Subway passenger flow volume prediction method, system and device considering subway-bus transfer and medium

Abstract

The invention discloses a subway passenger flow prediction method, system and device considering subway-bus transfer, and a medium, and the method comprises the following steps: obtaining the historical passenger flow of a subway station in a certain region within a period of time and the historical passenger flow of a corresponding bus station around each subway station, and forming a data set; dividing the data set into a training set and a test set according to a time sequence; constructing a dynamic space-time decomposition and fusion network, wherein the dynamic space-time decomposition and fusion network comprises a time sequence decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module and a dual-path space-time adaptive fusion module; and training the dynamic space-time decomposition and fusion network by using the training set and the test set, and predicting the subway passenger flow volume by using the optimized dynamic space-time decomposition and fusion network by using a mean absolute error, a root-mean-square error and a weighted mean absolute error percentage as performance evaluation indexes. The prediction precision is improved, and the practicability and generalization ability of the model are enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and particularly relates to a subway passenger flow prediction method, system, device and medium considering subway-bus transfer. BACKGROUND

[0002] In the urban rail transit system, short-time passenger flow prediction of subway stations is of great significance for ensuring real-time scheduling, safe operation and resource allocation of subways. Accurate passenger flow prediction can not only alleviate peak congestion and improve passenger travel experience, but also provide decision support for operators, thereby realizing fine management.

[0003] Traditional passenger flow prediction methods are mostly based on time series modeling or machine learning algorithms. These methods usually focus on historical passenger flow data of a single station and predict the periodicity and trend over time through modeling. However, in a real urban transportation network, subway passenger flow is not only affected by its own historical passenger flow, but also significantly disturbed by bus transfer behavior around the station. For example, when there is a sudden change in bus station passenger flow, the passenger flow of nearby subway stations will also fluctuate synchronously. This linkage between multiple modes is difficult for traditional models to capture. In addition, although existing deep learning methods introduce spatiotemporal graph convolution networks that consider time and space characteristics, they mostly model spatial dependence based only on the topological structure of the transportation network, ignoring the variability of spatial weights over time, which leads to insufficient prediction ability of the model in complex transfer scenarios. SUMMARY

[0004] The purpose of the present application is to provide a subway passenger flow prediction method considering subway-bus transfer, which improves prediction accuracy and enhances the practicality and generalization ability of the model.

[0005] The second purpose of the present application is to provide a subway passenger flow prediction system considering subway-bus transfer.

[0006] The third purpose of the present application is to provide an electronic device.

[0007] The fourth purpose of the present application is to provide a computer-readable storage medium.

[0008] Technical solution: To achieve the above purpose, the present application discloses a subway passenger flow prediction method considering subway-bus transfer, comprising the following steps:

[0009] (1) Obtain historical passenger flow D of subway stations in a certain area within a period of time t and historical passenger flow G of bus stations corresponding to each subway station around the subway station t form a data set, and divide the data set into a training set and a test set according to time sequence;

[0010] (2) Construct a dynamic spatio-temporal decomposition and fusion network, the dynamic spatio-temporal decomposition and fusion network comprising a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module and a double-path spatio-temporal adaptive fusion module; the historical passenger flow D t of the subway station and the historical passenger flow G t of the bus station are input into the time series decomposition module to perform structural decomposition and extract multi-scale time sequence features of the subway and the bus, and output fluctuation items and trend items of the subway and the bus; the fluctuation items and the trend items of the subway and the bus are input into the corresponding dynamic graph convolution module to dynamically extract spatial correlation features between stations, and output subway features D H and bus features G H ; and the subway features D H and the bus features G H are input into the gated cross-modal attention fusion module to extract subway and bus transfer behavior features, and output cross-modal fusion features H F ; and the cross-modal fusion features H F are input into the double-path spatio-temporal adaptive fusion module to integrate features in the time dimension and the space dimension, and finally output a prediction result.

[0011] (3) The dynamic spatio-temporal decomposition and fusion network is trained by using a training set and a test set, the mean absolute error, the root mean square error and the weighted average absolute error percentage are used as performance evaluation indexes, and the optimized dynamic spatio-temporal decomposition and fusion network is used for subway passenger flow prediction.

[0012] Optionally, the following steps are specifically performed in the time series decomposition module: the historical passenger flow D t of the subway station and the historical passenger flow G t of the bus station are input into the corresponding time series decomposition module, respectively, to perform a moving average operation and decompose the historical passenger flow into a long-term trend item and a short-term fluctuation item, and then pass through a multi-layer perception machine (MLP) to obtain a subway trend item a subway fluctuation item a bus trend item and a bus fluctuation item

[0013] Optionally, the following steps are specifically performed in the dynamic graph convolution module: a subway topology structure diagram of a certain area is obtained and converted into a static adjacency matrix the subway trend item MetroTrend is input into a 2-layer one-dimensional convolution layer, a 2-layer linear layer and a Sigmoid activation function in sequence to obtain a first subway dynamic feature F D14MetroRemainder is input into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function in sequence to obtain a second subway dynamic feature F D24 D14 D24 static D1 D2 D1 D1 D2 D2 D1 D2 H ;

[0014] BusTrend is input into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function in sequence to obtain a first bus dynamic feature F G14 BusRemainder is input into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function in sequence to obtain a second bus dynamic feature F G24 G14 G24 static G1 G2 G1 G1 G2 G2 G1 G2 BusTrend is input into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function in sequence to obtain a first bus dynamic feature F​​​​​​​​​​​​​​​​​​​​​​​H .

[0015] Optionally, the gated cross-modal attention fusion module specifically performs the following steps: [The text abruptly shifts to a different topic] ...metro feature D... H and public transport features G H After dimensional concatenation, the input is processed through a multi-head attention mechanism, and the output is a weighted feature M(D). H G H Subway Feature D H and public transport features G H The summation is then input into a gating network, processed by a nonlinear gating function and a sigmoid activation function, and outputs a fused feature F; the weighted feature M(D) is further processed. H G H After performing element-wise multiplication of the fusion feature F with the subway feature D, H The summation outputs the cross-modal fusion feature H. F .

[0016] Optionally, the dual-path spatiotemporal adaptive fusion module specifically performs the following steps: introducing initial temporal attention weights. Spatial attention weights Cross-modal fusion feature H F The initial temporal attention weight Ta0 and spatial attention weight Sa0 are added to obtain the mixed input feature X; temporal attention is calculated on the mixed input feature X to obtain the temporal attention feature Ta(X), and spatial attention is calculated on the mixed input feature X to obtain the spatial attention feature Sa(X); temporal attention weights and spatial attention weights are extracted from the mixed input feature X to generate temporal and spatial weights G. t G s ∈(0,1) B×T×N The temporal attention feature Ta(X) and the spatial attention feature Sa(X) are multiplied by the temporal and spatial weights G, respectively. t G s The spatiotemporal fusion feature F is then added to the mixed input feature X to output the spatiotemporal fusion feature F. X Spatiotemporal fusion feature F X After nonlinear transformation via a multilayer perceptron (MLP) and residual connections, the final output feature H(X) is used as the prediction result for subway traffic flow.

