A subway station fare area passenger flow distribution short-time prediction method based on big data mining

Abstract

This invention discloses a short-term prediction method for passenger flow distribution in the paid area of ​​a subway station based on big data mining. First, using historical AFC (Automatic Fare Collection) data, the K-shortest-circuit algorithm is employed to calculate the number of paths between origin-destination (OD) pairs, and the AFC data for passengers on a single OD pair is extracted. Second, using this data, a computational graph model is used to calculate the time for walking edges, transfer edges, and waiting nodes. Finally, combined with real-time AFC card swipe data, the passenger flow distribution in the paid area of ​​the station is predicted. The technical solution proposed in this invention is based on the mining of historical big data to calculate the time for walking edges, transfer edges, and nodes in the paid area of ​​the station, which is more accurate and has smaller errors than the current mainstream manual measurement methods. Simultaneously, by utilizing real-time card swipe data and combining the calculation results with historical data, the passenger flow distribution in the paid area is predicted, providing strong data support and scientific basis for passenger service, operation scheduling, and safety emergency response.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of urban rail transit operation control, in particular to a short-time prediction of passenger distribution in a fare zone based on big data mining. BACKGROUND

[0002] As a basic unit and carrier of urban rail transit operation, the safe operation of a station plays an irreplaceable role in the safe operation of a rail transit network. Meanwhile, a station is the starting point of passenger flow inflow and the boundary line of passenger flow outflow. Therefore, accurately depicting the short-time passenger flow distribution in the fare zone of a subway station plays a crucial role in improving the quality of subway operation services, avoiding the occurrence of large passenger flow events, and strengthening subway operation safety.

[0003] Station passenger flow prediction can be divided into two types in physical space: one is passenger flow prediction outside the fare zone, which is relatively mature at home and abroad. There are two prediction ideas: one is to start from the global perspective and mainly use the traditional "four-stage method". The research of related models mainly focuses on traffic mode division and traffic distribution. The other is to focus on local problems, taking a single station as the research object, analyzing the influencing factors of station passenger flow, including the correlation between land use, traffic conditions, resident travel behavior, and station inflow and outflow, obtaining the influencing factor data, and constructing a mathematical model to predict the station passenger flow. The other is passenger flow prediction in the fare zone, i.e., the prediction of passenger flow distribution in the entire platform and station hall before passengers take the train after checking in. At present, there are few studies on this part, but this part of passenger flow distribution is an important object and indispensable part of subway operation services, which has important reference significance and guiding role for avoiding large-scale stampede of passengers in the fare zone of the station and strengthening passenger service strength. SUMMARY

[0004] The present application provides a short-time prediction method for passenger flow distribution in the fare zone of a subway station based on big data mining, which uses historical Automatic Fare Collection (AFC) and train operation timetable to mine the time consumed by each edge and node in the fare zone of the station, and on this basis, combines real-time card swiping data to predict the passenger flow in the fare zone of the station.

[0005] In the present application, the definition of the fare zone is: the area excluding trains that passengers pass through after checking in and before checking out, including platforms, station halls, and transfer areas.

[0006] A short-time prediction method for passenger flow distribution in the fare zone of a subway station based on big data mining, comprising the following steps:

[0007] Step 1: based on historical AFC (Automatic Fare Collection) data, the data of single path passengers between OD (Origin-Destination) pairs is extracted by using K-shortest path algorithm;

[0008] Step 2: the time of walking edges, transfer edges and waiting nodes is calculated by using a computational graph model, specifically including;

[0009] 2.1) constructing a passenger travel cost mathematical optimization model;

[0010] 2.2) vectorizing the passenger travel cost mathematical optimization model;

[0011] 2.3) solving the passenger travel cost mathematical optimization model by using a computational graph model;

[0012] Step 3: according to the time of walking edges, transfer edges and waiting nodes, and in combination with real-time AFC card data, the passenger flow distribution in the station fare area is predicted.

