A method for predicting urban travel demand based on prior causal transfer

By introducing a deep learning method of prior causal transmission in urban travel demand prediction, a dynamic causal graph is constructed and a graph convolution network is used to generate and enhance spatiotemporal causal features, the problem of insufficient causal expression in the existing technology is solved, and the prediction is improved Accuracy and stability.

CN118627702BActive Publication Date: 2024-11-12BEIHANG UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411117155.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-12
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

The existing technology is difficult to effectively express complex potential causal relationships and time-varying causal relationships, resulting in insufficient accuracy and stability of urban travel demand prediction models.

Method used

Deep learning method based on prior causal transmission is adopted to construct dynamic causal graphs through time position information, and distance graphs and interaction graphs are introduced as prior knowledge. Graph convolutional network generation enhances spatio-temporal causal features and improves the robustness of the prediction model.

Benefits of technology

It significantly improves the accuracy and robustness of urban travel demand forecasts, can more effectively capture the causal relationship between time and space, and provide stable prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118627702B_ABST
    Figure CN118627702B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of spatiotemporal data mining, and discloses a method for predicting urban travel demand based on prior causal transfer. The prediction method first collects historical urban travel demand data and constructs a training set; then constructs a deep learning model based on prior causal transfer, including a causal debiasing module, a causal graph learner, a prior causal transfer module, and a fusion prediction module; uses the training set to complete the training of the deep learning model; and finally applies the trained deep learning model to the prediction of future urban travel demand data. The prediction method of the present invention can effectively improve the accuracy and robustness of the prediction model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of spatiotemporal data mining, and specifically is a method for predicting urban travel demand based on a priori causal transmission. Background Art

[0002] Urban travel demand data can effectively reflect the travel behavior patterns of the current urban population. Discovering potential travel patterns of the population from urban travel demand data and giving accurate predictions can help city managers carry out urban planning and improve the quality of life of residents.

[0003] In the existing technology, most methods process spatiotemporal correlation based on graph networks and use static graph structures to describe the relationship between nodes. On the one hand, static graph structures are difficult to express complex potential causal relationships. On the other hand, the causal graph generation process lacks prior knowledge guidance and is often difficult to converge to support the model for stable prediction.

[0004] By learning the causal relationship in spatiotemporal data such as travel demand data, the accuracy and stability of the model can be effectively improved. Existing methods can be divided into prior-based methods, static causal-based methods, and prior-based methods. Prior-based methods use known differential equations to generate causal graphs. The Chinese invention patent with publication number CN115860798A discloses a system and method for predicting urban taxi demand, which combines a spectral domain graph convolutional neural network with a gated recurrent unit, considers the impact of time periodicity and functional similarity between regions on taxi demand, and realizes the prediction of taxi demand in a certain area in the future timestamp. The Chinese invention patent with publication number CN116051171A discloses a joint demand prediction method based on taxis and online car-hailing, extracts the intra-mode features and inter-mode features of taxis and online car-hailing, and can realize the fusion of intra-mode features and inter-mode features, realizing the joint demand prediction of taxis and online car-hailing. However, the causal relationship in the real world is time-varying, so a certain time-varying prior knowledge should be introduced to model the causal transmission process. Summary of the invention

[0005] Based on the deficiencies in the prior art, the present invention uses prior information such as time and position to learn a dynamic spatiotemporal causal graph, models dynamic causal relationships across time slices, and proposes a method for predicting urban travel demand based on prior causal transmission, thereby effectively improving the accuracy and robustness of the prediction model.

[0006] The technical solution of the present invention is as follows:

[0007] A method for predicting urban travel demand based on prior causal transfer, comprising the following steps:

[0008] S1: According to the prediction requirements, the historical urban travel demand data in the area to be predicted is collected and the training set is obtained after standardized processing;

[0009] S2: construct a deep learning model based on prior causal transfer, wherein the deep learning model includes a causal debiasing module, a causal graph learner, a prior causal transfer module and a fusion prediction module;

[0010] In the causal graph learner, a time embedding matrix is ​​constructed based on the time position information, and then a dynamic feature matrix is ​​constructed for the region to be predicted using the time embedding matrix, and finally a causal transfer matrix is ​​constructed based on the dynamic feature matrix;

[0011] In the prior causal transfer module, based on the causal transfer matrix, the distance graph and the interaction graph are fused as prior knowledge, and the graph convolutional network is used to construct the causal transfer process and generate enhanced spatiotemporal causal features;

[0012] S3: Use the training set in S1 to complete the training of the deep learning model;

[0013] S4: Apply the trained deep learning model to predict the desired urban travel demand data.

