Subway multi-task passenger flow prediction method fusing space-time multi-dimensional features

Abstract

The invention discloses a subway multi-task passenger flow prediction method fusing time-space multi-dimensional features, and the method comprises the steps: preprocessing original passenger flow data collected by an AFC system, and obtaining a passenger flow distribution matrix of all stations of a rail transit network; collecting multi-dimensional features, including external dynamic environment and external geographical environment features, station space relationships, periodic hour codes, date features and peak code features; constructing an end-to-end multi-dimensional feature integration model CompLSTM (CompGCN-Attention-LSTM), combining a multi-graph adjacency matrix, and integrating origin-destination passenger flow, a station space connection relationship, a station distance relationship and a station function relationship; and adopting Attention-LSTM (Long Short Term Memory) to mine space-time characteristics of the passenger flow matrix fused with the multi-dimensional characteristics. According to the method, multi-dimensional features are fused, an end-to-end integrated model is constructed, spatio-temporal data characteristics are excavated comprehensively, the subway network multi-station origin-destination short-time passenger flow and the in-out short-time passenger flow can be accurately predicted, line expansion planning and network optimization design are facilitated, a subway operation strategy is optimized, and intelligent management of an urban rail transit network is achieved.

Description

Technical Field

[0001] The present invention relates to a subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features, and belongs to the field of artificial intelligence technology. Background Art

[0002] With the acceleration of urbanization, traffic congestion is becoming increasingly serious. As a high-capacity, highly efficient public transportation mode, rail transit plays an increasingly important role in urban transportation. Currently, my country's urban rail transit continues to develop steadily, with the network size and passenger volume ranking first globally. However, passenger flow in and out of urban rail transit networks is significantly affected by multiple factors, including geographic environment, operating model, urban spatial structure, and temporal fluctuations. Predicting short-term inbound and outbound passenger flow at each station in the rail transit network can help operational managers promptly identify and address peak congestion and sudden large passenger flow surges, effectively improving passenger transportation organization and management, and implementing targeted safety control measures to reduce the risk of accidents. Origin-destination passenger flow forecasting can identify passenger travel patterns, clarify travel purposes, and identify station functions. Origin-destination passenger flow forecasting can accurately identify differences in travel demand between stations, providing a key basis for optimizing route planning, improving service efficiency, and enhancing the passenger travel experience.

[0003] According to statistics, urban rail transit passenger flow is showing a rapid growth trend, especially during peak hours in the morning and evening. Rail transit carries a large number of urban commuters, increasing the network's operational load and placing higher demands on operational scheduling, service quality, and passenger experience. However, current rail transit passenger flow forecasting still has significant shortcomings: First, existing methods can generally only predict passenger flow at a single station, failing to fully cover all stations in the urban rail transit network; second, the forecasting task is single-minded, with most research methods only forecasting passenger flow at inbound and outbound stations or at origin and destination points, failing to predict both inbound and outbound passenger flow and origin and destination passenger flow simultaneously; third, the impact of multidimensional spatiotemporal characteristics on passenger flow is not fully considered, resulting in low prediction accuracy.

[0004] In fact, passengers' travel decisions are often influenced by multiple factors, including morning and evening peak hours, date characteristics, external environment, travel purpose, etc. Therefore, the introduction of multi-dimensional feature data is of great significance for improving passenger flow prediction performance. Summary of the Invention

[0005] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned existing technologies and propose a subway multi-task passenger flow prediction method that integrates spatiotemporal multi-dimensional features. This method can fully integrate multiple data sources and use advanced artificial intelligence technology to make high-precision predictions of the inbound and outbound passenger flow and the passenger flow at the starting and ending points of the rail transit network. This not only helps subway operators optimize scheduling and operation strategies, but also improves passengers' travel experience, reduces congestion, and thus achieves efficient and intelligent management of urban rail transit. By comprehensively utilizing multi-source data such as historical passenger flow data, weather conditions, traffic events, and urban activities, the present invention can more accurately capture the spatiotemporal variation patterns of passenger flow, thereby improving prediction accuracy and providing strong support for the refined management and sustainable development of the rail transit network.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a subway multi-task passenger flow prediction method integrating spatiotemporal multi-dimensional features, the method comprising the following steps:

[0007] Step 1: Preprocess the raw passenger flow data collected from the AFC system, including outlier processing and time granularity selection, to obtain historical subway passenger flow OD data and construct a passenger flow OD matrix. Data on external factors influencing subway passenger flow are obtained, primarily including spatial and temporal coding features. Spatial features include geographic environment characteristics, station distance matrix, station topology, and station function relationships. Temporal coding features include date characteristics, periodic hour codes, peak codes, and weather characteristics.