[0013] In the present application, firstly, the number of paths between OD pairs is calculated by using K-shortest path algorithm based on historical AFC data, and the AFC data of single path passengers between OD pairs is extracted; secondly, the time of walking edges, transfer edges and waiting nodes is calculated by using a computational graph model based on the data; and finally, the passenger flow distribution in the station fare area is predicted in combination with real-time AFC card data. The technical scheme proposed in the present application is based on the mining of historical big data, and the time of walking edges, transfer edges and nodes in the station fare area is calculated, which is more accurate and has smaller error than the current mainstream manual measurement method. Meanwhile, the passenger flow distribution in the station fare area is predicted by using real-time card data in combination with the calculation results of historical data, which provides strong data support and scientific basis for passenger service, operation scheduling and safety emergency.

[0014] Step 1: based on AFC historical data, the data of single path passengers between OD pairs is extracted by using K-shortest path algorithm, specifically including:

[0015] 1.1) the number of paths between OD pairs is calculated by using K-shortest path algorithm by traversing AFC historical data;

[0016] 1.2) the data of single path passengers between OD pairs is screened according to the number of paths between OD pairs.

[0017] In step 1.1), K-shortest path algorithm adopts Dijkstra algorithm, Floyd algorithm or A-star algorithm.

[0018] Step 2: the time of walking edges, transfer edges and waiting nodes is calculated by using a computational graph model, specifically including;

[0019] 2.1) constructing a passenger travel cost mathematical optimization model;

[0020] 2.2) vectorizing the passenger travel cost mathematical optimization model to obtain a vectorized passenger travel cost mathematical optimization model;

[0021] 2.3) solving the vectorized passenger travel cost mathematical optimization model by using a computational graph model to obtain the time of a walking edge, a transfer edge and a waiting node.

[0022] A computational graph is a kind of underlying computing framework in machine learning, which is a kind of "language" for describing a computing process, and contains nodes and edges, each node represents a variable or a computing operation on a variable, the variable can be a scalar, a vector, a matrix, etc., and the operation can be a recurrent neural network, a convolutional neural network, a fully connected network, etc., each edge represents the data dependency relationship between variables, that is, the flow relationship between variables, so the computational graph can also be called a data flow graph. The computational graph completely describes the forward propagation process of the variable and the backward propagation process of the error based on the computational graph. There are mainly two ways for calculating the traffic state in the existing traffic field, one is model-driven and the other is data-driven. Considering the complexity of the subway network, the diversity of the influencing factors of passenger travel behavior and the huge amount of passenger flow data of the network, the computational graph model has a particularly prominent advantage in the calculation. Therefore, the use of the computational graph model and the AFC historical data and train operation historical timetable data to solve the vectorized passenger travel cost mathematical optimization model can greatly reduce the calculation amount and improve the calculation efficiency.

[0023] In the present application, the actual subway operation network is described as an expression of graph theory, in which the walking edge refers to the unified name of the two walking areas of passengers from the entrance card swiping to the platform waiting and passengers from the platform to the exit card swiping, the transfer edge refers to the area walked by passengers between OD pairs due to transfer demand, and the waiting node refers to the position node of passengers between OD pairs due to waiting demand.

[0024] In step 3, according to the time of the walking edge, the transfer edge and the waiting node, the passenger flow distribution in the station fare area is predicted in combination with the real-time AFC card swiping data, which specifically includes:

[0025] According to the time of the walking edge, the transfer edge and the waiting node, in combination with the predicted passenger flow of the edge r of the station i in the time period t after the passenger swipes the card to enter the station The following formula is used for calculation:

[0026]

[0027] Wherein, is the passenger flow of the last statistical time, Dr represents the flow of the previous node or the previous section to the section r, Or represents the flow of the section r to the next node or section, is the flow of the section r in the station i in the k time period, and k is a limited division of the t time period, As shown in the following formula:

[0028]

[0029] The time of the walking section, the transfer section and the waiting node in the station payment area is calculated by using big data, and the predicted passenger flow of the section r in the station i in the time period t is calculated according to the real-time AFC card data is equal to the passenger flow in the previous statistical time period , plus the passenger flow of the section r in the station i in the t time period, and minus the passenger flow of the section r in the station i in the t time period, that is, the predicted passenger flow of the section r in the station i in the t time period

[0030] Compared with the prior art, the present application has the following advantages:

[0031] The technical scheme provided in the present application is based on the mining of historical big data, and the time of the walking section, the transfer section and the waiting node in the station payment area is calculated, which is more accurate and has smaller error than the current mainstream manual measurement method.

