Subway OD flow prediction method
By identifying date and false types, obtaining OD feature traffic, and constructing and extracting spatial and temporal dynamic features, the problem of dynamic prediction of OD traffic between subway stations is solved, and higher prediction accuracy and robustness are achieved.
Patent Information
- Application Number
- CN202510115797.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-05-27
AI Technical Summary
It is difficult for the prior art to accurately predict OD traffic dynamics between subway stations, especially in peak hours, emergencies and uneven passenger distribution. Simple site-level traffic prediction is difficult to meet the needs of prediction accuracy and traffic scheduling optimization.
A subway OD traffic prediction method is proposed. By identifying the date and false traffic types of the time period to be predicted, the OD feature traffic related to it is obtained, the site's adjacency matrix is constructed, the spatial and temporal dynamic features of the three-dimensional matrix are extracted, and the three-dimensional matrix is spliced for prediction.
This method can effectively deal with the complexity of subway OD traffic and improve the robustness and accuracy of prediction results, especially when dealing with differences in working days and holiday modes and differences in various sites' functions.
Smart Images
Figure CN120046782A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the technical field of subway OD flow prediction, and particularly relates to a subway OD flow prediction method. Background Art
[0002] In subway passenger flow prediction, the prediction of station-level passenger flow provides important support for the operation optimization of the subway system, but it cannot comprehensively reveal the travel paths and interaction patterns of passengers between different stations. In a complex subway network, especially in the face of peak hours, emergencies, and uneven passenger distribution, it is difficult to accurately grasp the flow dynamics between stations by simply relying on station-level flow prediction. Therefore, the origin-destination (OD) flow prediction is of greater significance in further improving the prediction accuracy and optimizing traffic dispatching.
[0003] Compared with the prediction of station-level passenger flow, OD flow prediction faces higher complexity, including the sparsity of the flow matrix, the diversity of spatio-temporal dependence relationships, and the impact of uncertain events on the flow. These problems not only increase the difficulty of data processing but also pose higher requirements for model design. Summary of the Invention
[0004] In view of this, the purpose of this application is to provide a subway OD flow prediction method, which fully considers the pattern differences between weekdays and holidays, as well as the correlation differences caused by different functions of each station, can effectively cope with the complexity of subway OD flow, and improve the robustness and accuracy of the prediction results.
[0005] This application provides a subway OD flow prediction method, including:
[0006] Identifying the date of the period to be predicted and the type of work attendance on that date, and obtaining the OD characteristic flow having a dependence relationship with the period to be predicted; wherein, the type of work attendance includes: holidays and weekdays;
[0007] For each station, extracting the OD characteristic flow from this station to each target station to construct the adjacency matrix of this station; wherein, the adjacency matrix is used to represent the flow interaction relationship between this station and each target station;
[0008] Constructing a three-dimensional adjacency matrix based on the adjacency matrices of each station, and extracting the spatial dynamic characteristics of the three-dimensional adjacency matrix through Einstein summation convention and gated linear unit;
[0009] Extracting the time dynamic characteristics through trend decomposition based on the OD characteristic flow from each station to each target station;
[0010] Concatenate the spatial dynamic features and the temporal dynamic features, and use the concatenated spatio-temporal dynamic features for subway OD flow prediction.
[0011] Further, obtaining the OD feature flows having a dependency relationship with the to-be-predicted time period based on the identified types of false attendance includes:
[0012] Based on the OD flows of each station, obtain the real-time truncated OD flows of the time periods adjacent to the to-be-predicted time period within this date, the outstanding OD flows of the to-be-predicted time period, and obtain the full-cycle OD flows of the same time period of the previous day and the full-cycle OD flows of the same time period of the previous week within the dates of the same type of false attendance;
[0013] Take the obtained real-time truncated OD flows, outstanding OD flows, full-cycle OD flows of the same time period of the previous day, and full-cycle OD flows of the same time period of the previous week as the OD feature flows having a dependency relationship with the to-be-predicted time period.