[0014] Preferably, the urban travel demand includes: taxi demand, bus load, shared bicycle demand or subway passenger flow.

[0015] Preferably, the standardization process in S1 is: dividing the region to be predicted into N The urban travel demand data of each sub-region in each time period is collected.

[0016] Preferably, the causal debiasing module separates all N The sub-regions were divided into p The parallel encoders are used to extract the spatiotemporal causal features of each type of region in different time periods, and then merge them to obtain the spatiotemporal causal features of different time periods.

[0017] Preferably, the process of constructing a causal transfer matrix by the causal graph learner includes:

[0018] (1) Construct a time embedding matrix based on the time position information and construct a projection matrix;

[0019]

[0020] (2) Construct the dynamic feature matrix of all regions:

[0021]

[0022] in For the tThe dynamic feature matrix of all regions in the time period, For the t -1 Dynamically enhanced feature matrix for all regions in the time period, For the t The spatiotemporal causal features extracted by time periods, For the t -1 period enhances the spatiotemporal causal characteristics, and are projection matrices, , , , is a learnable parameter, For the t Time period and location information, For the t -1 time period time location information;

[0023] (3) Based on and Generate from Time period to Causal transfer matrix of time period .

[0024] Preferably, the causal transfer process in the priori causal transfer module is as follows:

[0025] (1) The causal transmission process is established as follows:

[0026]

[0027] in, Indicates t The final output features of the period causal transfer module, represents the ReLU activation function, l Indicates l Layer causal transmission, L represents the total number of causal transmission layers, Indicates t The initial value of the period causal transfer process, Indicates t Time period The output features of the layer, Indicates The output features of the layer, Indicates Layer network parameter weights, Indicates Layer network bias;

[0028] (2) Incorporate the distance graph and interaction graph as prior knowledge into the causal transfer process to generate the first t Time segment enhances spatiotemporal causal characteristics :

[0029]

[0030] in, represents the causal transfer process of the distance graph, Represents the causal transfer process of the interaction graph, represents the adjacency matrix of the inter-region distance graph, Indicates t An interaction matrix of the time zones where the time slots are located.

[0031] Preferably, the The way to obtain is as follows:

[0032] (1) Divide a week into K time zones, and count the interactive traffic between regions in each time zone , Indicates time zone k Neiyudi i Departure and arrival at the area j Interaction traffic in the class area, k =1,2,3… K ;

[0033] (2) Yes After normalization, the normalized interaction traffic between regions in each time zone is obtained:

[0034]

[0035] (3) Constructing an interaction matrix for each time zone , No. t The interaction matrix of the time zone in which the time period is located is .

[0036] Preferably, the fusion prediction module takes the output of multiple series-connected enhanced spatiotemporal causal feature extraction modules as input, and generates the urban travel demand data prediction result using stacked fully connected layers and Relu activation function:

[0037]

[0038] in, For the t +1 period forecasted urban travel demand data, , , and are all learnable parameters. Enhanced spatiotemporal causal features extracted for the last period.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] The present invention proposes a method for predicting urban travel demand based on prior causal transfer. The prediction method uses prior knowledge to construct a causal graph of travel demand in various areas of the city to introduce inductive bias, and introduces it into the spatial causal transfer process across time slices. Combined with a learnable causal transfer matrix, the spatiotemporal causal representation is enhanced, thereby improving the accuracy and robustness of urban travel demand prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. The features and advantages of the present invention can be more clearly understood by referring to the drawings. The drawings are schematic and should not be understood as limiting the present invention in any way. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 It is a flow chart of the urban travel demand prediction method based on prior causal transmission proposed by the present invention.

[0043] Figure 2 It is a schematic diagram of time domain causal intervention in the present invention. DETAILED DESCRIPTION

[0044] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0046] The present invention provides a method for predicting urban travel demand based on prior causal transmission, such as Figure 1 As shown, first, according to the prediction requirements, the historical urban travel demand data in the area to be predicted is collected, and a training set is obtained after standardized processing; then a deep learning model based on prior causal transfer is constructed, and the deep learning model includes a causal debiasing module, a causal graph learner, a prior causal transfer module and a fusion prediction module; the training set is used to complete the training of the deep learning model; finally, the trained deep learning model is applied to the prediction of future urban travel demand data.

