period. From a global perspective, drawing on the ideas of the KDD paper, the present invention models the conversion relationship between the starting point and the destination site by learning a network-wide shared conversion matrix W, that is, In addition, given the negative correlation between the passenger flow distribution between two stations and the time spent between stations. As shown in the figure below, the longer the travel time, the smaller the proportion of the number of trips, that is, passengers tend to travel short distances. This is in line with the first law of geography: "Everything is related to everything else, but things that are close are more related than things that are far away." In the study of passenger flow conversion relationships, the present invention incorporates the influence of the time spent factor and expands the above formula to in It is used to calculate the impact of time factors. The idea is that the distribution relationship between time and travel volume follows an exponential distribution. λ1 and λ2 are parameters to be learned.

[0096] The innovative technical points of the present invention are at least:

[0097] 1) A general passenger flow prediction model from a travel destination to a destination is proposed, which is applicable to the transportation network from a travel destination to a destination. The invention takes the subway transportation network as an example, defines the transportation network as a directed weighted graph, and introduces the model in detail.

[0098] 2) The model proposed in the present invention takes into account the spatiotemporal characteristics of the travel destination and the destination station at the same time, extracts the spatiotemporal characteristics of the travel destination and the destination station separately, and finally combines these two characteristics to predict the destination station.

[0099] 3) The present invention takes into account the impact of various factors from multiple different perspectives, such as the time characteristics of the traffic network, the flow change characteristics of the travel destination, and other factors.

[0100] The point to be protected by the present invention is the use of node feature extraction method proposed by the present invention to mine the travel characteristics of the travel destination and the arrival characteristics of the destination station, and after combining the travel characteristics with the arrival characteristics, use the time series model to extract the spatiotemporal characteristics respectively, and finally multiply the travel destination and destination station characteristics to obtain the transfer matrix method from the travel destination to the destination station.

[0101] The advantages of the model proposed in the present invention are as follows: first, the prediction model proposed in the present invention is universal and applicable to the prediction of passenger flow from the travel destination to the destination station in the transportation network. In addition, compared with the disadvantage of low accuracy in the prediction of the destination station in the existing technology, the present invention focuses on designing a model combining the spatiotemporal characteristics of the travel destination and the spatiotemporal characteristics of the destination station. The prediction model uses theoretical knowledge such as graph convolutional neural networks in the field of deep learning to achieve accurate prediction of passenger flow from the travel destination to the destination station.

[0102] The present invention experiments the model on data sets generated by subway travel in Shenzhen, Hangzhou and Shanghai. The experimental results show that the model proposed by the present invention can accurately solve the destination station problem.

[0103] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0104] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0105] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the system embodiments described above are only schematic. For example, the division of units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0106] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed over multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0107] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0108] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