[0008] Step 2: Build an end-to-end multidimensional feature integration model, CompLSTM (CompGCN-Attention-LSTM). The CompGCN module processes time-coded subway passenger flow O / D data, station operation characteristics, station topology, and station distance matrix. The Attention-LSTM module further explores the relationship between passenger flow and other spatiotemporal data. The LSTM is used to process peak-coding features and some time series external dynamic environmental features, such as weather characteristics. The feature fusion layer processes these results along with external geographic environmental features, periodic hour codes, date, and important event features.

[0009] Step 3: Divide the OD passenger flow matrix, spatial features, and time coding features into a training set and a validation set; input the passenger flow distribution matrix and related spatiotemporal data features into the CompLSTM model, organize the predicted passenger flow matrix, and obtain the OD passenger flow prediction matrix and the IO passenger flow prediction matrix.

[0010] As an improvement of the present invention, the raw passenger flow data of the AFC system is extracted in step 1 of the present invention, and the factors affecting the subway passenger flow prediction are mined, including spatial features and time coding features. The historical subway passenger flow is obtained by the subway AFC system. When passengers enter and exit the station through the gate, the AFC system will record the passenger's entry time, entry station, passenger type, exit time, exit station and travel fee; the raw data recorded by the system may contain erroneous or incomplete information, and the raw data needs to be deduplicated, outliers and missing values ​​processed; the sample travel time is calculated and outliers are eliminated: ① the travel time exceeds 120 minutes or is less than 5 minutes, ② the entry station and exit station are the same, ③ the entry and exit time are inconsistent with the recording day; the cleaned data is standardized and divided into granularity with every 30 minutes as a time node, and finally a historical subway starting and ending point passenger flow distribution matrix is ​​generated;

[0011] The external influencing factor data of the present invention includes spatial features and time coding features. The spatial features include geographical environment features, station distance matrix, station topology relationship and station function relationship features.

[0012] The geographical environment features described in the present invention are composed of POI data, land mixing entropy, and traffic accessibility.

[0013] (1) POI data

[0014] Facilities around stations have a significant impact on travel behavior. POI data for residential areas, commercial areas, shopping centers, universities, medical facilities, and scenic spots within 400m, 500m, 600m, and 700m buffer zones around different stations were selected.

[0015] (2) Land mixing entropy

[0016] Based on the above six types of POI data, the land mixing entropy of different buffer zones near each station is calculated. The calculation formula is as follows:

[0017]

[0018] Among them, H represents the land mixing entropy of each station, n represents the POI data type, and p i Indicates the proportion of the i-th type POI in the buffer zone.

[0019] (3) Transportation accessibility

[0020] Transportation accessibility represents the difficulty of reaching other locations or regions from a certain location or region. The calculation formula is as follows:

[0021]

[0022] in, represents the accessibility of station i, W represents the weight of POI, T ij represents the travel cost between station i and POI data, θ is the decay coefficient, and j represents the POI category.

[0023] The above data are linearly spliced ​​to obtain geographical environment characteristics without the need for granularity division processing.

[0024] The site topology represents the connection relationship with sites as nodes. In the rail transit network, some sites are transfer stations, thus forming the connection relationship between different lines and forming a site connection matrix, as shown below:

[0025]

[0026] Among them, j and v represent two adjacent or non-adjacent nodes respectively.

[0027] The site distance matrix includes the site location distribution, site spacing, and network internal connectivity. The site distance matrix is ​​constructed by the inverse of the distance between sites, as shown below:

[0028]

[0029] Among them, Distance (j，v) is the actual spatial distance between two adjacent nodes.

[0030] Station functional relationships refer to determining the travel purpose of a trip consisting of origin and destination points based on historical station passenger flow and nearby POI data. K-shape cluster analysis was performed on the passenger flow at the origin and destination points, and eight clustering results were obtained using the elbow analysis method. The specific representation is as follows:

[0031]

[0032] Among them, C is the station function relationship matrix, f(x NN ) represents the clustering result of passenger flow at the origin and destination stations.

[0033] The nodes in the knowledge graph are created based on the geographical characteristics of the different stations mentioned above. A multi-graph adjacency matrix is ​​constructed based on the station distance matrix, station topology, and station functional relationship characteristics. Granularity is divided into 30-minute time nodes to obtain the model input. The multi-graph adjacency matrix is ​​represented by a four-tuple form:

[0034] G=(h,r,t,w hrt )

[0035] Wherein, h represents the starting node; r represents the relationship number, and in the present invention, r=0, 1, and 2 represent the geographical distance relationship, the site topological connection relationship, and the station function relationship respectively; t represents the end node; whrt Represents the edge weights of different association relationships.