[0017] Based on the same inventive concept, this invention discloses a subway passenger flow prediction system considering subway-bus transfers, comprising:

[0018] The historical data acquisition module is used to obtain the historical passenger flow D of a subway station in a certain area over a period of time. t Historical passenger flow G of the bus stops surrounding each subway station tForm a data set, and divide the data set into a training set and a test set according to a time sequence;

[0019] The network construction module is configured to construct a dynamic spatio-temporal decomposition and fusion network, which includes a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a double-path spatio-temporal adaptive fusion module. The time series decomposition module is configured to decompose historical passenger flow D t of the subway station and historical passenger flow G t of the bus station, and extract multi-scale time sequence features of the subway and the bus, and output fluctuation items and trend items of the subway and the bus. H The dynamic graph convolution module is configured to perform dynamic extraction on spatial correlation features between stations by inputting the fluctuation items and the trend items of the subway and the bus, respectively, and output subway features D H and bus features G H The gated cross-modal attention fusion module is configured to extract subway and bus transfer behavior features by inputting the subway features D H and the bus features G F , and output cross-modal fusion features H F The double-path spatio-temporal adaptive fusion module is configured to integrate features in the time dimension and the space dimension, and finally output a prediction result.

[0020] The network optimization module is configured to train the dynamic spatio-temporal decomposition and fusion network by using the training set and the test set, use mean absolute error, root mean square error, and weighted average absolute error percentage as performance evaluation indexes, and use the optimized dynamic spatio-temporal decomposition and fusion network to perform subway passenger flow prediction.

[0021] Optionally, in the gated cross-modal attention fusion module, the subway features D H and the bus features G H are dimensionally spliced and then input into a multi-head attention mechanism for processing to output weighted features M(D H ,G H ). The subway features D H and the bus features G H are added and then input into a gating network, processed by a nonlinear gating function and a Sigmoid activation function, and output fusion features F. The weighted features M(D H ,G H ) and the fusion features F are multiplied element by element, and then added to the subway features D H to output the cross-modal fusion features H F .

[0022] Optionally, in the double-path spatio-temporal adaptive fusion module, initial time attention weights and spatial attention weights Cross-modal fusion feature H F , the initial time attention weight Ta0 and the spatial attention weight Sa0 are added to obtain a mixed input feature X; the mixed input feature X is subjected to time attention calculation to obtain a time attention feature Ta(X), and the mixed input feature X is subjected to spatial attention calculation to obtain a spatial attention feature Sa(X); the mixed input feature X is subjected to time attention weight and spatial attention weight extraction to generate a time and spatial weight G t ,G s ∈(0, 1) B×T×N , the time attention feature Ta(X) and the spatial attention feature Sa(X) are multiplied by the time and spatial weight G t ,G s and are added to the mixed input feature X to output a space-time fusion feature F X ; the space-time fusion feature F X is subjected to nonlinear transformation via a multilayer perceptron MLP and a residual connection to finally output a feature H(X), which is taken as a prediction result of subway flow.

[0023] Based on the same inventive concept, the electronic device provided by the application comprises a processor and a storage medium;

[0024] The storage medium is used for storing instructions;

[0025] The processor is used for operating according to the instructions to perform the steps of the method as described above.

[0026] Based on the same inventive concept, the computer readable storage medium provided by the application has a computer program stored thereon, and the program is characterized in that when the program is executed by a processor, the steps of the method as described above are implemented.

[0027] Advantages: compared with the prior art, the application has the following remarkable advantages:

[0028] (1) The application aims at the problem that the existing method ignores multimodal transfer behavior, only considers static space-time characteristics, and has limited prediction accuracy, combines time series decomposition to capture the trend and volatility of subway and bus passenger flow, uses a dynamic graph convolution network to extract the spatial dynamic correlation between subway stations, and uses a cross-modal attention fusion network to capture the dynamic influence of bus and subway transfer behavior on subway passenger flow, and finally performs deep feature extraction in space and time through a double-path space-time adaptive mechanism, thereby realizing fine modeling of bus and subway transfer behavior, improving short-term passenger flow prediction accuracy, enhancing the adaptability of the model to actual traffic network changes, and having good engineering deployability and promotion value;

[0029] (2) The application enhances the modeling capability of passenger flow change law under different time scales through a time series decomposition module, and performs structural decomposition on the input subway and bus passenger flow time series; at the same time, a moving average filter is used to split the time series into a long-term trend item and a short-term fluctuation item, so as to extract stable patterns and local disturbances respectively, and to strengthen the retention of the structure of the time series by the model;

[0030] (3) The application uses a dynamic graph convolution module to learn the time-varying spatial dependence relationship between subway stations. Unlike the traditional graph convolution structure based on a static adjacency matrix, the application combines the subway topological structure and the real-time passenger flow mode to construct a dynamic adjacency matrix that can change over time; the dynamic graph convolution module simultaneously introduces a causal convolution mechanism to simulate the influence of historical passenger flow changes on the current prediction, especially the conduction effect in the presence of cross-modal delay (such as bus-to-subway transfer);

[0031] (4) The application uses a gated cross-modal attention fusion module to model the transfer relationship between the subway and the bus system, and designs a cross-modal fusion structure based on multi-head attention, which dynamically adjusts the information interaction intensity at different times and stations by introducing a gating mechanism; this mechanism ensures that bus information is fully integrated when transfer behavior is significant, and automatically suppresses redundant information when the influence is weak, thereby realizing fine-grained subway-bus fusion expression.

[0032] (5) The application further improves the modeling capability of heterogeneous spatial and temporal dependence through a dual-path spatio-temporal adaptive fusion module. The dual-path spatio-temporal adaptive fusion module uses a dual-branch structure to extract spatial and temporal features in subway passenger flow data, and fuses them through a gating mechanism; the spatial branch captures the structural dependence between different stations, and the temporal branch extracts periodic and trend changes, and finally the features are reorganized and optimized through a fully connected network and a residual connection to output the final passenger flow prediction result. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 The figure is a framework structure diagram of the dynamic spatio-temporal decomposition and fusion network in the application;

[0034] Figure 2 The figure is a framework structure diagram of the dynamic graph convolution module in the application;

[0035] Figure 3 The figure is a framework structure diagram of the dual-path spatio-temporal adaptive fusion module in the application;

[0036] Figure 4 Example 1 of the method proposed in the application for predicting passenger flow at different subway stations in Beijing;

[0037] Figure 5 Example 2 of the method proposed in the application for predicting passenger flow at different subway stations in Beijing. DETAILED DESCRIPTION

[0038] The technical solutions of the present application are further described below with reference to the accompanying drawings.