[0014] Further, the target station is determined in advance by the following method:
[0015] Obtain historical subway riding records and extract the passenger entry and exit information of each day;
[0016] For each station, respectively count the OD flows from this station to other stations, and sort the other stations according to the OD flows to select multiple target stations with higher OD flows.
[0017] Further, the adjacency matrix of this station is constructed by the following method:
[0018] Obtain the OD feature flows from this station to any two target stations respectively to obtain the flow sequences from this station to any two target stations;
[0019] Calculate the DTW distance between the flow sequences from this station to any two target stations respectively to obtain the correlation weight between this station and any two target stations;
[0020] Take the correlation weight between this station and any two target stations as matrix elements to construct the adjacency matrix of this station.
[0021] Further, extracting the spatial dynamic features of the three-dimensional adjacency matrix through Einstein summation convention and gated linear unit includes:
[0022] Based on the three-dimensional adjacency matrix, sum the matrix elements of each target station dimension respectively to obtain the first spatial feature representing the OD flow characteristics of each target station;
[0023] The dimension of the first spatial feature is mapped to twice using a linear transformation, and the first spatial features before and after mapping are respectively used as gates after being activated by sigmoid and directly as inputs, and then the GLU activation function is used to obtain the spatial dynamic feature.
[0024] Further, the time dynamic features are extracted by trend decomposition based on the OD feature flows from each station to each target station, including:
[0025] Based on the OD feature flows from each station to each target station, the OD feature flows of the time series are obtained;
[0026] The long-term trend of the OD feature flows of the time series is removed using first-order difference, and a filling operation is performed on the short-term fluctuations of the remaining OD feature flows of the time series to obtain the first-order difference features;
[0027] The long-term trend features of the OD feature flows of the time series are extracted through average pooling operation, and the short-term fluctuation features of the OD feature flows of the time series are extracted through max pooling operation, and the difference between the long-term trend features and the short-term fluctuation features is calculated to obtain the residual features;
[0028] The OD feature flows of the time series, the first-order difference features, the short-term fluctuation features, and the residual features are concatenated to obtain the time dynamic features.
[0029] The subway OD flow prediction method provided by this application fully considers the pattern differences between weekdays and holidays, as well as the correlation differences caused by the different functions of each station, can effectively handle the complexity of subway OD flows, and improve the robustness and accuracy of the prediction results. Brief Description of the Drawings
[0030] Figure 1 Shows the flowchart of the subway OD flow prediction method provided by the embodiments of this application. Detailed Description of the Embodiments
[0031] To make the objectives, technical solutions, and advantages of this technical solution clearer, the following further details this technical solution in conjunction with specific embodiments. It should be understood that these descriptions are exemplary and not intended to limit the scope of this technical solution.
[0032] Please refer to Figure 1 the flowchart of the subway OD flow prediction method provided by the embodiments of this application as shown. As Figure 1 shown, the method includes:
[0033] S101. Identify the date of the to-be-predicted period and the absenteeism type of this date, and obtain the OD feature flow that has a dependency relationship with the to-be-predicted period based on the identified absenteeism type.
[0034] Among them, the absenteeism types include: holidays and working days.
[0035] In this step, considering the temporal correlation of subway OD flow, that is, there are obvious pattern differences in the intra-day changes of subway passenger flow on working days and holidays. Specifically, the flow peaks on working days are concentrated in the morning and evening commuting periods, forming a bimodal pattern, while the flow fluctuations on holidays are relatively gentle, and the peak times are relatively scattered. Therefore, in the embodiments of the present application, by distinguishing the absenteeism type of the date where the to-be-predicted period is located, that is, identifying whether this date is a holiday or a working day, it is ensured that historical data with the same absenteeism type as the date where the to-be-predicted period is located can be obtained. In other words, if the date where the to-be-predicted period is located is a working day, the historical data of the previous day and the previous week must also come from working days. When the historical data of the previous day or the previous week is a holiday, it is necessary to continue searching forward until the closest historical data that is also a working day is found. This matching of absenteeism types ensures that the prediction of subway OD flow will not be disturbed by the differences in historical data of different absenteeism types, and can more accurately obtain the OD feature flow that has a dependency relationship with the to-be-predicted period from historical data.
