Urban peak demand forecasting method and system based on dynamic spatiotemporal hypergraph representation learning

Through dynamic spatiotemporal hypergraph representation learning method, a spatiotemporal dynamic weighted supergraph is constructed, weather and OD data are embedded, and convolutional modules are defined. The limitations of existing OD demand prediction methods in urban peak time and interaction influences between multiple factors are solved, and the limitations of existing OD demand prediction methods are improved. The accuracy of OD demand forecasting and traffic management decision support are provided.

CN118014135BActive Publication Date: 2024-12-06HARBIN INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410153926.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-02
Publication Date
2024-12-06
Estimated Expiration
2044-02-02

AI Technical Summary

Technical Problem

The existing OD demand prediction methods have limitations in dealing with the interaction effects of traffic demand and multivariate factors during peak time in cities, and it is difficult to accurately capture the changes in traffic demand and the interactive effects of multiple factors during special periods.

Method used

Using a method based on dynamic spatiotemporal hypergraph representation learning, by constructing a spatiotemporal dynamic weighted hypergraph, embed weather information and OD travel data, calculate the super edge weight sum, define the hypergraph spatial convolution module and the time convolution module, and combine it with deep learning Methods: The urban peak OD demand prediction model is constructed.

Benefits of technology

It realizes a deep-level representation of the spatiotemporal characteristics of urban traffic flow, improves the accuracy of OD demand prediction and captures multiple relationships during critical periods, and enhances the environmental adaptability of the model and traffic management decision support.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118014135B_ABST
    Figure CN118014135B_ABST
Patent Text Reader

Abstract

The present invention provides a method for predicting urban peak demand based on dynamic spatiotemporal hypergraph representation learning, the method comprising the following steps: S1, collecting and cleaning urban peak OD demand prediction data; S2, constructing a spatiotemporal dynamic weighted hypergraph; S3, constructing an urban peak OD demand prediction model to complete urban peak demand prediction. The present invention achieves a deep characterization of the spatiotemporal characteristics of urban traffic flow through multi-dimensional data fusion and characterization processing. Compared with the prior art, the present invention exhibits higher accuracy in capturing and representing complex spatiotemporal dependencies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of big data mining and traffic engineering, and specifically to a method and system for predicting urban peak demand based on dynamic spatiotemporal hypergraph representation learning. Background Art

[0002] Urban peak-time OD (Origin-Destination) travel is a major component of urban travel activities. Predicting changes in peak-time OD demand is important in the fields of traffic engineering and urban planning. Currently, existing technologies and patents mainly focus on full-time OD demand prediction. Due to the large fluctuations in OD demand during certain special periods (such as morning and evening peaks), in order to pursue the overall prediction effect, existing models often sacrifice the accuracy of predictions during these key periods, resulting in the inability of prediction models to effectively predict traffic demand during these periods, which in turn affects the efficiency and accuracy of traffic control and management decisions.

[0003] In addition, OD demand prediction not only involves temporal changes, but is also closely related to spatial factors. Many existing prediction models perform spatiotemporal prediction based on ordinary graph structures, which can provide good prediction results when only considering changes in road networks and travel volume. However, changes in OD demand are affected by many factors, including special OD areas, POI (point of interest) information, weather conditions, and weekday time information. These factors are difficult to fully reflect in traditional spatiotemporal graph features because ordinary graph structures can usually only represent binary relationships, while in fact, the formation of OD demand is the result of the interaction of multiple complex factors. For example, weather conditions may affect travel patterns in multiple areas at the same time, while weekday time information is closely related to travel demand in the entire city. Therefore, a more complex data structure is needed to represent this multivariate relationship and make accurate predictions on this basis.

[0004] In summary, the existing OD demand forecasting methods have obvious limitations in dealing with special time periods and comprehensive multiple factors. There is an urgent need for a forecasting method that can accurately capture the changes in urban peak traffic demand and the interactive effects of multiple factors. Summary of the invention

[0005] In order to solve technical problems such as the limitations of existing OD demand forecasting methods in dealing with urban peak-time traffic demand and the interactive influence of multiple factors, the present invention provides an urban peak-time demand forecasting method based on dynamic spatiotemporal hypergraph representation learning.