[0047] Example 1

[0048] For ease of understanding, the present invention is described in more detail by taking the taxi demand in the urban travel demand data as an example.

[0049] Step 1: Collect historical taxi order information through the taxi online booking platform. The order information includes the following eight data: order number, departure longitude, departure latitude, arrival longitude, arrival latitude, departure time, arrival time, and travel time. Then standardize the order information and divide the regions to be predicted into N For each sub-region, we remove the order data with travel time less than 0, and count the taxi demand in each time period in each sub-region as the standard data set.

[0050] No. t The taxi demand in all sub-regions during the period is recorded as , a total of Q Historical taxi demand in all sub-regions during the time period as a training set.

[0051] Step 2: Build a deep learning model based on prior causal transfer.

[0052] The deep learning model based on prior causal transfer is composed of multiple enhanced spatiotemporal causal feature extraction modules connected in series and a fusion prediction module, wherein each enhanced spatiotemporal causal feature extraction module includes a causal debiasing module, a causal graph learner and a prior causal transfer module.

[0053] The causal debiasing module stratifies the input historical taxi demand into clusters with different confounding factors through a potential confounding estimator, extracts spatiotemporal features from each cluster through a private encoder, and constructs unbiased spatiotemporal causal features using backdoor adjustment.

[0054] The causal graph learner constructs a time embedding matrix based on time position information, then uses the time embedding matrix to construct a dynamic feature matrix for the region to be predicted, and finally constructs a causal transfer matrix based on the dynamic feature matrix.

[0055] The prior causal transfer module, based on the causal transfer matrix, fuses the distance graph and the interaction graph as prior knowledge, uses a graph convolutional network to construct the causal transfer process and generate enhanced spatiotemporal causal features.

[0056] Finally, the enhanced spatiotemporal causal features are input into the fusion prediction module, mapped to the taxi demand space, and the predicted value of taxi demand in the future period is output.

[0057] (1) Causal debiasing module

[0058] The schematic diagram of time domain causal intervention in the causal debiasing module is as follows Figure 2 As shown, there are six variables in total.t -1 period (i.e. the previous period) taxi demand , No. t The demand for taxis during the current period , No. t +1 period (i.e. the next period) predicts the taxi demand, t The time period (i.e. the current time period) enhances the spatiotemporal causal characteristics , No. t -1 period (i.e. the previous period) enhances the spatiotemporal causal characteristics , confounding factors C .

[0059] Through modeling Make predictions, where To integrate the predictor, the enhanced spatiotemporal causal features of the current period are mapped to the taxi demand space, thereby outputting the predicted value of taxi demand for the next period, which is generally achieved through a fully connected layer. , confounding factors C Through the backdoor path exist and If the confounding factors C According to the value of the stratification is broken down into:

[0060]

[0061] in, Indicates t Taxi demand during the period, confounding factors C Obeying discrete uniform distribution, p Confounding factors C Total p Class, coming soon N Sub-regions are divided into p Class area, , Indicates the input i Category Area t -1 taxi demand during a period The corresponding deep learning model, where i =1,2,… p , Indicates i Category Area t The time period enhances the spatiotemporal causal characteristics, Indicates i Category Area t -1 period to enhance the spatiotemporal causal characteristics, Indicates i Confounding factors, express The co-occurrence probability of .

[0062] Due to the imbalance of the dataset, Confounding factors C cannot be combined with equal probability, resulting in the extracted It tends to express frequently occurring general timing dependencies. In order to eliminate false correlations, a backdoor adjustment formula is used. replace , thus cutting off Backdoor path to eliminate confounding factors C The false correlation It can be decomposed by the following formula:

[0063]

[0064] In order to Confounding factors C Fair combination, setting up confounding factors C It obeys the discrete uniform distribution, that is, , is the total number of selectable values ​​for the confounding factors.