[0032] The present application uses real-time card data and combines the calculation results of historical data to predict the passenger flow distribution of the payment area, thereby providing strong data support and scientific basis for passenger service, operation scheduling and safety emergency. DETAILED DESCRIPTION

[0033] Figure 1 The present application is a short-time prediction method for passenger flow distribution in a subway station payment area based on big data mining;

[0034] Figure 2 The calculation steps based on the calculation graph model are shown in the following table:

[0035] Figure 3 The calculation graph framework: forward propagation-reverse propagation schematic diagram is shown in the following table:

[0036] Figure 4 The variable forward propagation-reverse propagation schematic diagram is shown in the following table. DETAILED DESCRIPTION

[0037] As Figure 1As shown, the subway station fare area passenger flow distribution short-time prediction method based on big data mining includes: based on historical AFC data, using K-short path algorithm to extract OD pair single path passenger data; secondly, using these data, calculating the walking edge, transfer edge and waiting node time by using the calculation graph framework; finally, combining the real-time AFC card data, predicting the passenger flow distribution in the station fare area.

[0038] As shown in Figure 2 , the calculation steps based on the calculation graph model mainly include: constructing a passenger travel cost mathematical optimization model; vectorizing the passenger travel cost mathematical optimization model; solving the passenger travel cost mathematical optimization model by using the calculation graph model;

[0039] 2.1) Constructing a passenger travel cost mathematical optimization model

[0040] Based on the extraction of OD pair single path passenger AFC data by K-short path algorithm, considering that there is no path selection problem for such passengers, according to the actual travel process of passengers, the generalized travel cost can be defined as the time spent in completing the OD pair process, as shown in formula (1).

[0041]

[0042] Among them is the generalized cost of passenger travel, r represents the entrance, s represents the exit, rs∈P OD , P OD is the set of OD pairs, a∈E is the edge set in the subway network graph G, v∈V is the station node set in the subway network graph G, h∈H=(h1, h2, h3…) is the time interval interval when the passenger enters the subway network, t a is the time consumed by the passenger when passing through the edge segment a, is the waiting time of the passenger at the station node v in the time interval h. α, β and γ are the coefficients of t a and , which are defined as shown in formula (2).

[0043]

[0044] This step mainly estimates the edge travel time and node waiting time by using AFC data. The objective function is to minimize the error between the estimated path travel time and the real path travel time, and the decision variable is the edge travel time and the station waiting time, wherein the edge travel time includes the travel time on the entrance edge, the exit edge, the train running edge and the transfer edge, and the node waiting time includes the waiting time at the starting station and the transfer station. First, define the real path travel time and the estimated path travel time, and then build the optimization model to form the final estimation framework.

[0045] (1) Real travel time

[0046] The path travel time is divided into passenger entry and exit station walking time, waiting time at path starting station and transfer station, train running time and transfer walking time. The entry station, entry time, exit station and exit time recorded in individual card swiping data are used to estimate passenger entry and exit station walking time, waiting time at path starting station and transfer station, train running time and transfer walking time, and then the path generalized cost is obtained for passenger flow distribution.

[0047] For each trip of passenger i, the entry station is r, the entry time is the exit station is s, and the exit time is Therefore, the real travel time is shown in formula (3).

[0048]

[0049] There are multiple trip records in the same OD pair in the same time period. Due to different entry times, walking speeds and other reasons, the trip records usually have different travel times. In order to reduce the calculation amount, the average travel time of all trip records of the OD pair in the time period is selected as the real travel time of the OD pair in the time period, as shown in formula (4).

[0050]

[0051] wherein, is the number of trip records of OD pair rs in time period h, N is the number of nodes of the subway network, and h∈H is the time period divided in a day. is the real travel time of the i-th trip record of the specific OD pair rs in the time period h.

[0052] (2) Estimate path travel time

[0053] For the OD pair of a single path, there is no path selection probability, and the estimated path travel time is the generalized cost of passenger completing an OD trip

[0054] (3) Build optimization model

[0055] The objective function is to minimize the error between the real travel time and the expected travel time of individual trips, as shown in formula (5).