[0006] The method for predicting urban peak demand based on dynamic spatiotemporal hypergraph representation learning of the present invention comprises the following steps:

[0007] S1. Collect and clean urban peak OD demand forecast data;

[0008] S2, constructing a spatiotemporal dynamic weighted hypergraph;

[0009] S3. Build a city peak OD demand forecasting model to complete the city peak demand forecasting.

[0010] Furthermore, S2 specifically includes the following steps:

[0011] Determine the dynamic weighted hypergraph infrastructure; dynamically embed weather information into node V; dynamically integrate OD travel data into hyperedge E; calculate hypergraph G h The hyperedge weight W and degree.

[0012] Furthermore, weather information is dynamically embedded into the node, specifically including the following steps:

[0013] Step 2-2-1: Determine the weather conditions t and intensity i t

[0014] Weather conditions c at different time periods t t Including: clear, cloudy, haze, rain, snow;

[0015] Weather conditions at different times t Intensity t Includes: mild, moderate and severe;

[0016] Step 2-2-2: Embed weather information into the attributes of node V

[0017] The weather status and intensity of different road network partitions v determined in step 2-2-1 are respectively embedded in two layers E cond (c t ) and E int (i t ) is encoded and converted into a numerical feature vector, which is dynamically loaded into the attributes of node v along with the step size.

[0018] Furthermore, the OD travel data is dynamically integrated into the hyperedge E, which specifically includes the following steps:

[0019] Step 2-3-1: Construct the OD trip peak time recording matrix.

[0020] With the help of the topological structure of the urban road network, the Dijkstra algorithm is used to reconstruct each travel route in the OD record, and the 0-1 matrix H of the OD travel record is constructed with the road network partition number as the column and each travel record as the row. OD ,

[0021]

[0022] Where m represents the total number of travel records and n represents the total number of road network partitions. m,nIndicates whether the mth OD trip record passes through the nth road network partition. If so, then h m,n =1; otherwise, h m,n =0;

[0023] Step 2-3-2: Calculate H OD The Hamming distance matrix H of the matrix han

[0024] Use Hamming distance to measure the OD trip peak time record matrix H OD , whose Hamming distance matrix H han The calculation formula is as follows:

[0025]

[0026] Among them, H han (i,j) represents the Hamming distance between the i-th row and the j-th row, N is the number of columns, and H OD (i,k) and H OD (j,k) are H OD The value of the i-th row and j-th row in the k-th column of the matrix;

[0027] Step 2-3-3: Cluster OD travel record matrix H OD

[0028] Cluster H OD

[0029] Set the hierarchical clustering threshold th according to the data size, and apply the Ward method to the OD peak time record matrix H obtained in step 2-2-1. OD Perform hierarchical clustering;

[0030] Classification sum H OD

[0031] According to the clustering results, H OD The matrix elements are classified and summed, and the summation results after each clustering are used as a hypergraph G h The hyperedge e.

[0032] Furthermore, we calculate the hypergraph G h The hyperedge weight W and degree include the following steps:

[0033] Step 2-4-1: Calculate the hypergraph G h The weighted correlation matrix H ve

[0034] In the hypergraph G h In the example, the weighted association degree between vertex v and hyperedge e is calculated using the weighted association matrix H ve Representation; Matrix H ve Element h veThe calculation formula is as follows:

[0035]

[0036] Step 2-4-2: Train the hyperedge weight W

[0037] Super Graph G h The hyperedge weight matrix W is composed of the weight w of each hyperedge e e Composition; e By using deep learning methods, the weighted correlation matrix H is obtained by analyzing steps 2-4-1. ve and time feature training;