[0036] External dynamic environmental feature data includes temperature and rainfall characteristics, holiday date characteristics, dynamic time coding characteristics, and peak coding. The holiday date characteristics and dynamic time coding characteristics in the above data are linearly spliced ​​to obtain time influencing factor data, and the temperature and rainfall characteristics and peak coding are LSTM encoded; the time influencing factor data is standardized and divided into granularity with every 30 minutes as a time node.

[0037] The time coding features described in the present invention include date features, periodic hour codes, peak codes and weather features.

[0038] (1) Weather characteristics

[0039] The temperature and rainfall data corresponding to the passenger flow are collected and standardized, and the granularity is divided into 30-minute time nodes. The output is expressed as Y weather .

[0040] (2) Holiday date characteristics

[0041] Collect the holiday, weekday, and rest day features of the passenger flow corresponding to the day and perform one-hot encoding. The formula is as follows:

[0042] Y e =E embedding (E weekday , E weekend , E holiday )

[0043] Among them, E embedding represents the combination of dynamic time coding features, E weekday Indicates the code for the working day, E weekend The code for weekend, E holiday Indicates the holiday code.

[0044] (3) Periodic hour coding

[0045] Dynamic time features are encoded and integrated into the feature fusion layer. Periodicity is retained in the time domain, and trigonometric functions are used to time-code hourly series. Peak time features are extracted and integrated with date and important event features. Granularity is then divided into 30-minute time nodes. The formula is as follows:

[0046]

[0047] Where t is the input time series, sin(t) and cos(t) are the trigonometric function encoding results.

[0048] (4) Peak coding

[0049] The OD passenger flow between different stations has a periodic pattern due to residents' commuting behavior. Based on the OD historical passenger flow matrix, a peak coding matrix is ​​constructed with a granularity of 30 minutes as a time node to extract the passenger flow change trend of different stations. The formula is as follows:

[0050]

[0051] Among them, OD t Represents the OD passenger flow data at time step t; ∈ is a very small positive number.

[0052] As an improvement of the present invention, the specific method of integrating and fusing multi-dimensional feature models in step 2 of the present invention to construct a short-term passenger flow prediction module for the origin and destination includes: dividing the OD passenger flow matrix, spatial features and time coding features into a training set and a validation set. A multi-graph adjacency matrix is ​​constructed using geographical environment features, station distance matrix, station topology relationship and station functional relationship features as the input of the CompGCN module, and the OD passenger flow information code is output. The Attention-LSTM module is further used to process the OD passenger flow information code of the time series, and effectively identifies and assigns higher weights to key time steps through a dynamic attention mechanism. The LSTM module processes part of the time coding feature data including temperature data, rainfall data and peak coding; finally, the OD passenger flow information code processed by Attention-LSTM, the weather features and peak coding features processed by LSTM, the date features after one-hot coding and the periodic hour coding are linearly superimposed, and the final prediction result is output through a fully connected network and activation function module, and its dimension is consistent with the dimension of the passenger flow OD matrix in the input data. The specific mathematical expression formula is as follows:

[0053]

[0054] Among them, Y OD Represents the predicted OD passenger flow matrix; OD in and OD out The passenger flow OD matrix, station operation characteristics and station spatial relationship are converted into a multi-graph adjacency matrix quadruple input CompGCN; Y weather Contains temperature and rainfall; Y e Including holiday date features and dynamic time coding; and They represent the peak codes of the origin and destination matrices respectively.

[0055] As an improvement of the present invention, in step 3 of the present invention, the predicted OD passenger flow matrix is ​​sorted and separated by time intervals to obtain the starting point passenger flow matrix and the destination point passenger flow matrix, which are specifically expressed as follows:

[0056]

[0057] According to the above matrix, the inbound passenger flow of the departure station can be obtained from the row sum of the starting point passenger flow matrix, and the outbound passenger flow of the arrival station can be obtained from the column sum of the destination passenger flow matrix.

[0058]

[0059] in, and They represent the starting point passenger flow matrix and the ending point passenger flow matrix of the t-th time interval respectively; N represents the total number of stations; represents the passenger flow from the starting station i to j in the tth time interval; inflow and outflow represent the vectors of inflow and outflow passenger flows respectively; T represents the time interval of all predictions.