[0039] Subway Passenger Flow refers to the number of passengers entering and leaving each subway station within a specific time interval, which can be obtained through passenger card records (such as IC card or mobile payment records). This indicator is a core variable in subway operation scheduling and passenger flow prediction. Subway Graph is an undirected graph structure, where the nodes of the graph represent subway stations, including transfer stations (connecting multiple lines) and non-transfer stations; the edges of the graph connect adjacent stations, reflecting the track connection relationship between stations; this graph is encoded by a symmetric static adjacency matrix, not only describing the connection between stations on the same line, but also indirectly expressing the transfer relationship between lines through the sharing of transfer station node identities. Subway-Bus Transfer refers to the travel mode of passengers transferring between subway and bus systems during their commute; to characterize the impact of transfer behavior on subway passenger flow, the present application introduces the boarding and alighting passenger flow at bus stops within a 500-meter radius around subway stations as spatial-temporal features into the model to assist in modeling the interaction between subway and bus systems.

[0040] The present application aims to achieve accurate prediction of short-term passenger flow (entry / exit number) at subway stations by modeling the spatio-temporal dependence in historical passenger flow data and the cross-modal interaction between subway and bus systems. Specifically, given a historical subway passenger flow sequence covering multiple past time steps, the prediction target is the passenger flow sequence for future time steps. The present application models the input through a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a dual-path spatio-temporal adaptive fusion module, ultimately realizing the prediction process. By constructing a dynamic spatio-temporal decomposition and fusion network, the present application fuses historical subway and bus passenger flow data to achieve high-precision prediction of future passenger flow at subway stations; the present application aims to provide efficient real-time passenger flow prediction capabilities for urban traffic scheduling, optimize the allocation of transport resources, and improve the intelligent level of urban public transportation systems.

[0041] As shown in Figure 1 , the present application discloses a subway passenger flow prediction method considering subway-bus transfer, comprising the following steps:

[0042] (1) Obtain the historical passenger flow D t of subway stations in a certain area within a period of time and the historical passenger flow G t of the corresponding bus stops around each subway station to form a data set, the historical passenger flow D tHistorical passenger flow G at bus stops t The datasets are all in the form of three-dimensional tensors [B,T,N], where B is the batch size, T is the length of the historical time window, and N is the number of sites; and the datasets are divided into training and testing sets according to the time order.

[0043] (2) Construct a dynamic spatiotemporal decomposition and fusion network, which includes a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a dual-path spatiotemporal adaptive fusion module; the time series decomposition module is used to analyze the historical passenger flow D of subway stations. t Historical passenger flow G at bus stops t Structural decomposition is performed to extract multi-scale temporal features of subway and bus systems, outputting fluctuation and trend terms for both. These fluctuation and trend terms are then input into corresponding dynamic graph convolutional modules to dynamically extract spatial correlation features between stations, outputting the subway feature D. H and public transport features G H ; and will subway feature D H and public transport features G H The input gated cross-modal attention fusion module extracts features of subway and bus transfer behavior and outputs cross-modal fusion feature H. F Then fuse the cross-modal features H F The input dual-path spatiotemporal adaptive fusion module integrates features in the temporal and spatial dimensions, and finally outputs the prediction result.

[0044] The time series decomposition module uses a moving average filter to structurally decompose historical passenger flow sequences, effectively extracting multi-scale temporal features from subway and bus passenger flow data. The core idea of ​​this module is to decompose the original passenger flow time series into long-term trend terms and short-term fluctuation terms. The long-term trend term reflects the slow change trend of passenger flow over a longer period of time, while the short-term fluctuation term captures instantaneous fluctuations and sudden changes. Through time series decomposition, the model's ability to perceive periodic and trend changes can be enhanced, while maintaining the sequential characteristics of the time series and avoiding information loss.

[0045] The time series decomposition module specifically performs the following steps:

[0046] Historical passenger flow data of subway stations D t Historical passenger flow G at bus stops t The data are input into the corresponding time series decomposition modules, where moving average operations are performed to decompose historical passenger flow into long-term trend terms and short-term fluctuation terms. These are then processed by a multilayer perceptron (MLP) to obtain the subway trend term. Subway fluctuation items Public transport trends With public transport fluctuations

[0047]

[0048] MetroRemainder t = D t -MetroTrend t ,

[0049] where MetroTrend t represents the subway trend item at the t time, MetroRemainder t represents the subway fluctuation item at the t time, and ω is the size of the sliding window, when ω is 3, the value of i is 0, 1, 2; represents the subway historical passenger flow sequence at the t time, D t-i represents the subway historical passenger flow sequence at the t-i time,

[0050]

[0051] BusRemainder t = G t -BusTrend t

[0052] where, represents the bus historical passenger flow sequence at the t time, G t-i represents the bus historical passenger flow sequence at the t-i time, BusTrend t represents the bus trend item at the t time, BusRemainder t represents the bus fluctuation item at the t time.

[0053] The application firstly enhances the modeling ability of passenger flow change law at different time scales through a time series decomposition module, and performs structural decomposition on the input subway and bus passenger flow time series; at the same time, a moving average filter is used to split the time series into a long-term trend item and a short-term fluctuation item, so as to extract stable patterns and local disturbances respectively, and to strengthen the retention of the structure of the time series by the model.

[0054] Dynamic graph convolution (Dynamic Graph Convolution) is a convolution operation based on graph structure, which integrates static topological structure and dynamic passenger flow mode, so as to learn the spatial dependence relationship between stations changing over time; since the subway passenger flow has time-varying characteristics, the traditional static graph structure cannot effectively model the dynamic conduction process, so it is necessary to introduce a dynamic graph convolution mechanism to enhance the adaptability to spatial relationship.