[0038] Step 2-4-3: Calculate the degree of the hypergraph

[0039] Super Graph G h The degree includes the degree of vertex v and the degree of hyperedge e, which represent the connection strength of the vertex and the coverage of the hyperedge respectively;

[0040] The degree of a vertex v is the weight w of all hyperedges connected to it. e The calculation formula is as follows:

[0041]

[0042] Among them, e represents a hyperedge, w e represents the weight of hyperedge e, h ve is the degree of association between vertex v and hyperedge e;

[0043] The degree of a hyperedge e refers to the number of vertices connected by the hyperedge, and its calculation formula is as follows:

[0044]

[0045] Among them, v is the vertex, h ve is the degree of association between vertex v and hyperedge e.

[0046] Furthermore, S3 specifically includes the following steps:

[0047] Step 3-1 Define the hypergraph spatial convolution module

[0048] Define the hypergraph Laplacian matrix Δ, the calculation formula of the hypergraph Laplacian matrix Δ is as follows:

[0049]

[0050] Define the hypergraph convolution layer. The definition of the hypergraph convolution layer is:

[0051]

[0052] Among them, X (l)represents the vertex feature matrix of the lth layer, Θ (l) represents the learnable parameter matrix of layer l.

[0053] Furthermore, S3 specifically includes the following steps:

[0054] S3-2 Define Hypergraph Temporal Convolution Module

[0055] In order to effectively extract temporal features and allow parallel processing of features, a three-layer convolutional neural network sequence is used in the time block. Each layer of CNN recognizes temporal dependencies with a more efficient architecture through independent convolution operations. The extraction formula of temporal features is as follows:

[0056] X l+1 =σ(σ(θ1·X l +σ(θ2·X l ))+θ3·X l )

[0057] Among them, σ is a nonlinear activation function, X l and X l+1 is the temporal feature of OD travel data, θ1, θ2 and θ3 are convolution kernels.

[0058] Furthermore, S3 specifically includes the following steps:

[0059] S3-3 Define data fusion module

[0060] The features extracted by the specified time and space blocks are integrated using the fusion process to encapsulate the comprehensive spatiotemporal dependencies in the network. The mathematical representation of the data fusion mechanism is as follows:

[0061] X=σ(concat(X OD_temp ·Θ1,X OD_spat ·Θ2,X hour ·Θ3,X dow ·Θ4,X weather ·Θ5))

[0062] Among them, X OD_temp and X OD_spat represent the temporal and spatial characteristics of the OD travel data obtained through steps 3-1 and 3-2, respectively, hour , X dow and X weather They correspond to the time period hour, day of the week and the time characteristics of the weather conditions obtained in step 3-1 respectively. Θ1, Θ2, Θ3, Θ4 and Θ5 represent the learning parameters of the fusion process respectively.

[0063] Furthermore, S3 specifically includes the following steps:

[0064] S3-4 Define the loss function

[0065] Minimize the mean squared error (MSE) loss. This loss is calculated as the average of the squares of the differences between the predicted and actual OD demands. The loss function is expressed as follows:

[0066]

[0067] Where n represents the number of OD pairs, y i represents the peak OD observation value of the i-th sample, It represents the model's predicted value of peak OD demand for the i-th sample.

[0068] The present invention also relates to a city peak demand forecasting system based on dynamic spatiotemporal hypergraph representation learning, comprising a computer module, which applies the above-mentioned city peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning.

[0069] Beneficial Effects

[0070] 1. The present invention achieves a deep characterization of the spatiotemporal characteristics of urban traffic flow through multi-dimensional data fusion and characterization processing. Compared with the prior art, the present invention shows higher accuracy in capturing and representing complex spatiotemporal dependencies.

[0071] 2. The present invention adopts a hypergraph representation learning method, which can more effectively capture and express multi-dimensional relationships in traffic flow. Compared with traditional graph models, the present invention has obvious advantages in processing multi-point relationships in urban traffic networks.