[0060] Beneficial effects:

[0061] 1. The present invention integrates multi-dimensional features, such as external geographical environment features, weather features, date features, and station operation features, to comprehensively explore the relationship between spatiotemporal data and rail transit network passenger flow, thereby achieving simultaneous prediction of subway inbound and outbound passenger flow and passenger flow at origin and destination points.

[0062] 2. This paper proposes a CompLSTM model for multi-task subway passenger flow forecasting. Inbound and outbound passenger flow forecasting helps subway station operators optimize scheduling and operational strategies, improves passenger travel experience, and reduces congestion. Origin-destination passenger flow forecasting can accurately identify differences in travel demand between stations, providing an important basis for optimizing route planning, improving service efficiency, and enhancing the passenger travel experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Flow chart of the method of the present invention.

[0064] Figure 2 This is a diagram of the internal structure of the model of the present invention. DETAILED DESCRIPTION

[0065] The present invention will be described in further detail below with reference to the accompanying drawings and specific implementations.

[0066] Example 1

[0067] like Figure 1 As shown, the present invention provides a subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features, which specifically includes the following:

[0068] Obtain the original passenger flow data collected by the AFC system of all subway stations, and remove duplicates, outliers and missing values ​​from the original passenger flow data.

[0069] Passenger travel time is calculated and outliers are eliminated: ① travel time exceeds 120 minutes or is less than 5 minutes, ② the entry and exit stations are the same, ③ the entry and exit times are inconsistent with the recorded day; the cleaned data is standardized and divided into granularity with every 30 minutes as a time node, and finally a historical subway origin and destination passenger flow distribution matrix is ​​generated.

[0070] Obtain the external spatial influencing factors of subway passenger flow distribution, including POI data, land mixing entropy data, traffic buffer data, and traffic accessibility data, and construct a geographical environment matrix. This matrix does not require granularity division and only requires normalization operations;

[0071] Time-coded feature data is obtained, including date features, periodic hourly codes, peak codes, and weather features. Date features consist of weekdays, weekends, and holidays and are one-hot encoded. Weather features consist of temperature and rainfall, divided into 30-minute time intervals. Periodic hourly codes are obtained by time-coding the hourly sequence using trigonometric functions. A peak code matrix is ​​constructed based on historical passenger flow changes at the origin and destination points, with 30-minute intervals as the granularity.

[0072] The adjacency relationship between each subway station is obtained, including the station distance matrix and the station topological relationship. According to the changes in the OD passenger flow matrix, k-shape clustering is used, and the station function relationship matrix is ​​obtained by combining the POI data near the station.

[0073] The OD passenger flow matrix, spatial features, and time-coded features are divided into training and validation sets. A multi-graph adjacency matrix is ​​constructed using geographic features, station distance matrix, station topology, and station functional relationship features. This matrix is ​​then fed into the CompGCN module as a four-tuple input to produce the output OD passenger flow information encoding. The Attention-LSTM module is further used to process the OD passenger flow information encoding in the time series. A dynamic attention mechanism effectively identifies and assigns higher weights to key time steps. A three-layer LSTM module and a two-layer fully connected neural network are used to process peak-coded features and weather features, respectively. Finally, the OD passenger flow information encoding processed by the Attention-LSTM, the weather features and peak-coded features processed by the LSTM and fully connected neural network, the one-hot-encoded date features, and the periodic hour encoding are linearly superimposed. The final prediction result is output through the fully connected network and activation function module. Its dimensions match the OD passenger flow matrix in the input data.

[0074] The predicted OD passenger flow matrix is ​​sorted and separated by time intervals to obtain the starting point passenger flow matrix and the destination point passenger flow matrix.

[0075] Model indicator verification uses MAE, MAPE, and RMSE indicators to compare the final predicted data with the actual passenger flow data. The formulas for MAE, MAPE, and RMSE are:

[0076]

[0077]

[0078] Where n represents the length of sample data, y i is the actual value of the sample, is the sample prediction value.

[0079] The present invention integrates multi-dimensional features to realize multi-task prediction of subway in-and-out passenger flow and passenger flow at origin and destination points. Through spatiotemporal data analysis, external factors such as temperature, rainfall, POI, land mixing entropy, traffic accessibility, and inter-station distance are introduced to quantitatively analyze the time, space, and station characteristics of the transportation railway network; an end-to-end multi-dimensional feature model CompLSTM is constructed, which will help subway operators optimize scheduling and operation strategies, improve passengers' travel experience, reduce congestion, and thus realize efficient and intelligent management of urban rail transit.