[0036] In specific implementation, the OD feature flow that has a dependency relationship with the to-be-predicted period can be obtained based on the identified absenteeism type in the following way:
[0037] Step 1011. Based on the OD flow of each station, obtain the real-time truncated OD flow of the time period adjacent to the to-be-predicted period within this date, the outstanding OD flow of the to-be-predicted period, and the full-cycle OD flow of the same time period of the previous day and the full-cycle OD flow of the same time period of the previous week within the dates of the same absenteeism type.
[0038] In this step, the real-time truncated OD flow refers to the passengers who have completed getting off the train within a specific time period. That is, the real-time truncated flow provides a "partial" data perspective and can only reflect the passengers who have completed the whole journey within the given time window. The outstanding OD flow refers to the passengers who have entered the station but have not yet gotten off the train within a certain time period. That is, by capturing the ongoing journey, it provides an early warning for future OD flow. The full-cycle OD flow refers to the complete travel records of all passengers from entering the station to getting off the train within a certain time period. The key concept of this is "completeness", that is, regardless of when the passengers get off the train, as long as they enter the station within the specified time period, no matter when their journey ends, they will be included in the full-cycle flow.
[0039] Step 1012: Use the obtained real-time truncated OD flow, pending OD flow, full-cycle OD flow of the same period of the previous day, and full-cycle OD flow of the same period of the previous week as OD feature flows that have a dependency relationship with the period to be predicted.
[0040] In this step, considering that the OD flow of the period to be predicted has a high similarity with the flow patterns of the same period of the previous day and the same period of the previous week, especially during the morning and evening rush hours on weekdays, this similarity is more significant. Therefore, the OD flow of the previous day is used to reflect short-term periodic changes, and the OD flow of the previous week is used to reflect longer-term periodic changes.
[0041] Before extracting the OD feature flows from this station to each target station to create the adjacency matrix of this station, the method further includes: determining the target stations in advance by the following method:
[0042] Step 201: Obtain historical subway riding records and extract the daily passenger entry and exit information.
[0043] In this step, the historical subway riding records include important information related to OD flow statistics, such as: card swiping time, passenger ID card number, entry and exit status, the station where the transaction occurs, and unimportant information, such as: traffic type, transaction nature, line ID, device ID, transaction nature, etc. In the embodiments of the present application, the entry and exit stations and times of each passenger are extracted from the historical subway riding records. Also, since for the subway system, the data of each day is a separate cycle flow, it is necessary to separately count the daily passenger entry and exit information.
[0044] Step 202: For each station, separately count the OD flow from this station to other stations, and sort the other stations according to the OD flow to select multiple target stations with higher OD flow.
[0045] In this step, considering the spatial correlation of subway OD flow, that is, different stations have different roles and importance in the entire subway network, which is specifically manifested as: 1) Most OD flows are concentrated in a few destination stations, which reflects the existence of main passenger flow corridors in the subway network; 2) The flow of most OD pairs (from the starting station to the destination station) is very small, and only a few OD pairs contribute the main flow. Therefore, in the embodiments of the present application, for each station, only the stations with higher OD flow are selected as target stations, and further historical data of the target stations are extracted for prediction, while the remaining stations with smaller OD flow are combined into a whole for prediction. This method improves the calculation efficiency by reducing the number of stations to be predicted and the amount of historical data processing, and also ensures the prediction accuracy by retaining the main OD flow.
[0046] S102. For each station, extract the OD characteristic traffic from this station to each target station to create the adjacency matrix of this station.
[0047] Among them, the adjacency matrix is used to characterize the traffic interaction relationship between this station and each target station.
[0048] In this step, on the basis of considering the time correlation of subway OD traffic, the spatial correlation of subway OD traffic is further considered, that is, extract the OD characteristic traffic from this station to each target station from the obtained OD characteristic traffic that has a dependency relationship with the period to be predicted; then create the adjacency matrix of this station based on the extracted OD characteristic traffic from this station to each target station.