[0065] Since the confounding factors are not visible in the data set, the present invention designs a potential confounding estimator part to stratify the historical taxi demand according to the confounding factors, and assumes a learnable stratification matrix ,in N is the total number of sub-regions. The present invention performs regularization on the hierarchical matrix based on the Gumbel Softmax method, so that only one element in each row of the hierarchical matrix is ​​1, and the rest of the elements are 0. Regularized hierarchical matrix The historical taxi demand in each region is divided into clusters with different confounding factors. After that, the time domain causal features of each region are extracted through parallel time domain encoders. t Confounding factors within time period The value is i Spatial-temporal causal characteristics of region-like , the expression is as follows:

[0066]

[0067] in, Indicates i Category Area t -1 period extracted enhanced spatiotemporal causal features, represents the sigmoid activation function, , , , , and are all learnable weights, Indicates i Category Area t The demand for taxis during the time period, represents the tanh activation function, To initialize the zero matrix, represents the temporal position embedding layer, Indicates t Time period and location information.

[0068] Finally, the spatiotemporal causal features extracted from various regions are combined to obtain the t Time-space causal characteristics :

[0069]

[0070] in, t Indicates t Time period, Indicates i Category Area t The spatiotemporal causal features extracted by time periods, i =1,2,… p , represents the matrix concatenation operation on the node dimension, BatchNorm Represents the batch normalization operation, which prevents gradient disappearance and speeds up the convergence of deep learning models.

[0071] (2) Causal Graph Learner

[0072] How to model the dynamic nature of causal relationships and how to add prior knowledge to guide the generation of causal graphs are the keys to recovering causal relationships from the demand for urban taxis. In order to model dynamic causal relationships, the present invention first generates a dynamic feature matrix for all regions:

[0073]

[0074] in For the t The dynamic feature matrix of all regions in the time period, For the t -1 Dynamically enhanced feature matrix for all regions in the time period, For the t The spatiotemporal causal features extracted by time periods, For the t -1 period enhances the spatiotemporal causal characteristics, and are all projection matrices;

[0075] In order to enhance the interpretability of the model and reduce the difficulty of training, the time position information matrix is ​​used Guide the projection matrix generation process:

[0076]

[0077] in , , , is a learnable parameter, For the t Time period and location information, For the t -1 Time period location information.

[0078] First For example, the Class area to Causal Effects of Region-like As follows:

[0079]

[0080] in, express No. i OK, express No. j OK.

[0081] Use the softmax function to normalize the causal effect:

[0082]

[0083] The final generated from arrive Causal transfer matrix of time period .

[0084] (3) Priori Causal Transfer Module

[0085] In order to transfer the information in the spatiotemporal causal features of the previous period to the current period, the present invention designs a priori causal transfer module across time slices.

[0086] First Taking the class region as an example, the causal transfer process is defined as:

[0087]

[0088] For all regions, the causal transfer process can be generalized as:

[0089]

[0090] in, Indicates t The final output features of the cross-time slice causal transfer module, represents the ReLU activation function, l Indicates l Layer causal transmission, L represents the total number of causal transmission layers, Indicates The output features of the layer, , Indicates Layer network parameter weights, Indicates Layer network bias.

[0091] Considering that regions with close geographical distances are more likely to have causal relationships, the present invention integrates the distance map as prior knowledge into the causal transmission process, denoted as , represents the causal transfer process of the distance graph, An adjacency matrix representing the distance graph between regions.

[0092] In addition, the present invention also integrates the interaction graph as prior knowledge into the causal transmission process, which is recorded as , Represents the causal transfer process of the interaction graph, Indicates t An interaction matrix of the time zones where the time slots are located.

[0093] Among them Obtained through:

[0094] a) Divide a week into K time zones, and count the interactive traffic between regions in each time zone , Indicates time zone k Neiyudi i Departure and arrival at the area j Interaction traffic in the class area, k =1,2,3… K ;

[0095] b) Yes After normalization, the normalized interaction traffic between regions in each time zone is obtained:

[0096]

[0097] c) Build an interaction matrix for each time zone , No. t The interaction matrix of the time zone in which the time period is located is .

[0098] Finally, generate the t Time segment enhances spatiotemporal causal characteristics :

[0099] .

[0100] (4) Fusion prediction module.

[0101] The present invention maps the extracted enhanced spatiotemporal causal features to the taxi demand space to generate the taxi demand forecast result for the next period. Specifically, the taxi demand forecast result is generated using stacked fully connected layers and Relu activation functions, as shown in the following formula:

[0102]

[0103] in, For the t +1 period predicted taxi demand, , , and are all learnable parameters. For the Q Enhanced spatiotemporal causal features extracted within the period (i.e., the last period).