[0080] The present invention includes pre-processing the original passenger flow data collected by the AFC system to obtain the passenger flow distribution matrix of all stations in the rail transit network; collecting multi-dimensional features, including: external dynamic environment features, external geographical environment features, station spatial relationships, periodic hour codes, date features and peak code features; constructing an end-to-end multi-dimensional feature integration model CompLSTM (CompGCN-Attention-LSTM), combining multi-graph adjacency matrices, integrating the passenger flow of origin and destination points, station spatial connection relationships, station distance relationships and station functional relationships; using Attention-LSTM to mine the spatiotemporal characteristics of the passenger flow matrix that integrates multi-dimensional features; based on the passenger flow increase and decrease rate of each OD unit between the current moment and the previous moment, constructing a peak state dynamic matrix, and combining the LSTM module to extract time series features, thereby enhancing the model's fitting ability for peak periods; using trigonometric functions to time-code the hourly series, integrating date and important event features, and enhancing the model's dependence on time; and organizing the predicted output passenger flow matrix into IO passenger flow and OD passenger flow matrices of different nodes. The present invention integrates multi-dimensional features to construct an end-to-end integrated model CompLSTM, which comprehensively explores the spatiotemporal data characteristics: geographical environment characteristics, dynamic environment characteristics, topological connection relationships between sites within the network, site distance characteristics and station functional relationships, periodic hourly coding and peak characteristics of origin and destination points. It can accurately predict the short-term passenger flow at the origin and destination points of multiple sites in the subway network and the short-term passenger flow in and out of the station, which is not only helpful for line expansion planning and network optimization design, but also can optimize subway operation strategies and realize intelligent management of urban rail transit networks.

[0081] Example 2

[0082] The present invention provides a subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features, the method comprising the following steps:

[0083] Step 1: Preprocess the raw passenger flow data collected from the AFC system, including outlier processing and time granularity selection, to obtain historical subway passenger flow OD data and construct a passenger flow OD matrix. Data on external factors influencing subway passenger flow are obtained, primarily including spatial and temporal coding features. Spatial features include geographic environment characteristics, station distance matrix, station topology, and station function relationships. Temporal coding features include date characteristics, periodic hour codes, peak codes, and weather characteristics.

[0084] Step 2: Build an end-to-end multidimensional feature integration model, CompLSTM (CompGCN-Attention-LSTM). The CompGCN module processes time-coded subway passenger flow O / D data, station operation characteristics, station topology, and station distance matrix. The Attention-LSTM module further explores the relationship between passenger flow and other spatiotemporal data. The LSTM is used to process peak-coding features and some time series external dynamic environmental features, such as weather characteristics. The feature fusion layer processes these results along with external geographic environmental features, periodic hour codes, date, and important event features.

[0085] Step 3: Divide the OD passenger flow matrix, spatial features, and time coding features into a training set and a validation set; input the passenger flow distribution matrix and related spatiotemporal data features into the CompLSTM model, organize the predicted passenger flow matrix, and obtain the OD passenger flow prediction matrix and the IO passenger flow prediction matrix.

[0086] In step 1 of the present invention, the original passenger flow data of the AFC system is extracted, and the factors affecting the subway passenger flow prediction are explored, including spatial features and time coding features. The historical subway passenger flow is obtained by the subway AFC system. When passengers enter and exit the station through the gate, the AFC system will record the passenger's entry time, entry station, passenger type, exit time, exit station and travel fee; the original data recorded by the system may contain erroneous or incomplete information, and the original data needs to be deduplicated, outliers and missing values ​​are processed; the sample travel time is calculated and outliers are eliminated: ① the travel time exceeds 120 minutes or is less than 5 minutes, ② the entry station and exit station are the same, ③ the entry and exit time are inconsistent with the recording day; the cleaned data is standardized and divided into granularity with every 30 minutes as a time node, and finally a historical subway starting and ending point passenger flow distribution matrix is ​​generated;

[0087] The external influencing factor data of the present invention includes spatial features and time coding features. The spatial features include geographical environment features, station distance matrix, station topology relationship and station function relationship features.

[0088] The geographical environment features described in the present invention are composed of POI data, land mixing entropy, and traffic accessibility.

[0089] (3) POI data

[0090] Facilities around stations have a significant impact on travel behavior. POI data for residential areas, commercial areas, shopping centers, universities, medical facilities, and scenic spots within 400m, 500m, 600m, and 700m buffer zones around different stations were selected.

[0091] (4) Land mixing entropy

[0092] Based on the above six types of POI data, the land mixing entropy of different buffer zones near each station is calculated. The calculation formula is as follows:

[0093]

[0094] Among them, H represents the land mixing entropy of each station, n represents the POI data type, and p i Indicates the proportion of the i-th type POI in the buffer zone.