[0055] The dynamic graph convolution module is used to capture the spatial dependence and causal structure changing over time in the subway system, thereby solving the problem that the traditional static graph-based graph neural network cannot effectively model the time-varying topology relationship. At the same time, by introducing a low-rank compression mechanism, the module significantly reduces the complexity of graph convolution calculation while ensuring modeling capability. As shown in FIG. 8, the following steps are specifically performed in the dynamic graph convolution module: Figure 2

[0056] Obtain the subway topology structure diagram of a certain area and convert it into a static adjacency matrix The calculation formula of the adjacency relationship weight A between the i th station and the j th station in the static adjacency matrix is as follows: ij

[0057]

[0058] The first subway dynamic feature F is obtained by sequentially inputting the subway trend item MetroTrend into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function D14 The second subway dynamic feature F is obtained by sequentially inputting the subway fluctuation item MetroRemainder into 2 one-dimensional convolution layers, 2 linear layers and a Sigmoid activation function D24

[0059] F D11 = Conv1D(MetroTrend), F D12 = Conv1D(F D11 ), F D13 = Dense(Flatten(F D12 )),

[0060] F D14 = σ(Dense(Flatten(F D13 ))), F D21 = Conv1D(MetroRemainder),

[0061] F D22 = Conv1D(F D21 ), F D23 = Dense(Flatten(F D22 )), F D24 = σ(Dense(Flatten(F D23 ))),

[0062] Wherein, Conv1D is a one-dimensional convolution layer, Dense represents a fully connected layer, Flatten represents a flattening operation, and σ is a Sigmoid activation function;

[0063] The first subway dynamic feature F​​​D14 and the second subway dynamic feature F D24 respectively with the static adjacency matrix A static element-wise multiplication to obtain the first subway dynamic adjacency matrix A D1 and the second subway dynamic adjacency matrix A D2 ;

[0064] A D1 = A static ⊙F D14 ,

[0065] A D2 = A static ⊙F D24 ,

[0066] wherein, ⊙ is Hadamard product;

[0067] the first subway dynamic adjacency matrix A D1 and the subway trend item MetroTrend again through the graph convolution propagation mechanism, using dynamic adjacency matrix based feature propagation between nodes, to obtain subway trend feature H D1 ; the second subway dynamic adjacency matrix A D2 and the subway fluctuation item MetroRemainder again through the graph convolution propagation mechanism, using dynamic adjacency matrix based feature propagation between nodes, to obtain subway fluctuation feature H D2 , the subway trend feature H D1 and the subway fluctuation feature H D2 fusion to obtain subway feature D H ;

[0068] H D1 = φ(A D1 ·MetroTrend·W l1 ), H D2 = φ(A D2 ·MetroRemainder·W l2 ),

[0069] D H = H D1 +H D2 , wherein, W l1 , W l2 are learnable parameters, and φ is Elu activation function;

[0070] the bus trend item BusTrend is input into 2 one-dimensional convolution layers, 2 linear layers and Sigmoid activation function in turn to obtain the first bus dynamic feature F G14, the second bus dynamic feature F G24 ;

[0071] F G11 = Conv1D(BusTrend), F G12 = Conv1D(F G11 ), F G13 = Dense(Flatten(F G12 )),

[0072] F G14 = sigma(Dense(Flatten(F G13 ))), F G21 = Conv1D(BusRemainder),

[0073] F G22 = Conv1D(F G21 ), F G23 = Dense(Flatten(F G22 )), F G24 = sigma(Dense(Flatten(F G23 ))),

[0074] The first bus dynamic feature F G14 and the second bus dynamic feature F G24 respectively element-wise multiply the static adjacency matrix A static to obtain the first bus dynamic adjacency matrix A G1 and the second bus dynamic adjacency matrix A G2 ;

[0075] A G1 = A static ⊙ sigma(F G14 ), A G2 = A static ⊙ sigma(F G24 ),

[0076] The first bus dynamic adjacency matrix A G1 and the bus trend BusTrend are further propagated through the graph convolution propagation mechanism, and the feature propagation between nodes based on the dynamic adjacency matrix is used to obtain the bus trend feature H G1 , the second bus dynamic adjacency matrix A G2 and the bus fluctuation BusRemainder are further propagated through the graph convolution propagation mechanism, and the feature propagation between nodes based on the dynamic adjacency matrix is used to obtain the bus fluctuation feature H G2 , the bus trend feature H G1and bus fluctuation feature H G2 Fusion to obtain bus feature G H ;

[0077] H G1 = φ (A G1 ·BusTrend·W l3 ), H G2 = φ (A G2 ·BusRemainer·W l4 ), G H = H G1 + H G2 ,

[0078] Wherein, W l3 , W l4 are learnable parameters.

[0079] The subway station between the time-varying spatial dependence relationship is learned by the dynamic graph convolution module; unlike the traditional graph convolution structure based on static adjacency matrix, the subway topology structure and real-time passenger flow mode are combined to construct the dynamic adjacency matrix which can change with time; the dynamic graph convolution module simultaneously introduces the causal convolution mechanism, which is used to simulate the influence of historical passenger flow change on current prediction, especially the conduction effect under the condition of cross-modal delay (such as bus to subway).

[0080] The gating cross-modal attention fusion module can effectively integrate the transfer information between the subway and the bus system, which mainly extracts the interaction features through the multi-head attention mechanism: the subway features and bus features extracted according to the dynamic graph convolution network are applied to the multi-head attention mechanism to capture the dynamic dependence relationship between the two modalities. The following steps are specifically executed in the gating cross-modal attention fusion module:

[0081] The subway feature D H and the bus feature G H are dimensionally spliced and input into the multi-head attention mechanism for processing, and the weighted feature M(D H , G H ) is output;

[0082] Q i = Dense (Concat (D H , G H )), K i = Dense (Concat (D H , G H )),

[0083] V i = Dense (Concat (D H , G H )),

[0084] M(D H ,G H )=Concat(Head1,...,Head h )W O ,

[0085] where Head i i denotes the i-th attention weight, Attention(Q i ,K i ,V i ) denotes the multi-head attention mechanism, Q i ,K i ,V i denote the query, key, value matrices in the i-th attention head, d k denotes the dimension of the key vector in the attention, Concat denotes the dimension concatenation function in deep learning, and W O is the trainable weight parameter in the attention mechanism;

[0086] The subway feature D H and the bus feature G H are added and input into the gating network, and after being processed by the nonlinear gating function and the Sigmoid activation function, the fusion feature F is output;

[0087] M H =φ(W1(D H +G H )),F=σ(W2M H ),

[0088] where M H is the hidden layer feature in the gating network, F is the fusion feature output by the gating network, W1 and W2 are both learnable parameter matrices of the gating network, φ is the ELU activation function, and σ is the Sigmoid activation function;

[0089] The weighted feature M(D H ,G H ) and the fusion feature F are element-wise multiplied, and then added with the subway feature D H to output the cross-modal fusion feature H F ;

[0090] H F =D H +F⊙M(D H ,G H ),

[0091] where ⊙ is the Hadamard product.

[0092] The application utilizes a gated cross-modal attention fusion module to model the transfer relationship between the subway and the bus system, designs a cross-modal fusion structure based on multi-head attention, and dynamically adjusts the information interaction intensity at different times and stations by introducing a gating mechanism; the mechanism ensures that the bus information is fully fused when the transfer behavior is significant, and the redundant information is automatically inhibited when the influence is weak, thereby realizing fine-grained subway-bus fusion expression.