[0109] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A passenger flow distribution prediction method based on multi-view passenger flow change trend modeling, characterized in that: The following steps are involved: (1) Construction of passenger flow dynamic sequence diagram: For any time period t' in the first P (P>0) time period, construct a dynamic passenger flow graph Where V is the set of nodes in the graph, A∈R N×N Indicates the proximity relationship or edge features between nodes. Represents the passenger flow characteristics of each node in time period t', also known as node characteristics: Characteristics of nodes: represents the passenger flow information of each station at time t', where the i-th row represents the passenger flow characteristics of station i. For the four sub-modules with different passenger flow change trends, namely, the incoming passenger flow change trend module, the incoming passenger flow destination station allocation trend module, the outgoing passenger flow change trend module and the outgoing passenger flow starting station allocation trend module, the characteristics corresponding to station i are respectively Its characteristic dimension D f They are 1, N, 1, N respectively; Edge characteristics: E∈R N×N Represents the proximity relationship between stations, which is calculated based on the inverse of the time spent between two stations. E∈R N×N It is static, that is, the constructed passenger flow graph is shared; For each submodule, a dynamic traffic state timing diagram G = (G t-P-1 ,G t-P-2 ,…G t ); (2) Site embedding representation based on GAT graph attention mechanism: By aggregating the passenger flow information of neighboring nodes, the uncertainty of the model's learning of passenger flow changes caused by the randomness of local observation data is alleviated; Traffic state graph G for each observation period t' t' Each node v in ∈G i , aggregate all neighbor nodes N through graph convolutional neural network i = {v j |E ij >0}; In order to gather information about neighbor nodes, a graph attention mechanism is used to assign different attention scores to each neighbor; By extending the graph attention mechanism GAT to act on a graph with node features and edge features, the node features and edge features are fused together to calculate the v of each neighbor node j ∈N i v i Importance: The characteristics of node i and node j are and The edge feature is First, the input features are converted to high-dimensional features through a learnable linear transformation to obtain sufficient expressive power, including On this basis, perform self-attention w ij =a(z i ||z j ||z i,j ) calculates the attention coefficient, indicating the importance of the features of node j to node i; where || represents the connection operation, and the a(.) operation is implemented using a single-layer forward neural network and a Leaky ReLU nonlinear activation function. The softmax function is used to regularize the sparse attention of all neighboring nodes, that is: Finally, the normalization coefficient is used to calculate the aggregated neighbor passenger flow features. After Attention, the passenger flow feature vector of node j is: Use the multi-head attention mechanism to obtain the traffic feature representation of the node. If there are K attentions, concatenate the vectors generated by the K attentions, that is: Based on the above process, for each time period t', the feature representation of all nodes in the graph after aggregation is obtained For the dynamic traffic state sequence G in the first P time periods, G = (G t-P-1 ,G t-P-2 ,…G t ), and obtain the time series X represented by the node embedding = (X t-P-1 ,X t-P-2 ,…X t ); (3) GRU-based passenger flow temporal correlation learning: In view of the fact that the passenger flow change trend has the characteristics of recent period correlation and periodic correlation, firstly based on the dynamic passenger flow sequence X = (X t-P-1 ,X t-P-2 ,…X t ), learn the short-term temporal correlation implicit in the passenger flow sequence X, and then combine it with the periodically related time attributes of the predicted period t+1 to model the temporal correlation of the passenger flow; First, based on X=(X t-P-1 ,X t-P-2 ,…X t ) To learn the correlation of time series, we use the GRU model with fewer parameters and less risk of overfitting. GRU only uses two gate switches, achieving results close to those of LSTM. r t′ =σ(W r ·[h t′-1 ,X t′ ]+b r ) you t′ =σ(W u ·[h t′-1 ,X t′ ]+b u ) Among them, σ() is the sigmoid activation function, ⊙ represents element multiplication, r t is the reset gate and u t is the update gate, the weight vector W and the bias vector b are the parameters that need to be learned, and the passenger flow change trend representation h at the most recent time t is obtained based on the GRU model t , to simplify the period, it is expressed as h t = GRU(X); On this basis, combined with the processed discretized time attribute f of the forecast target period t+1 t+1 , after a layer of full connection operation, the passenger flow trend representation r under a specific view is obtained t =FC(f t+1 ,h t ); (4) Multi-view fusion passenger flow trend representation of the starting and destination stations: In the starting station dynamic feature learning module, based on the passenger flow change trend learner, the change trend of the inbound passenger flow at each station is obtained. t in , the changing trend of the destination station allocation of inbound passenger flow t oD ; The above two change trends are concatenated through the fully connected layer to obtain a dimension of N×D o The passenger flow change trend matrix O of all stations as the starting station t =fc([r t in ||r t oD ]), where the i-th row Indicates site i As a representation of the starting site; Similarly, in the destination station dynamic feature learning module, based on the passenger flow change trend learner, the passenger flow change trend representation r of each station as the destination station is obtained. t out , the changing trend of the destination station allocation of inbound passenger flow t dO ; The above two change trends are concatenated through the fully connected layer to obtain a dimension of N×D d The passenger flow change trend of each station in the whole network as the destination station is expressed as D t =fc([r t out ||r t dO ]), where D t The i-th row of Indicates that site i is the destination site; (5) Passenger flow conversion relationship learning: Based on the above method, the passenger flow change trend representation of each station as the starting station and the destination station is learned. t and D t ; In order to obtain the OD matrix of the next period, from a global perspective, a network-wide shared transformation matrix W is learned to model the transformation relationship between the starting point and the destination site, that is, Given the negative correlation between the passenger flow distribution between two stations and the time spent between stations: the longer the travel time, the smaller the proportion of the trip, that is, passengers are more inclined to travel short distances; in the study of passenger flow conversion relationship, the influence of the time spent factor is incorporated, and the above formula is expanded to in It is used to calculate the impact of time factors. The idea is that the distribution relationship between time and travel volume follows an exponential distribution. λ1 and λ2 are parameters to be learned.

2. A passenger flow distribution prediction device based on multi-view passenger flow change trend modeling using the passenger flow distribution prediction method based on multi-view passenger flow change trend modeling according to claim 1, characterized in that: include: The starting station dynamic feature learning module is used to construct a dynamic complete graph using nodes and edges from multiple view perspectives according to the current period; The destination station dynamic feature learning module is used to extract the travel destination features and destination station features of nodes and edges respectively, and to establish a connection between the original isolated travel destination nodes and destination station nodes through the relationship between nodes; The starting station-destination station transfer module is used to estimate the passenger flow distribution of the entire network in the next time period based on the constructed contact data.

3. The passenger flow distribution prediction device based on multi-view passenger flow change trend modeling according to claim 2 is characterized in that: Specifically include: The starting station dynamic feature learning module is used to construct a deep representation of the passenger flow change trend of each station as the starting station. Based on two types of real-time observed dynamic passenger flows related to the starting station and integrating the station proximity relationship, the passenger flow change trend of the starting station at different angles is modeled from two views, and finally the results of the two views are integrated as the passenger flow trend representation of the starting station; The destination station dynamic feature learning module is used to construct a deep representation of the passenger flow change trend of each station as a destination station, based on two types of real-time observed dynamic passenger flows related to the destination station, and integrating the station proximity relationship, to model the passenger flow change trend of the destination station at different angles from two views, and finally integrate the results of the two views as the passenger flow trend representation of the destination station; The starting station-destination station transfer module is used to construct a transfer matrix from the starting station to the destination station based on the proximity relationship between the stations, and to calculate the passenger flow distribution of the entire network in the next period based on the passenger flow change trends of the constructed starting station and destination station.

Images