[0056]

[0057] Substitute formula (1) and formula (4) into formula (5) to obtain formula (6), which is a model objective function for minimizing the error between the actual path travel time and the expected path travel time of individual passengers, and the decision variable is the time t consumed by passengers when passing through edge segment a a , the waiting time of passengers at the station node v By optimizing the objective function, the final t a is obtained, which is used for subsequent path calibration and cross-sectional flow acquisition.

[0058]

[0059]

[0060] 2.2) Vectorization of passenger travel cost mathematical optimization model

[0061] The above optimization model is placed in the computational graph framework to obtain t a , Before modeling the computational graph model, all variables in the optimization model need to be vectorized and then embedded in the computational graph, and then a supervised learning task under the computational graph framework is constructed with the help of deep learning frameworks such as PyTorch. The variables involved in formula (6) are explained one by one as follows.

[0062] (1) Real path travel time Vectorized representation

[0063] Based on formula (4), the vectorized representation of the real path travel time is shown in formula (7).

[0064]

[0065] Where N is the number of subway network nodes, h∈H is the time period divided in a day, is the real travel time of a specific OD pair rs in time period h, is the real travel time of all OD pairs in all time periods, which is also the vectorized real path travel time variable required in this section, N rs is the number of all OD pairs in the network.

[0066] (2) Edge vectorization representation

[0067] The travel time of all edges in this section is independent of the time period, that is, it is assumed that the travel time of all edges is the same in different time periods. The number of inbound edges, outbound edges, train running edges and transfer edges in the network is N a , the number of network nodes is N, and the vectorized representation of the edges of the OD pairs is shown in formula (8) and (9).

[0068]

[0069]

[0070] where, is the vectorization of the edge between OD pair rs, and is the vectorization of the edge between OD pair rs, and is the vectorization of the edge between OD pair rs, and N is the number of OD pairs in the network. rs is the number of OD pairs in the network.

[0071] (3) Vectorization of nodes

[0072] The waiting time of all nodes is related to the time period, i.e., the waiting time of the same node in different directions is different in different time periods. Let the number of waiting nodes in different directions of all nodes in the network be N v , then the vectorization of the waiting nodes between OD pairs is shown in Equations (10) and (11).

[0073]

[0074]

[0075] where, and is the vectorization of the edge between OD pair rs, and is the vectorization of the waiting nodes on the path between OD pair rs in time period h, is the vectorization of the waiting nodes on the path between OD pair rs in time period h, and N is the number of OD pairs in the network. rs is the number of OD pairs in the network. v is the number of waiting nodes in different directions of all nodes in the network.

[0076] (4) Joint vectorization of edges and waiting nodes

[0077] The B and M h in different time periods are jointly vectorized as shown in Equation (12).

[0078]

[0079] where, is the joint vectorization of edges and waiting nodes, and N is the number of OD pairs in the network. rs is the number of OD pairs in the network.

[0080] (5) Vectorization of the estimated edge travel time t a

[0081] The vectorization of the travel time of all edges is shown in Equation (13).​

[0082]

[0083] where is one of the decision variables, t a where a represents all segments.

[0084] (6) The vectorized representation of the waiting time of the nodes to be estimated Vectorized representation

[0085] The vectorized representation of the waiting time of all nodes is shown in equation (14)

[0086]

[0087] where is one of the decision variables, the waiting time of the waiting node is related to the time period h, where v represents all nodes.

[0088] (7) Joint vectorized representation of segment travel time and node waiting time:

[0089] The joint vectorization of and h in different time periods is shown in equation (15)

[0090]

[0091] where, is the joint vectorized representation of segment travel time and node waiting time.

[0092] (8) Vectorized representation summary

[0093] All the above scalar vectorized representations are summarized in Table 1:

[0094] Table 1 Vectorized representation of variables

[0095]

[0096] (9) Vectorized representation of optimization model

[0097] Based on Table 1, the vectorized representation of equation (6) is shown in equation (4-16).

[0098]

[0099] s.t.t≥0

[0100] The main decision variable of the above optimization model is t, which contains the travel time of all segments and the waiting time of nodes in different time periods, and when the matrix A column is full rank, the optimization problem has a unique feasible solution.