[0093] The dual-path space-time adaptive fusion module considers the heterogeneity of spatial dependence and temporal dependence, decouples spatial modeling and temporal modeling, models through a dual-branch structure, and introduces a gated attention mechanism in the fusion stage. As shown in Figure 3 the following steps are performed in the dual-path space-time adaptive fusion module:

[0094] First, the position encoding is introduced, and trainable position embedding initial time attention weight and spatial attention weight cross-modal fusion feature H F are introduced for time and space respectively, and the initial time attention weight Ta0 and the spatial attention weight Sa0 are added to obtain the mixed input feature X, X=H F +Ta0+Sa0.

[0095] The mixed input feature X is calculated to obtain the time attention feature Ta(X), and the mixed input feature X is calculated to obtain the spatial attention feature Sa(X),

[0096]

[0097] wherein, the input Q=K=V=X of the attention mechanism indicates that the self-attention mechanism is adopted, Q, K, V represent the query, key and value matrices in the attention mechanism, d k represents the dimension of the key vector in the attention; X T represents the time dimension and the space dimension of the tensor X are transposed to extract the spatial features;

[0098] The mixed input feature X is calculated to obtain the time attention weight and the spatial attention weight, and the time and spatial weight G t ,G s ∈(0,1) B×T×N The time attention feature Ta(X) and the spatial attention feature Sa(X) are multiplied by the time and spatial weight G t ,G s and then added to the mixed input feature X to output the space-time fusion feature F X .

[0099] G t ,G sSplit(sigma(W1delta(W g X))) + F X = X + G t ⊙Ta(X) + G s ⊙Sa(X),

[0100] where W g denotes the trainable weight matrix for attention weight extraction, delta is a Swish activation function, W1 is a linear layer trainable weight matrix, sigma is a Sigmoid activation function for limiting the weight value in the interval [0, 1], and is a Hadamard product;

[0101] The spatio-temporal fusion feature F X is finally outputted after nonlinear transformation via a multi-layer perception MLP and a residual connection, and is taken as the prediction result of the subway flow;

[0102]

[0103] wherein, MLP is a multi-layer perception, used for feature reconstruction, is a dimension transposition on the fusion feature F X , processing spatial structure information, and H(X) represents the final output feature of the double-path spatio-temporal adaptive fusion module, in the format of [B, T future , N], T future is the length of the future time window.

[0104] The application further improves the modeling capability of heterogeneous space and time dependence through the double-path spatio-temporal adaptive fusion module, which adopts a double-branch structure to extract spatial and temporal features in the subway passenger flow data respectively, and fuses them through a gating mechanism; the spatial branch captures the structural dependence between different stations, the temporal branch extracts periodic and trend changes, and finally the feature is reorganized and optimized through a fully connected network and a residual connection to output the final passenger flow prediction result.

[0105] (3) The dynamic spatio-temporal decomposition and fusion network is trained by using the training set and the test set, the mean absolute error MAE, the root mean square error RMSE and the weighted average absolute error percentage WMAPE are used as performance evaluation indexes, and the optimized dynamic spatio-temporal decomposition and fusion network is used for subway passenger flow prediction.

[0106] The experimental data selected in this embodiment is derived from the rail transit and bus system in Beijing, China, covering the full passenger card record from May 1 to May 31, 2019. This embodiment selects the central area of Beijing, with a geographical range of 115.97° to 116.75° east longitude and 39.67° to 40.24° north latitude, covering an area of about 92.5 square kilometers, including 329 subway stations and bus stations within a 500-meter range; the total data amount is about 99.09 million subway card records and 60.19 million bus card records.

[0107] The data set is divided according to time sequence, i.e. according to the following time nodes, the training set from May 1 to May 23, 2019, and the test set from May 24 to May 31, 2019; each trip record contains passenger ID, boarding / alighting station, departure / arrival time and corresponding latitude and longitude coordinates; to improve data quality, abnormal records from 00:00 to 05:00 every day are excluded. Subway and bus passenger flow data are aggregated at 15-minute intervals to generate inbound and outbound sequences. All model training and inference processes are completed on a workstation configured with Intel i9-12900 CPU, 64GB memory, NVIDIA RTX 5080 GPU. In addition, to assist in analyzing the time variation and spatial distribution characteristics of subway passenger flow, radial column charts are used to visualize the inbound flow of 329 subway stations to observe the flow changes in different time periods within a day and the spatial distribution of peak stations.

[0108] The present application adopts mean absolute error (MAE), root mean square error (RMSE) and weighted average absolute error percentage (WMAPE) as performance evaluation indicators, and verifies the independent contribution of each module through ablation experiment.

[0109] The dynamic spatio-temporal decomposition and fusion network proposed in the present application performs excellent performance in the task of short-term subway passenger flow prediction. To verify the effectiveness of each core module of the present application in the task of subway passenger flow prediction, an ablation experiment of the system is designed to test the combination of the time series decomposition module, the graph convolution module (static or dynamic), the gated cross-modal attention fusion module and the double-path spatio-temporal adaptive fusion module. Table 1 shows the ablation experiment analysis of each module of the model.

[0110] Table 1 Ablation experiment analysis

[0111]

[0112] As can be seen from Table 1, after replacing the graph convolution with a dynamic graph convolution, the model has obvious improvement in various indicators, indicating that dynamic modeling of inter-site dependence plays an important role in depicting real passenger flow changes. At the same time, the introduction of the time series decomposition module can further improve the prediction accuracy, especially under the dynamic graph convolution, by separating long-term trends and short-term disturbances, effectively enhancing the model's understanding of the time series structure. In addition, the addition of the gated cross-modal attention fusion module significantly improves the model's ability to model bus and subway transfer behavior, by dynamically adjusting the fusion weight of different modal information, enabling the model to more accurately reflect the complex interaction between multi-modal transportation systems. In various configurations, the introduction of the dual-path spatio-temporal adaptive fusion module brings stable performance improvement, verifying the universality and robustness of the module in spatio-temporal decoupling modeling. Overall, the four modules work best when used together, achieving the lowest MAE (14.38), RMSE (24.09), and WMAPE (10.69) on the Beijing real data set, fully demonstrating the significant advantages of the invention in modeling transfer behavior and dynamic spatio-temporal relationships.

[0113] Figure 4 Figure 1 shows the spatial distribution of subway passenger flow prediction error for 329 subway stations in Beijing on May 25, 2019 from 8:00 to 8:15 am, Figure 5 Figure 2 shows the spatial distribution of subway passenger flow prediction error for 329 subway stations in Beijing on May 25, 2019 from 12:00 to 12:15 pm. Station markers are color-coded using a cool-to-warm color gradient, with cooler colors (blue) indicating lower prediction error and warmer colors (red) representing higher error. Marker sizes are scaled to highlight outliers; stations with MAE over 30 are displayed at 60% of the standard point size. Analysis shows that 75% of the stations have MAE values below 20, indicating consistent prediction accuracy during peak and off-peak hours.