[0095] (3) Transportation accessibility

[0096] Transportation accessibility represents the difficulty of reaching other locations or regions from a certain location or region. The calculation formula is as follows:

[0097]

[0098] in, represents the accessibility of station i, W represents the weight of POI, T ij represents the travel cost between station i and POI data, θ is the decay coefficient, and j represents the POI category.

[0099] The above data are linearly spliced ​​to obtain geographical environment characteristics without the need for granularity division processing.

[0100] The site topology represents the connection relationship with sites as nodes. In the rail transit network, some sites are transfer stations, thus forming the connection relationship between different lines and forming a site connection matrix, as shown below:

[0101]

[0102] Among them, j and v represent two adjacent or non-adjacent nodes respectively.

[0103] The site distance matrix includes the site location distribution, site spacing, and network internal connectivity. The site distance matrix is ​​constructed by the inverse of the distance between sites, as shown below:

[0104]

[0105] Among them, Distance (j，v) is the actual spatial distance between two adjacent nodes.

[0106] Station functional relationships refer to determining the travel purpose of a trip consisting of origin and destination points based on historical station passenger flow and nearby POI data. K-shape cluster analysis was performed on the passenger flow at the origin and destination points, and eight clustering results were obtained using the elbow analysis method. The specific representation is as follows:

[0107]

[0108] Among them, C is the station function relationship matrix, f(x NN ) represents the clustering result of passenger flow at the origin and destination stations.

[0109] The nodes in the knowledge graph are created based on the geographical characteristics of the different stations mentioned above. A multi-graph adjacency matrix is ​​constructed based on the station distance matrix, station topology, and station functional relationship characteristics. Granularity is divided into 30-minute time nodes to obtain the model input. The multi-graph adjacency matrix is ​​represented by a four-tuple form:

[0110] G=(h,e,t,w hrt )

[0111] Wherein, h represents the starting node; r represents the relationship number, and in the present invention, r=0, 1, and 2 represent the geographical distance relationship, the site topological connection relationship, and the station function relationship respectively; t represents the end node; w hrt Represents the edge weights of different association relationships.

[0112] External dynamic environmental feature data includes temperature and rainfall characteristics, holiday date characteristics, dynamic time coding characteristics, and peak coding. The holiday date characteristics and dynamic time coding characteristics in the above data are linearly spliced ​​to obtain time influencing factor data, and the temperature and rainfall characteristics and peak coding are LSTM encoded; the time influencing factor data is standardized and divided into granularity with every 30 minutes as a time node.

[0113] The time coding features described in the present invention include date features, periodic hour codes, peak codes and weather features.

[0114] (1) Weather characteristics

[0115] The temperature and rainfall data corresponding to the passenger flow are collected and standardized, and the granularity is divided into 30-minute time nodes. The output is expressed as Y weather .

[0116] (2) Holiday date characteristics

[0117] Collect the holiday, weekday, and rest day features of the passenger flow corresponding to the day and perform one-hot encoding. The formula is as follows:

[0118] Y e =E embedding (E weekday , E weekend , E holiday )

[0119] Among them, E embedding represents the combination of dynamic time coding features, E weekday Indicates the code for the working day, E weekend The code for weekend, E holiday Indicates the holiday code.

[0120] (3) Periodic hour coding

[0121] Dynamic time features are encoded and integrated into the feature fusion layer. Periodicity is retained in the time domain, and trigonometric functions are used to time-code hourly series. Peak time features are extracted and integrated with date and important event features. Granularity is then divided into 30-minute time nodes. The formula is as follows:

[0122]

[0123] Where t is the input time series, sin(t) and cos(t) are the trigonometric function encoding results.

[0124] (4) Peak coding

[0125] The OD passenger flow between different stations has a periodic pattern due to residents' commuting behavior. Based on the OD historical passenger flow matrix, a peak coding matrix is ​​constructed with a granularity of 30 minutes as a time node to extract the passenger flow change trend of different stations. The formula is as follows:

[0126]

[0127] Among them, OD t Represents the OD passenger flow data at time step t; ∈ is a very small positive number.