[0101] 2.3) Solving the passenger travel cost mathematical optimization model by using the computational graph model;

[0102] Solving the above optimization problem by using the computational graph framework, by regarding the optimization problem as a machine learning task and placing it under the computational graph framework, and by using the forward propagation-backward propagation algorithm, the optimal solution of the optimization problem is obtained, Figure 3 The computational graph framework of the optimization problem is shown, which includes two parts, the forward propagation part and the backward propagation part.

[0103] The forward propagation part is shown by the solid line in the figure. The process assumes that the node waiting time and the arc travel time are fixed, calculates the estimated OD travel time by calculating the path travel time, compares the estimated OD travel time with the observed OD travel time, and obtains the estimation error. The backward propagation part propagates the error back to the original path to adjust the node waiting time and the arc travel time, thus completing one iteration process. Such iterative calculation is repeated until the iteration stops, and the final estimated node waiting time and arc travel time are obtained for subsequent cross-section short-term passenger flow prediction.

[0104] Figure 4 For the variable propagation process of the framework, the solid line represents forward propagation and the dashed line represents backward propagation. Through the process of continuously iterating the variable forward propagation and error backward propagation, the final is obtained, which is composed of node waiting time and arc travel time. The forward propagation and backward propagation processes are described in detail below.

[0105] Forward propagation: with randomly initialized node waiting time and edge travel time t a as the model input, and combined with the path representation A, the estimated OD travel time c, the whole process is shown in formula (17).

[0106] c = At (17)

[0107] The final goal of the forward propagation process is to obtain the estimated OD travel time c. Regarding the objective function in formula (16) as the loss function L (Loss function) in the machine learning task, it can be decomposed as formula (18).

[0108]

[0109] c = At

[0110] where, is the observed average OD travel time obtained from the AFC data, and c is the estimated OD travel time.

[0111] Back propagation: The path estimation time is obtained according to the initialized node and arc travel time, and the objective function is written as formula (19), and the gradient of the objective function can be obtained, as shown in formula (20).

[0112]

[0113]

[0114] Integrating formula (20), the gradient of the objective function with respect to time can be obtained, as shown in formula (21).

[0115] After obtaining the gradient of the objective function, the decision variable t can be optimized by using the back propagation algorithm.

[0116]

[0117] The forward propagation and back propagation are iterated until the approximate optimal solution of the optimization problem shown in formula (16) is obtained.

Claims

1.A metro station fare zone passenger flow distribution short-time prediction method based on big data mining, characterized in that, The method comprises the following steps: Step 1: based on historical AFC data, AFC data of passengers with single path between OD pairs is extracted by using K-shortest path algorithm, specifically comprising: 1.1) AFC historical data is traversed, and the number of paths between OD pairs is calculated by using K-shortest path algorithm; 1.2) data of passengers with single path between OD pairs is screened according to the number of paths between OD pairs; Step 2: time of walking edges, transfer edges and waiting nodes is calculated by using a computation graph model, specifically comprising: 2.1) a passenger travel cost mathematical optimization model is constructed; 2.2) the passenger travel cost mathematical optimization model is vectorized to obtain a vectorized passenger travel cost mathematical optimization model; 2.3) the vectorized passenger travel cost mathematical optimization model is solved by using a computation graph model to obtain time of walking edges, transfer edges and waiting nodes; Step 3: according to the time of walking edges, transfer edges and waiting nodes, and in combination with real-time AFC card swiping in and out of station data, station fare collection area passenger flow distribution is predicted, specifically comprising: According to the time of the running side, the transfer side and the waiting node, in combination with the passenger instant card swiping into the station, the predicted passenger flow of the station side section in the time period is calculated The following formula is used for calculation:​​​ ； wherein, is the passenger flow at the last statistical time, denotes the flow into the station at the t period, edge segment passenger flow, Dr denotes the flow from the last node or the last edge segment to the edge segment r, denotes the flow out of the station at the t period, edge segment passenger flow, Or denotes the flow out of the edge segment r to the next node or edge segment, is the flow of the station selected at the time period t, edge segment k is the finite partition of the statistical t period, as shown in the following formula: 。 2.The metro station fare zone passenger flow distribution short-time prediction method based on big data mining of claim 1, wherein, In step 1.1), the K-shortest path algorithm adopts Dijkstra algorithm, Floyd algorithm or A-star algorithm.

Images