[0072] 3. The present invention incorporates the influence of weather and calendar data into the model, providing a novel way to integrate environmental factors that affect traffic flow and enhancing the environmental adaptability of the model when predicting urban traffic flow.

[0073] 4. The present invention significantly improves the value of the model in practical applications, especially for decision support for traffic management and planning, by specifically analyzing and predicting the OD demand during key time periods (such as morning and evening peak hours). BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 This is a flow chart of urban peak OD demand prediction based on dynamic spatiotemporal hypergraph representation learning according to the present invention.

[0075] Figure 2 This is a schematic diagram of the spatiotemporal dynamic weighted hypergraph model for urban peak OD demand prediction according to the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0077] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0078] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0079] Step S1: Collect and clean city peak OD demand forecast data

[0080] Collect multi-dimensional data including road network partition center coordinates, historical OD travel data, city peak hours, weekdays and historical weather information.

[0081] Step S2: Constructing a spatiotemporal dynamic weighted hypergraph

[0082] Step 2-1: Determine the dynamic weighted hypergraph infrastructure

[0083] After data preparation is completed, construct the hypergraph G h =(V,E,W) to represent the urban road network. The road network partitions are regarded as vertices v∈V in the hypergraph, and the OD travel routes are hyperedges e∈E connecting these vertices.

[0084] Step 2-2: Dynamically embed weather information into node V

[0085] Step 2-2-1: Determine the weather conditions t and intensity i t

[0086] Weather conditions t : The weather conditions c of different time periods t considered by the present invention t Mainly includes: clear, cloudy, haze, rain, snow, etc.

[0087] Weather intensity t : The present invention considers different time periods t weather conditions c t The intensity of t Mainly include: mild, moderate and severe etc.

[0088] Step 2-2-2: Embed weather information into the attributes of node V

[0089] The weather status and intensity of different road network partitions v determined in step 2-2-1 are respectively embedded in two layers Econd (c t ) and E int (i t ) is encoded and converted into a numerical feature vector, which is dynamically loaded into the attributes of node v along with the step size.

[0090] Step 2-3: Dynamically integrate OD travel data into hyperedge E

[0091] In order to balance the detail and computational performance of OD travel data, the present invention first screens OD travel records by hour and converts them into a peak time record matrix for road network partitions, and then clusters the OD travel peak time record matrix using Ward hierarchical clustering, and the sum of each clustering result is used as a hyperedge. The process of OD travel records to the hypergraph is dynamically loaded according to the step size, rather than completed in one go.

[0092] Step 2-3-1: Construct the OD trip peak time recording matrix.

[0093] With the help of the topological structure of the urban road network, the Dijkstra algorithm is used to reconstruct each travel route in the OD record, and the 0-1 matrix H of the OD travel record is constructed with the road network partition number as the column and each travel record as the row. OD ,

[0094]

[0095] Where m represents the total number of travel records and n represents the total number of road network partitions. m,n Indicates whether the mth OD trip record passes through the nth road network partition. If so, then h m,n =1; otherwise, h m,n =0.

[0096] Step 2-3-2: Calculate H OD The Hamming distance matrix H of the matrix han

[0097] The present invention uses the Hamming distance to measure the OD trip peak time record matrix H OD , whose Hamming distance matrix H han The calculation formula is as follows:

[0098]

[0099] Among them, H han (i,j) represents the Hamming distance between the i-th row and the j-th row, N is the number of columns, and H OD (i,k) and H OD (j,k) are H OD The value in the kth column of the matrix with the i-th row and j-th row.

[0100] Step 2-3-3: Cluster OD travel record matrix H OD

[0101] Cluster H OD . With the help of the Hamming distance matrix H from step 2-2-2 han , and the road network partition ID of step 1, set the hierarchical clustering threshold th reasonably according to the data size, and apply the Ward method to the OD travel peak time record matrix H obtained in step 2-2-1 OD Perform hierarchical clustering.

[0102] Classification sum H OD According to the clustering results, H OD The matrix elements are classified and summed, and the summation results after each clustering are used as a hypergraph G h The hyperedge e.