[0128] As an improvement of the present invention, the specific method of integrating and fusing multi-dimensional feature models in step 2 of the present invention to construct a short-term passenger flow prediction module for the origin and destination includes: dividing the OD passenger flow matrix, spatial features and time coding features into a training set and a validation set. A multi-graph adjacency matrix is ​​constructed using geographical environment features, station distance matrix, station topology relationship and station functional relationship features as the input of the CompGCN module, and the OD passenger flow information code is output. The Attention-LSTM module is further used to process the OD passenger flow information code of the time series, and effectively identifies and assigns higher weights to key time steps through a dynamic attention mechanism. The LSTM module processes part of the time coding feature data including temperature data, rainfall data and peak coding; finally, the OD passenger flow information code processed by Attention-LSTM, the weather features and peak coding features processed by LSTM, the date features after one-hot coding and the periodic hour coding are linearly superimposed, and the final prediction result is output through a fully connected network and activation function module, and its dimension is consistent with the dimension of the passenger flow OD matrix in the input data. The specific mathematical expression formula is as follows:

[0129]

[0130] Among them, Y OD Represents the predicted OD passenger flow matrix; OD in and OD out The passenger flow OD matrix, station operation characteristics and station spatial relationship are converted into a multi-graph adjacency matrix quadruple input CompGCN; Y weather Contains temperature and rainfall; Y e Including holiday date features and dynamic time coding; and They represent the peak codes of the origin and destination matrices respectively.

[0131] As an improvement of the present invention, in step 3 of the present invention, the predicted OD passenger flow matrix is ​​sorted and separated by time intervals to obtain the starting point passenger flow matrix and the destination point passenger flow matrix, which are specifically expressed as follows:

[0132]

[0133] According to the above matrix, the inbound passenger flow of the departure station can be obtained from the row sum of the starting point passenger flow matrix, and the outbound passenger flow of the arrival station can be obtained from the column sum of the destination passenger flow matrix.

[0134]

[0135] in, and They represent the starting point passenger flow matrix and the ending point passenger flow matrix of the t-th time interval respectively; N represents the total number of stations; represents the passenger flow from the starting station i to j in the tth time interval; inflow and outflow represent the vectors of inflow and outflow passenger flows respectively; T represents the time interval of all predictions.

[0136] The above description is merely an example of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features, characterized by: The method comprises the following steps: Step 1: Preprocess the collected raw passenger flow data from the AFC system, including outlier processing and time granularity selection, to obtain the OD data of historical subway passenger flow and construct the passenger flow OD matrix; Obtain data on external factors that affect subway passenger flow, including spatial characteristics and time-coded characteristics. Spatial characteristics include geographical environment characteristics, station distance matrix, station topology relationship, and station function relationship. Time-coded characteristics include date characteristics, periodic hour code, peak code, and weather characteristics. Step 2: Build an end-to-end multidimensional feature integration model, CompLSTM (CompGCN-Attention-LSTM). The CompGCN module processes time-coded subway passenger flow OD data, station operation characteristics, station topology relationships, and station distance matrices. The Attention-LSTM module further explores the relationship between passenger flow and other spatiotemporal data. The LSTM is used to process peak coding features and some time series external dynamic environmental features, such as weather characteristics. The feature fusion layer processes these results along with external geographic environmental features, periodic hour codes, dates, and important event features. Step 3: Input the passenger flow distribution matrix and related spatiotemporal data features into the CompLSTM model, organize the predicted passenger flow matrix, and obtain the OD passenger flow prediction matrix and IO passenger flow prediction matrix.