[0049] In specific implementation, the adjacency matrix of this station can be constructed in the following way:
[0050] Step 1021. Obtain the OD characteristic traffic from this station to any two target stations respectively to obtain the traffic sequences from this station to any two target stations respectively.
[0051] Step 1022. Calculate the DTW distance between the traffic sequences from this station to any two target stations respectively to obtain the correlation weight between this station and any two target stations respectively.
[0052] Step 1023. Use the correlation weight between this station and any two target stations respectively as matrix elements to construct the adjacency matrix of this station.
[0053] The process of constructing the adjacency matrix of this station based on the above steps 1021 - 1023 can be mapped to a mathematical function. As an example, first, set this station as Ori and set the number of target stations as 4, and set any two target stations as T j and T k , which represent the jth target station and the kth target station respectively; second, the traffic sequences from this station Ori to any two target stations T j and T k are defined as OD(Ori, T j ) and OD(Ori, T k ), then the DTW distance between the traffic sequences from this station to any two target stations is DTW(OD(Ori, T j ), OD(Ori, T k )); generally, the smaller the DTW distance, the higher the similarity between the two traffic sequences; third, the traffic sequence OD(Ori, T j) and the correlation weight between OD (Ori, T k ) Finally, the correlation weights between each pair of traffic sequences are used as matrix elements to construct the adjacency matrix of the site.
[0054]
[0055] In the formula, σ is a parameter used to adjust the sensitivity of similarity. Generally, a smaller σ value makes the similarity more sensitive to changes in distance.
[0056] S103. Construct a three-dimensional adjacency matrix based on the adjacency matrix of each site, and extract the spatial dynamic features of the three-dimensional adjacency matrix through Einstein summation convention and gated linear unit.
[0057] In this step, first, a three-dimensional adjacency matrix Adj OD is constructed based on the adjacency matrix AdjOD[Ori] of each site, where each site is all sites in the subway network to be predicted; then, Einstein summation convention and gated linear unit are used together to form a graph convolution module for extracting the spatial dynamic features of the three-dimensional adjacency matrix.
[0058] In specific implementation, the spatial dynamic features of the three-dimensional adjacency matrix can be extracted through the following method by Einstein summation convention and gated linear unit:
[0059] Step 1031. Based on the three-dimensional adjacency matrix, sum the matrix elements of each target site dimension respectively to obtain the first spatial feature for characterizing the OD traffic feature of each target site.
[0060] Step 1032. Use linear transformation to map the dimension of the first spatial feature to twice, and use the first spatial feature before and after mapping as the gate after sigmoid activation respectively, and directly as the input, and then use the GLU activation function to obtain the spatial dynamic feature.
[0061] The steps for extracting the spatial dynamic features of the three-dimensional adjacency matrix based on the above steps 1031 - 1032 can be mapped to a mathematical function as shown in formula (2):
[0062]
[0064] In the formula, represents the output feature tensor from site i to the jth target site, that is, the OD traffic feature, is the element of the input tensor H of the lth layer, representing the OD traffic hidden feature from site i to the kth target site, It represents the correlation degree between the OD characteristic flow from site i to the jth target site and the OD characteristic flow from site i to the kth target site.
[0065] In the above formula (2), by summing the matrix elements in the k dimension (i.e., the target site dimension), the output H of each site ij will take into account the input characteristics between site i and each target site, as well as the similarity relationship between target sites.
[0066] Among them, the GLU activation function in the above formula (2) is shown as the following formula (3):
[0067] GLU(X) = X 1 ⊙σ(X 2 ); (3)
[0069] In the formula, where X 1 and X 2 are two parts of the input characteristics (i.e., the first spatial characteristics), usually calculated separately through a linear layer, and σ is the sigmoid activation function.
[0070] The above formula (3) maps the input characteristics to twice the original characteristic dimension through a linear transformation, and then divides it into two parts. One part is used as a gate after being activated by sigmoid, and the other part is directly used as an input. Compared with a simple non - linear function, the gating mechanism of GLU can adaptively select which node characteristics need to be enhanced and which characteristics should be suppressed, thereby improving the expressive ability of graph convolution.
[0071] S104. Based on the OD characteristic flows from each site to each target site, extract time - dynamic characteristics through trend decomposition.