[0103] Step 2-4: Compute the hypergraph G h The hyperedge weight W and degree

[0104] After steps 2-3, the hypergraph G has been completed h Before using graph neural networks to predict OD demand, it is necessary to calculate the weights and degrees of vertices and hyperedges in the hypergraph to understand the structure and characteristics of the hypergraph.

[0105] Step 2-4-1: Calculate the hypergraph G h The weighted correlation matrix H ve

[0106] In the hypergraph G h In the example, the weighted association degree between vertex v and hyperedge e is calculated using the weighted association matrix H ve This weighted representation not only shows whether there is a connection between vertices and hyperedges, but also quantifies the strength of this connection. ve Element h ve The calculation formula is as follows:

[0107]

[0108] Step 2-4-2: Train the hyperedge weight W

[0109] Super Graph G h The hyperedge weight matrix W is composed of the weight w of each hyperedge e e Composition. e By using deep learning methods, the weighted correlation matrix H is obtained by analyzing steps 2-4-1. ve and time feature training.

[0110] Step 2-4-3: Calculate the degree of the hypergraph

[0111] Super Graph G h The degree includes the degree of vertex v and the degree of hyperedge e, which represent the connection strength of the vertex and the coverage of the hyperedge respectively.

[0112] The degree of vertex v. The degree of vertex v refers to the weights w of all hyperedges connected to this vertex. e This reflects the connection strength of the vertex in the hypergraph. The calculation formula is as follows:

[0113]

[0114] Among them, e represents a hyperedge, w e represents the weight of hyperedge e, h ve is the degree of association between vertex v and hyperedge e.

[0115] The degree of hyperedge e. The degree of hyperedge e refers to the number of vertices connected by the hyperedge. This represents the coverage of the hyperedge in the hypergraph. Its calculation formula is as follows:

[0116]

[0117] Among them, v is the vertex, h ve is the degree of association between vertex v and hyperedge e.

[0118] Step S3: Build a city peak OD demand forecasting model to complete the city peak demand forecasting

[0119] Based on step S1, time and weather data are obtained, and step S2, a hypergraph G is obtained. h Based on the characteristics of the proposed method, a spatiotemporal dynamic weighted hypergraph model for urban peak OD demand prediction is constructed. The structure of the model is as follows: Figure 2 As shown, it mainly includes six temporal convolution modules, one spatial convolution module, one data fusion module and one fully connected layer.

[0120] Step 3-1: Define the hypergraph spatial convolution module

[0121] Define the hypergraph Laplacian matrix Δ. The present invention uses the hypergraph Laplacian matrix Δ to perform spatial convolution processing, which involves the identity matrix I, the association matrix H ve (Step 2-3-1), hyperedge weight matrix W (Step 2-3-2), vertex degree matrix D v and the hyperedge degree matrix D e (Step 2-3-3). The calculation formula of the hypergraph Laplacian matrix Δ is as follows:

[0122]

[0123] Define the hypergraph convolution layer. The definition of the hypergraph convolution layer is:

[0124]

[0125] Among them, X (l) represents the vertex feature matrix of the lth layer, Θ (l) represents the learnable parameter matrix of layer l.

[0126] Step 3-2: Define the hypergraph temporal convolution module

[0127] In order to effectively extract temporal features and allow parallel processing of features, the present invention uses a three-layer convolutional neural network (CNN) sequence within a time block. Each layer of CNN uses independent convolution operations to identify temporal dependencies with a more efficient architecture. The temporal feature extraction formula is as follows:

[0128] X l+1 =σ(σ(θ1·X l +σ(θ2·X l ))+θ3·X l )

[0129] Among them, σ is a nonlinear activation function, Xl and Xl +1 is the temporal feature of OD travel data, θ1, θ2 and θ3 are convolution kernels.