2. The subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features according to claim 1 is characterized in that: The step 1 comprises: Based on the original data from the subway's AFC system, we extracted the passengers' entry time, entry station, exit time, and exit station, calculated the sample travel time, and eliminated outliers: ① the travel time exceeded 120 minutes or was less than 5 minutes, ② the entry and exit stations were the same, and ③ the entry and exit times were inconsistent with the recorded day. The cleaned data was then used to generate different OD passenger flow matrices based on the entry and exit times with a time granularity of 30 minutes. The external influencing factor data includes spatial characteristics and time coding characteristics, and the spatial characteristics include geographical environment characteristics, station distance matrix, station topological relationship and station functional relationship characteristics; Geographical environment features consist of POI data, land mixing entropy, and transportation accessibility; (1) POI data; The facilities around the station have a significant impact on travel behavior. POI data for residential areas, commercial areas, shopping centers, colleges and universities, medical facilities, and scenic spots within 400m, 500m, 600m, and 700m buffer zones around different stations were selected. (2) Land mixing entropy; Based on the above six types of POI data, the land mixing entropy of different buffer zones near each station is calculated. The calculation formula is as follows: Among them, H represents the land mixing entropy of each station, n represents the POI data type (6 in this paper), p i Indicates the proportion of the i-th type POI in the buffer zone; (3) transportation accessibility; Transportation accessibility represents the difficulty of reaching other locations or regions from a certain location or region. The calculation formula is as follows: in, represents the accessibility of station i, W represents the weight of POI, T ij represents the travel cost between station i and POI data, θ is the decay coefficient, and j represents the POI category; The site topology represents the connection relationship with sites as nodes. In the rail transit network, some sites are transfer stations, thus forming the connection relationship between different lines and forming a site connection matrix. The site distance matrix includes the site location distribution, site spacing, and internal network connectivity. The site distance matrix is ​​constructed by the inverse of the distance between sites, as shown below: Among them, j and v represent two adjacent or non-adjacent nodes respectively, and Distance (j，v) is the actual spatial distance between two adjacent nodes; The station functional relationship refers to determining the trip purpose consisting of the origin and destination points based on the station's historical passenger flow and nearby POI data. The passenger flow of the origin and destination points is subjected to k-shape cluster analysis, and eight clustering results are obtained using the elbow analysis method. The specific representation is as follows: Among them, C is the station function relationship matrix, f(x NN ) represents the clustering result of passenger flow at the origin and destination stations; The time coding features include date features, periodic hour codes, peak codes and weather features; (1) Weather characteristics; Collect the temperature and rainfall data corresponding to the passenger flow and perform normalization processing. The output is represented as Y weather ; (2) Holiday date characteristics; Collect the holiday, weekday, and rest day features of the passenger flow at the corresponding time and perform one-hot encoding. The formula is as follows: Y e =E embedding (E weekday ,E weekend ,E hoiday ) Among them, E embedding represents the combination of dynamic time coding features, E weekday Indicates the code for the working day, E weekend The code for weekend, E holiday Indicates holiday code; (3) Periodic hour coding; Dynamic time features are encoded and integrated into the feature fusion layer. Periodicity is retained in the time domain. Trigonometric functions are used to time-code hourly series. Peak time features are extracted and fused with date and important event features. The formula is as follows: Where t is the input time series, sin(t) and cos(t) are the trigonometric function encoding results; (4) Peak coding; The OD passenger flow between different stations has a periodic pattern due to the commuting behavior of residents. Based on the OD historical passenger flow matrix, a peak coding matrix is ​​constructed to extract the passenger flow change trend of different stations. The formula is as follows: Among them, OD t Represents the OD passenger flow data at time step t; ∈ is a very small positive number.

3. The subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features according to claim 1 is characterized in that: The specific method for integrating and fusing the multidimensional feature model in step 2 to construct an end-to-end multidimensional feature integration model CompLSTM (CompGCN-Attention-LSTM) includes: dividing the OD passenger flow matrix, spatial features, and time coding features described in step 1 into a training set and a validation set, using geographical environment features, station distance matrix, station topology relationship, and station functional relationship features to construct a Knowledge Graph as the input of the CompGCN module, and outputting the OD passenger flow information code. The Attention LSTM module is further used to process the OD passenger flow information code of the time series, effectively identifying and assigning higher weights to key time steps through a dynamic attention mechanism, and the LSTM module processes part of the time coding feature data including temperature data, rainfall data, and peak coding; finally, the OD passenger flow information code processed by Attention-LSTM, the weather features and peak coding features processed by LSTM, the date features after one-hot encoding, and the periodic hour coding are linearly superimposed, and the final prediction result is output through a fully connected network and an activation function module. Its dimension is consistent with the dimension of the passenger flow OD matrix in the input data. The specific mathematical expression formula is as follows: Among them, Y OD Represents the predicted OD passenger flow matrix; OD in and OD out The passenger flow OD matrix, station operation characteristics and station spatial relationship are converted into Knowledge Graph quadruple input CompGCN; Y weather Contains temperature and rainfall; Y e Including holiday date features and dynamic time coding; and They represent the peak codes of the origin and destination matrices respectively.

4. The subway multi-task passenger flow prediction method integrating spatiotemporal multidimensional features according to claim 1 is characterized in that: The specific method of step 3 is: The predicted OD passenger flow matrix is ​​sorted and separated by time intervals to obtain the starting point passenger flow matrix and the destination point passenger flow matrix, which are specifically expressed as follows: According to the above matrix, the inbound passenger flow of the departure station is obtained by the row sum of the starting point passenger flow matrix, and the outbound passenger flow of the arrival station is obtained by the column sum of the destination passenger flow matrix: in, and They represent the starting point passenger flow matrix and the ending point passenger flow matrix of the t-th time interval respectively; N represents the total number of stations; represents the passenger flow from the starting station i to j in the tth time interval; inflow and outflow represent the vectors of inflow and outflow passenger flows respectively; T represents the time interval of all predictions.

Images