[0072] In this step, the OD characteristic flows often contain both long - term trends and short - term fluctuations. Processing only one type of these characteristics may lead to information loss or a decrease in prediction accuracy. To solve this problem, the embodiments of this application decompose the long - term trends and short - term fluctuations, then capture the long - term trend characteristics and short - term fluctuation characteristics respectively, and splice all the characteristics to form a complete time - dynamic characteristic.
[0073] In specific implementation, the following method can be used to extract time - dynamic characteristics through trend decomposition based on the OD characteristic flows from each site to each target site:
[0074] Step 1041. Based on the OD characteristic flows from each site to each target site, obtain the OD characteristic flows of the time series.
[0075] Step 1042: Remove the long-term trend of the OD feature flow of the time series using first-order difference, and perform a filling operation on the short-term fluctuations of the OD feature flow of the remaining time series to obtain the first-order difference feature.
[0076] Step 1043: Extract the long-term trend feature of the OD feature flow of the time series through average pooling operation, and extract the short-term fluctuation feature of the OD feature flow of the time series through max pooling operation, and calculate the difference between the long-term trend feature and the short-term fluctuation feature to obtain the residual feature.
[0077] Step 1044: Concatenate the OD feature flow of the time series, the first-order difference feature, the short-term fluctuation feature, and the residual feature to obtain the time dynamic feature.
[0078] The process of extracting the time dynamic feature based on the above steps 1041-1044 can be mapped to a mathematical function. Specifically: 1) The long-term trend changes in the time series data often mask the short-term dynamic changes, thus affecting the capture of short-term dynamic features in the prediction process. Therefore, the long-term trend of the OD feature flow of the time series is removed through first-order difference to focus on the short-term fluctuations of the OD feature flow of the time series.
[0079] As an example, for the OD feature flow X(t) of the time series, calculate the difference ΔX(t) between adjacent time steps through first-order difference as shown in formula (4):
[0080] ΔX(t) = X(t) - X(t - 1); (4)
[0082] Here, since the number of time steps after the first-order difference is reduced by one, in order to keep the data dimension after the first-order difference consistent with the original data dimension, the difference result of the last time step is usually used to fill the result after the first-order difference to obtain the first-order difference feature as shown in formula (5):
[0083] ΔX padded = [ΔX 2 ,..., ΔX T , ΔX T . (5)
[0084] 2) The core purpose of trend decomposition is to extract features of different time scales from the time series data. On the one hand, through the sliding window method, use the average pooling algorithm shown in formula (6) below to calculate the average value T avg (t) of the OD feature flow X(t) of the time series in a relatively long time period to extract the long-term trend feature;
[0085]
[0086] On the other hand, first obtain the peak feature T of the OD feature flow X(t) of the time series max (t); then, in the way of a sliding window, use the max pooling algorithm shown in the following formula (7) to calculate the local maximum value S of the OD feature flow X(t) of the time series max (t) to extract short-term fluctuation features;
[0087] S max (t) = T max (t) - X(t); (7)
[0089] After extracting the long-term trend features and short-term fluctuation features, calculate the average value through the following formula (8)
[0090] T avg (t) and the difference between the local maximum value S max (t), that is, calculate the difference between the long-term trend features and short-term fluctuation features to obtain residual features;
[0091] T residual (t) = T max (t) - T avg (t); (8)
[0093] 3) Concatenate the OD feature flow X(t), the first-order difference feature ΔX padded (t), the short-term fluctuation feature S max (t), and the residual feature T residual (t) of the time series to form a complete multi-dimensional feature representation, and use this feature representation as the time dynamic feature.
[0094] S105. Concatenate the spatial dynamic feature and the time dynamic feature, and use the concatenated spatio-temporal dynamic feature for subway OD flow prediction.
[0095] In this step, since the concatenated spatio-temporal dynamic feature fully considers the pattern differences between weekdays and holidays, and the correlation differences caused by the different functions of each station, therefore, based on the concatenated spatio-temporal dynamic feature, the complexity of subway OD passenger flow can be effectively dealt with, and the robustness and accuracy of the prediction result can be improved.