[0114] Embodiment 2: A subway passenger flow prediction system considering subway-bus transfer in the present invention, comprising:

[0115] a historical data acquisition module for acquiring historical passenger flow D of subway stations in a certain area within a period of time t and historical passenger flow G of corresponding bus stations around each subway station t forming a data set and dividing the data set into a training set and a test set according to time sequence.

[0116] The network construction module is used to build a dynamic spatiotemporal decomposition and fusion network. This network includes a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a dual-path spatiotemporal adaptive fusion module. The time series decomposition module is used to analyze the historical passenger flow D at subway stations. t Historical passenger flow G at bus stops t Structural decomposition is performed to extract multi-scale temporal features of subway and bus systems, outputting fluctuation and trend terms for both. These fluctuation and trend terms are then input into corresponding dynamic graph convolutional modules to dynamically extract spatial correlation features between stations, outputting the subway feature D. H and public transport features G H ; and will subway feature D H and public transport features G H The input-gated cross-modal attention fusion module extracts features of subway and bus transfer behavior and outputs cross-modal fusion features H. F Then fuse the cross-modal features H F The input dual-path spatiotemporal adaptive fusion module integrates features in the temporal and spatial dimensions, and finally outputs the prediction result.

[0117] The time series decomposition module will analyze the historical passenger flow D of subway stations. t Historical passenger flow G at bus stops t The data are input into the corresponding time series decomposition modules, where moving average operations are performed to decompose historical passenger flow into long-term trend terms and short-term fluctuation terms. These are then processed by a multilayer perceptron (MLP) to obtain the subway trend term. Subway fluctuation items Public transport trends With public transport fluctuations

[0118]

[0119] MetroRemainder t =D t -MetroTrend t ,

[0120] Among them, MetroTrend t MetroRemainder represents the metro trend term at time t. t Let i represent the subway fluctuation term at time t, and ω be the sliding window size. When ω is 3, the value of i is 0, 1, or 2. Let D represent the historical passenger flow sequence of the subway at time t. t-i This represents the historical passenger flow sequence of the subway at time ti.

[0121]

[0122] BusRemainder t = G t -BusTrend t

[0123] wherein, MetroTrend represents the bus historical passenger flow sequence at the t time, G t-i MetroTrend represents the bus historical passenger flow sequence at the t-i time, BusTrend t MetroTrend represents the bus trend item at the t time, BusRemainder t MetroRemainder represents the bus fluctuation item at the t time.

[0124] obtain the subway topology structure diagram of a certain area in the dynamic graph convolution module and convert it into a static adjacency matrix The adjacency relationship weight A between the i th station and the j th station in the static adjacency matrix ij The calculation formula is as follows:

[0125]

[0126] After the metro trend item MetroTrend is sequentially input into the 2-layer one-dimensional convolution layer, the 2-layer linear layer and the Sigmoid activation function, the first subway dynamic feature F is obtained D14 , the second subway dynamic feature F is obtained after the metro fluctuation item MetroRemainder is sequentially input into the 2-layer one-dimensional convolution layer, the 2-layer linear layer and the Sigmoid activation function D24 ;

[0127] F D11 = Conv1D(MetroTrend), F D12 = Conv1D(F D11 ), F D13 = Dense(Flatten(F D12 )),

[0128] F D14 = σ(Dense(Flatten(F D13 ))), F D21 = Conv1D(MetroRemainder),

[0129] F D22 = Conv1D(F D21 ), F D23 = Dense(Flatten(F D22 )), F D24 = σ(Dense(Flatten(F D23))),

[0130] wherein Conv1D is a one-dimensional convolution layer, Dense represents a fully connected layer, Flatten represents a flattening operation, and σ is a Sigmoid activation function;

[0131] a first subway dynamic feature F D14 and a second subway dynamic feature F D24 respectively element-wise multiply with the static adjacency matrix A static to obtain a first subway dynamic adjacency matrix A D1 and a second subway dynamic adjacency matrix A D2 ;

[0132] A D1 = A static ⊙F D14 ,

[0133] A D2 = A static ⊙F D24 ,

[0134] wherein ⊙ is a Hadamard product;

[0135] The first subway dynamic adjacency matrix A D1 and the subway trend term MetroTrend are further propagated through the graph convolution propagation mechanism, and the feature propagation between nodes based on the dynamic adjacency matrix is used to obtain subway trend features H D1 ; the second subway dynamic adjacency matrix A D2 and the subway fluctuation term MetroRemainder are further propagated through the graph convolution propagation mechanism, and the feature propagation between nodes based on the dynamic adjacency matrix is used to obtain subway fluctuation features H D2 , the subway trend features H D1 and the subway fluctuation features H D2 are fused to obtain subway features D H ;

[0136] H D1 = φ(A D1 ·MetroTrend·W l1 ), H D2 = φ(A D2 ·MetroRemainder·W l2 ),

[0137] D H = H D1 + H D2 , wherein W l1 , W l2 are learnable parameters, and φ is an Elu activation function;

[0138] The bus trend item BusTrend is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a first bus dynamic feature F G14 The bus fluctuation item BusRemainder is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a second bus dynamic feature F G24 ;

[0139] F G11 =Conv1D(BusTrend), F G12 =Conv1D(F G11 ), F G13 =Dense(Flatten(F G12 )),

[0140] F G14 =σ(Dense(Flatten(F G13 ))), F G21 =Conv1D(BusRemainder),

[0141] F G22 =Conv1D(F G21 ), F G23 =Dense(Flatten(F G22 )), F G24 =σ(Dense(Flatten(F G23 ))),

[0142] The first bus dynamic feature F G14 and the second bus dynamic feature F G24 are respectively element-wise multiplied with the static adjacency matrix A static to obtain a first bus dynamic adjacency matrix A G1 and a second bus dynamic adjacency matrix A G2 ;

[0143] A G1 =A static ⊙σ(F G14 ), A G2 =A static ⊙σ(F G24 ),

[0144] The first bus dynamic adjacency matrix A G1 and the bus trend item BusTrend are further propagated through a graph convolution propagation mechanism, and the feature propagation between nodes is performed based on the dynamic adjacency matrix to obtain a bus trend feature H G1 , and the second bus dynamic adjacency matrix A G2With the bus fluctuation term BusRemainer, the bus fluctuation characteristics H are obtained by using the dynamic adjacency matrix based on the node feature propagation mechanism through the graph convolution propagation mechanism G2 The bus trend characteristics H G1 are obtained by using the dynamic adjacency matrix based on the node feature propagation mechanism through the graph convolution propagation mechanism G2 The bus trend characteristics H H are obtained by using the dynamic adjacency matrix based on the node feature propagation mechanism through the graph convolution propagation mechanism H ;

[0145] H G1 = φ(A G1 · BusTrend· W l3 ), H G2 = φ(A G2 · BusRemainer· W l4 ), G H = H G1 + H G2 ,

[0146] wherein W l3 , W l4 are learnable parameters.