[0130] Step 3-3: Define data fusion module

[0131] The features extracted by the specified time and space blocks are integrated using the fusion process to encapsulate the comprehensive spatiotemporal dependencies in the network. The mathematical representation of the data fusion mechanism is as follows:

[0132] X=σ(concat(X OD_temp ·Θ1,X OD_spat ·Θ2,X hour ·Θ3,X dow ·Θ4,X weather ·Θ5))

[0133] Among them, X OD_temp and X OD_spat represent the temporal and spatial characteristics of the OD travel data obtained through steps 3-1 and 3-2, respectively, hour , X dow and X weather They correspond to the time period hour, day of the week and the time characteristics of the weather conditions obtained in step 3-1 respectively. Θ1, Θ2, Θ3, Θ4 and Θ5 represent the learning parameters of the fusion process respectively.

[0134] Step 3-4: Define the loss function

[0135] The training objective of the model is to minimize the mean squared error (MSE) loss. This loss is calculated as the average of the squares of the differences between the predicted and actual OD demands. The loss function is expressed as follows:

[0136]

[0137] Where n represents the number of OD pairs, y i represents the peak OD observation value of the i-th sample, It represents the model's predicted value of peak OD demand for the i-th sample.

[0138] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A method for predicting urban peak demand based on dynamic spatiotemporal hypergraph representation learning, the method comprising the following steps: S1. Collect and clean urban peak OD demand forecast data; S2, constructing a spatiotemporal dynamic weighted hypergraph; S3. Build a city peak OD demand forecasting model to complete the city peak demand forecasting; S2 specifically includes the following steps: Determine the dynamic weighted hypergraph infrastructure; dynamically embed weather information into vertices; dynamically integrate OD travel data into hyperedges; calculate the hypergraph G h The hyperedge weight matrix W and degree of ; Dynamically embedding weather information into vertices includes the following steps: Step 2-2-1: Determine the weather conditions t and intensity i t ; Weather conditions c at different time periods t t include: Clear, cloudy, foggy, rainy, snowy; Weather conditions at different times t Intensity t Includes: mild, moderate and severe; Step 2-2-2: Embed weather information into the attributes of vertex v; The weather status and intensity of different road network partitions determined in step 2-2-1 are respectively embedded in two layers E cond (c t ) and E int (i t ) is encoded and converted into a numerical feature vector, which is dynamically loaded into the attributes of vertex v along with the step size.

2. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 1 is characterized in that: Dynamically integrate OD travel data into hyperedge e, including the following steps: Step 2-3-1: Construct the OD trip peak time record matrix; With the help of the topological structure of the urban road network, the Dijkstra algorithm is used to reconstruct each travel route in the OD record, and the OD travel peak time record matrix H is constructed with the road network partition number as the column and each travel record as the row. OD , Where m represents the total number of travel records, n represents the total number of road network partitions; h m,n Indicates whether the mth OD travel record passes through the nth road network partition; if so, h m,n =1; otherwise, h m,n =0; Step 2-3-2: Calculate H OD The Hamming distance matrix H of the matrix han ; Use Hamming distance to measure the OD trip peak time record matrix H OD , whose Hamming distance matrix H han The calculation formula is as follows: Among them, H han (i,j) represents the Hamming distance between the i-th row and the j-th row, N is the number of columns, and H OD (i,k) and H OD (j,k) are H OD The value of the i-th row and j-th row in the k-th column of the matrix; Step 2-3-3: Cluster OD trip peak time record matrix H OD ; Cluster H OD ; Set the hierarchical clustering threshold th according to the data size, and apply the Ward method to the OD peak time record matrix H obtained in step 2-2-1. OD Perform hierarchical clustering; Classification sum H OD ; According to the clustering results, H OD The matrix elements are classified and summed, and the summation results after each clustering are used as a hypergraph G h The hyperedge e.

3. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 1 is characterized in that: Compute the hypergraph G h The hyperedge weight matrix W and degree of the , specifically including the following steps: Step 2-4-1: Calculate the hypergraph G h The weighted correlation matrix H ve ; In the hypergraph G h In the example, the weighted association degree between vertex v and hyperedge e is calculated using the weighted association matrix H ve Representation; Matrix H ve Element h ve The calculation formula is as follows: α is a learnable parameter, is the normalized travel frequency of node v, is the normalized travel distance of the hyperedge e; Step 2-4-2: Train the hyperedge weight matrix W; Super Graph G h The hyperedge weight matrix W is composed of the weight w of each hyperedge e e Composition; e By using deep learning methods, the weighted correlation matrix H is obtained by analyzing steps 2-4-1. ve and time feature training; Step 2-4-3: Calculate the degree of the hypergraph; Super Graph G h The degree includes the degree of vertex v and the degree of hyperedge e, which represent the connection strength of the vertex and the coverage of the hyperedge respectively; The degree of a vertex v is the weight w of all hyperedges connected to it. e The calculation formula is as follows: Among them, e represents a hyperedge, w e represents the weight of hyperedge e, h ve is the degree of association between vertex v and hyperedge e; The degree of a hyperedge e refers to the number of vertices connected by the hyperedge, and its calculation formula is as follows: Among them, v is the vertex, h ve is the degree of association between vertex v and hyperedge e.

4. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 1 is characterized in that: S3 specifically includes the following steps: Step 3-1 defines the hypergraph spatial convolution module; Define the hypergraph Laplacian matrix Δ, the calculation formula of the hypergraph Laplacian matrix Δ is as follows: Define the hypergraph convolution layer. The definition of the hypergraph convolution layer is: Among them, X (l) represents the vertex feature matrix of the lth layer, σ s represents the nonlinear activation function of the spatial convolution module, Θ (l) represents the learnable parameter matrix of layer l.

5. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 4 is characterized in that: S3 specifically includes the following steps: S3-2 defines the hypergraph temporal convolution module; In order to effectively extract temporal features and allow parallel processing of features, a three-layer convolutional neural network sequence is used in the time block. Each layer of CNN recognizes temporal dependencies with a more efficient architecture through independent convolution operations. The extraction formula of temporal features is as follows: X l+1 =s t (s t (θ1·X l +s t (θ2·X l ))+θ3·X l ) Among them, σ t is the nonlinear activation function of the temporal convolution module, X l and X l+1 is the temporal feature of OD travel data, θ1, θ2 and θ3 are convolution kernels.

6. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 5 is characterized in that: S3 specifically includes the following steps: S3-3 defines data fusion module; The features extracted by the specified time and space blocks are integrated using the fusion process to encapsulate the comprehensive spatiotemporal dependencies in the network. The mathematical representation of the data fusion mechanism is as follows: X=σ f (concat(X OD_temp ·Θ1,X OD_spat ·Θ2,X hour ·Θ3,X dow ·Θ4,X weather ·Θ5)) Among them, σ f represents the nonlinear activation function of the data fusion module, X OD_temp and X OD_spat represent the temporal and spatial characteristics of the OD travel data obtained through steps 3-1 and 3-2, respectively, hour , X dow and X weather They correspond to the time period hour, day of the week and the time characteristics of the weather conditions obtained in step 3-1 respectively. Θ1, Θ2, Θ3, Θ4 and Θ5 represent the learning parameters of the fusion process respectively.

7. The urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning according to claim 1 is characterized in that: S3 specifically includes the following steps: S3-4 defines the loss function; Minimize the mean squared error (MSE) loss: This loss is calculated as the average of the squares of the differences between the predicted and actual OD demands; the loss function is expressed as follows: Where q represents the number of OD pairs, y f represents the peak OD observation value of the fth sample, It represents the model's predicted value of peak OD demand for the f-th sample.

8. A city peak demand forecasting system based on dynamic spatiotemporal hypergraph representation learning, characterized in that: It comprises a computer module, which applies the urban peak demand forecasting method based on dynamic spatiotemporal hypergraph representation learning as described in any one of claims 1 to 7.