[0096] The above content is only the preferred embodiment of the present invention. For those of ordinary skill in the art, according to the idea of the technical content of the present invention, many changes can be made in the specific implementation manner and application scope. As long as these changes do not deviate from the concept of the present invention, they all belong to the protection scope of this patent.
Claims
1. A subway OD flow prediction method, characterized in that: The method comprises: Identify the date of the time period to be predicted and the type of leave and attendance on the date, and obtain the OD characteristic flow that has a dependency relationship with the time period to be predicted; wherein the leave and attendance type includes: holidays and working days; For each site, extract the OD characteristic flow from the site to each target site to construct the adjacency matrix of the site; wherein the adjacency matrix is used to characterize the flow interaction relationship between the site and each target site; A three-dimensional adjacency matrix is constructed based on the adjacency matrix of each site, and spatial dynamic features of the three-dimensional adjacency matrix are extracted through the Einstein summation convention and the gated linear unit; Based on the OD characteristic flow from each station to each target station, the temporal dynamic characteristics are extracted through trend decomposition; The spatial dynamic features and the temporal dynamic features are spliced together, and the spliced spatial and temporal dynamic features are used to predict the subway OD flow.
2. The method according to claim 1, characterized in that The step of obtaining the OD characteristic flow rate having a dependency relationship with the time period to be predicted based on the identified absence and attendance type includes: Based on the OD flow of each site, obtain the real-time cut-off OD flow of the time period adjacent to the time period to be predicted and the pending OD flow of the time period to be predicted on the date, and obtain the full-cycle OD flow of the same time period of the previous day and the full-cycle OD flow of the same time period of the previous week on the date of the same leave type; The acquired real-time cut-off OD flow, pending OD flow, full-cycle OD flow in the same period of the previous day, and full-cycle OD flow in the same period of the previous week are taken together as OD characteristic flows having a dependency relationship with the period to be predicted.
3. The method according to claim 1, characterized in that The target site is predetermined by: Obtain historical subway ride records and extract daily passenger entry and exit information; For each site, the OD flows from the site to other sites are counted respectively, and other sites are sorted according to the OD flows, so as to select multiple target sites with higher OD flows.
4. The method according to claim 1, characterized in that The adjacency matrix for this site is constructed as follows: Obtain the OD characteristic flow from the site to any two target sites respectively, so as to obtain the flow sequence from the site to any two target sites respectively; Calculate the DTW distance between the traffic sequences from the site to any two target sites to obtain the correlation weights from the site to any two target sites; The weights of the associations between the site and any two target sites are used as matrix elements to construct the adjacency matrix of the site.
5. The method according to claim 1, characterized in that The method of extracting the spatial dynamic features of the three-dimensional adjacency matrix by using the Einstein summation convention and the gated linear unit includes: Based on the three-dimensional adjacency matrix, the matrix elements of each target site dimension are summed up respectively to obtain a first spatial feature for characterizing the OD flow characteristics of each target site; The dimension of the first spatial feature is mapped to two times by linear transformation, and the first spatial features before and after mapping are respectively activated by sigmoid as gates and directly used as inputs, and then the GLU activation function is used to obtain the spatial dynamic features.
6. The method according to claim 1, characterized in that The temporal dynamic characteristics are extracted by trend decomposition based on the OD characteristic flow from each station to each target station, including: Based on the OD characteristic flow from each station to each target station, the OD characteristic flow of the time series is obtained; The long-term trend of the OD characteristic flow of the time series is removed by using the first-order difference, and the short-term fluctuation of the OD characteristic flow of the retained time series is filled to obtain the first-order difference feature; Extracting the long-term trend feature of the OD feature flow of the time series through an average pooling operation, extracting the short-term fluctuation feature of the OD feature flow of the time series through a maximum pooling operation, and calculating the difference between the long-term trend feature and the short-term fluctuation feature to obtain a residual feature; The OD characteristic flow, first-order difference characteristics, short-term fluctuation characteristics and residual characteristics of the time series are spliced to obtain the time dynamic characteristics.