[0147] In the gated cross-modal attention fusion module, the subway characteristics D H and the bus characteristics G H are dimensionally spliced and then input into the multi-head attention mechanism for processing to output the weighted characteristics M(D H , G H ).

[0148] Q i = Dense(Concat(D H , G H )), K i = Dense(Concat(D H , G H )),

[0149] V i = Dense(Concat(D H , G H )),

[0150] M(D H , G H ) = Concat(Head1, …, Head h ) W O ,

[0151] wherein Head i represents the i-th attention weight, Attention(Q i , K i , V idenotes the multi-head attention mechanism,

[0152] Q i ,K i ,V i denotes the query, key, value matrix in the i-th attention head, d k denotes the dimension of the key vector in the attention, Concat denotes the dimension concatenation function in deep learning, W O is a trainable weight parameter in the attention mechanism;

[0153] subway feature D H and bus feature G H are added to input the gating network, and after being processed by the nonlinear gating function and the Sigmoid activation function, the fusion feature F is output;

[0154] M H = φ (W1 (D H + G H )), F = σ (W2M H ),

[0155] wherein, M H is the hidden layer feature in the gating network, F is the fusion feature output by the gating network, and the control fusion strength, W1 and W2 are both the learnable parameter matrix of the gating network, φ is the ELU activation function, and σ is the Sigmoid activation function;

[0156] The weighted feature M (D H , G H ) and the fusion feature F are element-wise multiplied, and then added to the subway feature D H to output the cross-modal fusion feature H F ;

[0157] H F = D H +F⊙M (D H , G H ),

[0158] wherein, ⊙ is the Hadamard product.

[0159] The initial time attention weight and the spatial attention weight are introduced in the double-path space-time adaptive fusion module The cross-modal fusion feature H F , the initial time attention weight Ta0 and the spatial attention weight Sa0 are added to obtain the mixed input feature X, X = H F + Ta0 + Sa0.

[0160]

[0160] The mixed input feature X is subjected to time attention calculation to obtain a time attention feature Ta(X), and the mixed input feature X is subjected to space attention calculation to obtain a space attention feature Sa(X),

[0161]

[0162] wherein the input Q = K = V = X of the attention mechanism indicates that a self-attention mechanism is adopted, Q, K, and V represent query, key, and value matrices in the attention mechanism, d k represents the dimension of the key vector in the attention; X T represents that the time dimension and the space dimension of the tensor X are transposed to extract spatial features;

[0163] The mixed input feature X is subjected to time attention weight and space attention weight extraction to generate time and space weights G t ,G s ∈(0, 1) B×T×N The time attention feature Ta(X) and the space attention feature Sa(X) are respectively multiplied by the time and space weights G t ,G s and then added to the mixed input feature X to output a space-time fusion feature F X ;

[0164] G t ,G s = Split(σ(W1δ(W g X))), F X = X + G t ⊙ Ta(X) + G s ⊙ Sa(X),

[0165] wherein W g represents a trainable weight matrix for attention weight extraction, δ is a Swish activation function, W1 is a linear layer trainable weight matrix, σ is a Sigmoid activation function for limiting the weight value in the interval [0, 1], and ⊙ is a Hadamard product;

[0166] The space-time fusion feature F X is subjected to nonlinear transformation via a multi-layer perception MLP and a residual connection to finally output a feature H(X), which serves as a prediction result of the subway flow.

[0167]

[0168] wherein MLP is a multi-layer perception for feature reconstruction, is a dimension transposition on the fusion feature F X to process spatial structure information, and H(X) represents the final output feature of the double-path space-time adaptive fusion module.

[0169] A network optimization module is configured to train the dynamic spatio-temporal decomposition and fusion network by using the training set and the test set, and to use mean absolute error (MAE), root mean square error (RMSE) and weighted mean absolute percentage error (WMAPE) as performance evaluation indexes, and to use the optimized dynamic spatio-temporal decomposition and fusion network to perform subway passenger flow prediction.

[0170] Embodiment 3: An electronic device according to the present application, comprising a processor and a storage medium;

[0171] The storage medium is configured to store instructions;

[0172] The processor is configured to operate according to the instructions to perform the steps of the method as described above.

[0173] Embodiment 4: A computer readable storage medium according to the present application, having stored thereon a computer program which, when executed by a processor, implements the steps of the method as described above.

Claims

1. A subway passenger flow prediction method considering subway-bus interchange, characterized by, The method comprises the following steps: (1) Obtain the historical passenger flow D of a certain area of subway stations in a period of time t and the historical passenger flow G of the corresponding bus stations around each subway station t Form a data set, and divide the data set into a training set and a test set according to the time sequence; (2) Construct a dynamic spatiotemporal decomposition and fusion network, which includes a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a dual-path spatiotemporal adaptive fusion module; the time series decomposition module is used to analyze the historical passenger flow D of subway stations. t Historical passenger flow G at bus stops t Structural decomposition is performed to extract multi-scale temporal features of subway and bus systems, outputting fluctuation and trend terms for both. These fluctuation and trend terms are then input into corresponding dynamic graph convolutional modules to dynamically extract spatial correlation features between stations, outputting the subway feature D. H and public transport features G H ; and will subway feature D H and public transport features G H The input gated cross-modal attention fusion module extracts features of subway and bus transfer behavior and outputs cross-modal fusion feature H. F Then fuse the cross-modal features H F The input dual-path spatiotemporal adaptive fusion module integrates features in the temporal and spatial dimensions, and finally outputs the prediction result. (3) training the dynamic spatio-temporal decomposition and fusion network by using the training set and the test set, taking the mean absolute error, the root mean square error and the weighted average absolute error percentage as performance evaluation indexes, and predicting the subway passenger flow by using the optimized dynamic spatio-temporal decomposition and fusion network. 2.The metro passenger flow prediction method considering metro-bus transfer according to claim 1, wherein: The time series decomposition module specifically performs the following steps: the historical passenger flow D t and the historical passenger flow G t of the bus station are respectively input into the corresponding time series decomposition module, the historical passenger flow is decomposed into a long-term trend item and a short-term fluctuation item through a moving average operation, and then the subway trend item subway fluctuation item bus trend item and bus fluctuation item 3.The metro passenger flow prediction method considering metro-bus transfer according to claim 1, wherein: The specific steps executed in the dynamic graph convolution module are as follows: obtaining a subway topology structure diagram of a certain area and converting it into a static adjacency matrix The subway trend item MetroTrend is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a first subway dynamic feature F D14 The subway fluctuation item MetroRemainder is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a second subway dynamic feature F D24 ; the first subway dynamic feature F D14 and the second subway dynamic feature F D24 are respectively element-wise multiplied with the static adjacency matrix A static to obtain a first subway dynamic adjacency matrix A D1 and a second subway dynamic adjacency matrix A D2 ; the first subway dynamic adjacency matrix A D1 and the subway trend item MetroTrend are subjected to graph convolution propagation mechanism, and feature propagation between nodes is performed based on the dynamic adjacency matrix to obtain a subway trend feature H D1 ; the second subway dynamic adjacency matrix A D2 and the subway fluctuation item MetroRemainder are subjected to graph convolution propagation mechanism, and feature propagation between nodes is performed based on the dynamic adjacency matrix to obtain a subway fluctuation feature H D2 The subway trend feature H D1 and the subway fluctuation feature H D2 are fused to obtain a subway feature D H ; The bus trend item BusTrend is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a first bus dynamic feature F G14 The bus fluctuation item BusRemainder is sequentially input into a two-layer one-dimensional convolution layer, a two-layer linear layer, and a Sigmoid activation function to obtain a second bus dynamic feature F G24 The first bus dynamic feature F G14 and the second bus dynamic feature F G24 are respectively element-wise multiplied with the static adjacency matrix A static to obtain a first bus dynamic adjacency matrix A G1 and a second bus dynamic adjacency matrix A G2 The first bus dynamic adjacency matrix A G1 and the bus trend item BusTrend are further propagated through a graph convolution propagation mechanism to propagate features between nodes based on the dynamic adjacency matrix to obtain a bus trend feature H G1 The second bus dynamic adjacency matrix A G2 and the bus fluctuation item BusRemainder are further propagated through a graph convolution propagation mechanism to propagate features between nodes based on the dynamic adjacency matrix to obtain a bus fluctuation feature H G2 The bus trend feature H G1 and the bus fluctuation feature H G2 are fused to obtain a bus feature G H . 4.The metro passenger flow prediction method considering metro-bus transfer according to claim 1, wherein: The specific steps executed in the gating cross-modal attention fusion module are as follows: subway features D H and bus features G H are input into a multi-head attention mechanism after dimension splicing for processing and output weighted features M(D H ,G H ); subway features D H and bus features G H are added and input into a gating network, and fusion features F are output after processing by a nonlinear gating function and a Sigmoid activation function; Weighted Feature M(D H ,G H ) and the fusion feature F are multiplied element by element, and then added with the subway feature D H , to output the cross-modal fusion feature H F . 5.The metro passenger flow prediction method considering metro-bus transfer according to claim 1, characterized in that: The double-path space-time adaptive fusion module specifically performs the following steps: introducing an initial time attention weight and the spatial attention weight The cross-modal fusion feature H F The initial time attention weight Ta0 and the spatial attention weight Sa0 are added to obtain a mixed input feature X; The input feature X is subjected to time attention calculation to obtain a time attention feature Ta(X), and the mixed input feature X is subjected to spatial attention calculation to obtain a spatial attention feature Sa(X); the mixed input feature X is subjected to time attention weight and spatial attention weight extraction to generate time and spatial weights G t ,G s ∈(0,1) B×T×N The time attention feature Ta(X) and the spatial attention feature Sa(X) are respectively multiplied by the time and spatial weights G t ,G s After being added to the mixed input feature X, a spatio-temporal fusion feature F is output X ; the spatio-temporal fusion feature F X is subjected to nonlinear transformation via a multi-layer perception MLP and a residual connection to finally output a feature H(X), which is taken as a prediction result of the subway flow.

6. A subway passenger flow prediction system considering subway-bus interchange, characterized by, The method comprises the following steps: a historical data acquisition module configured to acquire historical passenger flow D of a subway station in a certain area within a period of time t and historical passenger flow G of a bus station corresponding to each subway station t forming a data set and dividing the data set into a training set and a test set according to time sequence; The network construction module is used to build a dynamic spatiotemporal decomposition and fusion network. This network includes a time series decomposition module, a dynamic graph convolution module, a gated cross-modal attention fusion module, and a dual-path spatiotemporal adaptive fusion module. The time series decomposition module is used to analyze the historical passenger flow D at subway stations. t Historical passenger flow G at bus stops t Structural decomposition is performed to extract multi-scale temporal features of subway and bus systems, outputting fluctuation and trend terms for both. These fluctuation and trend terms are then input into corresponding dynamic graph convolutional modules to dynamically extract spatial correlation features between stations, outputting the subway feature D. H and public transport features G H ; and will subway feature D H and public transport features G H The input-gated cross-modal attention fusion module extracts features of subway and bus transfer behavior and outputs cross-modal fusion features H. F Then fuse the cross-modal features H F The input dual-path spatiotemporal adaptive fusion module integrates features in the temporal and spatial dimensions, and finally outputs the prediction result. The network optimization module is configured to train the dynamic spatio-temporal decomposition and fusion network by using the training set and the test set, take the mean absolute error, the root mean square error and the weighted average absolute error percentage as performance evaluation indexes, and predict the subway passenger flow by using the optimized dynamic spatio-temporal decomposition and fusion network.

7. The subway passenger flow prediction system considering subway-bus transfer according to claim 1, characterized in that: The subway feature D H and the bus feature G H are input into a multi-head attention mechanism after dimension splicing, processed, and output as weighted features M(D H ,G H ); the subway feature D H and the bus feature G H are added, input into a gating network, processed by a nonlinear gating function and a Sigmoid activation function, and output as a fusion feature F; Weighted Feature M(D H ,G H ) and the fusion feature F are multiplied element by element, and then added with the subway feature D H to output the cross-modal fusion feature H F . 8.The subway passenger flow forecasting system considering subway-bus transfer according to claim 1, wherein: The initial temporal attention weight is introduced in the dual-path space-time adaptive fusion module The spatial attention weight The cross-modal fusion feature H F The initial temporal attention weight Ta0 and the spatial attention weight Sa0 are added to obtain a mixed input feature X; The input feature X is subjected to time attention calculation to obtain a time attention feature Ta(X), and the mixed input feature X is subjected to spatial attention calculation to obtain a spatial attention feature Sa(X); the mixed input feature X is subjected to time attention weight and spatial attention weight extraction to generate time and spatial weights G t ,G s ∈(0,1) B×T×N The time attention feature Ta(X) and the spatial attention feature Sa(X) are respectively multiplied by the time and spatial weights G t ,G s After being added to the mixed input feature X, a spatio-temporal fusion feature F is output X The spatio-temporal fusion feature F X is subjected to nonlinear transformation via a multi-layer perception MLP and a residual connection to finally output a feature H(X), which is taken as a prediction result of the subway flow.

9. An electronic device, comprising: The device comprises a processor and a storage medium. The storage medium is configured to store instructions. The processor is configured to operate according to the instructions to perform the steps of the method of any one of claims 1-5.

10. A computer readable storage medium having stored thereon a computer program, characterized in that The program, when executed by the processor, implements the steps of the method of any one of claims 1-5.

Images