[0104] Step 3: Use the training set to train the deep learning model.

[0105] The present invention uses L2-norm as the prediction loss , and designed a multi-objective loss function to train the deep learning model, as shown below:

[0106]

[0107] in, represents the total loss function, is the sparseness loss, , which aims to make the causal transfer matrix sparse to prevent overfitting as a hyperparameter controlling the balance.

[0108] Step 4: Input the historical taxi demand of the area to be predicted into the trained deep learning model and output the taxi demand for the next period in each area.

[0109] The present invention uses the taxi order data set of Xicheng District, Beijing and the taxi order data set of Tongzhou District, Beijing to train the deep learning model. The Xicheng District, Beijing data set used in the present invention intercepts the order data from January 1, 2022 to December 31, 2023, which is divided into 235 sub-areas. The Tongzhou District, Beijing data set intercepts the order data from January 1, 2022 to December 31, 2023, which is divided into 106 sub-areas.

[0110] (1) Each order data includes the following eight types of information: order number, departure longitude, departure latitude, arrival longitude and arrival latitude, departure time, arrival time, and travel time. The data set is cleaned and converted into a regional taxi demand data set in the manner of step 1, and the time period is divided into 30 minutes. The data set is divided into a training set (80%), a validation set (10%), and a test set (10%) in the time dimension. The present invention regards the demand forecasting task as a traffic flow forecasting task, and uses the historical 3-hour regional demand data to predict the regional demand in the next 30 minutes.

[0111] (2) Training the deep learning model of the present invention. The data is Z-score standardized, and all parameters in the deep learning model are randomly initialized.

[0112] Training was performed on an Intel(R) Core(TM) i9-10980XE CPU and a GeForce RTX3090 GPU on a Linux operating system, with the batch parameter set to 64 and the initial learning rate set to 0.001.

[0113] The model was trained for 200 cycles on the complete dataset using the Adam optimization algorithm and the exponential decay dynamic learning rate strategy. The model was validated using the loss function in each training cycle, and then the optimal model was saved based on the loss function value. The Early Stopping strategy was used during training, and training was terminated early when the loss function value did not decrease for 50 consecutive cycles.

[0114] (3) The method of the present invention is compared with the prior art, and the regional taxi demand forecast is performed on the same data set. The comparison results are shown in Tables 1 and 2. The mean absolute error (MAE), root mean square error (RMSE) and mean percentage absolute error (MAPE) are used to evaluate the forecast results. The lower the error, the better the forecast effect. The present invention compares eight prior art solutions.

[0115] The first one is HGCN, which simultaneously captures spatial correlations of different scales through hierarchical GCN.

[0116] The second one is GraphWaveNet, which mines the potential space through an adaptive graph structure learning process and uses it for traffic flow prediction, further improving the prediction performance.

[0117] The third one is AGCRN, which models temporal dependencies by assigning independent parameter spaces to different regions, but lacks the process of adjusting the weights of features extracted from various regions from a causal perspective.

[0118] The fourth one is DGCRN, which uses a dynamic graph generation module to model the dynamics of the traffic system, but still models the temporal dependencies of different areas in a shared space.

[0119] The fifth one is DMSTGCN, which assigns a separate graph structure to each time slice within a day to model the dynamics of the traffic system and utilizes auxiliary information to improve the prediction accuracy.

[0120] The sixth is RGSL, which proposes a regularized graph learning method that integrates explicit and implicit graphs to simultaneously model spatial dependencies.

[0121] The seventh is CCHMM, which mines the potential physical variables in spatiotemporal data through variational inference and reconstructs their generation process.

[0122] The eighth is STNSCM, which performs causal debiasing through front-gate adjustment and learns causal graphs without the guidance of prior knowledge.

[0123] Table 1 Comparison of taxi demand forecast results in Xicheng District, Beijing

[0124]

[0125] Table 2 Comparison of taxi demand forecast results in Tongzhou District, Beijing

[0126]

[0127] It can be seen from the experimental results that the prediction effect of the present invention is better than that of the existing model.

[0128] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0129] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0130] In the present invention, the terms "first", "second", "third", "fourth" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. The term "plurality" refers to two or more, unless otherwise clearly defined.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting urban travel demand based on prior causal transfer, characterized in that: The following steps are involved: S1: According to the prediction requirements, the historical urban travel demand data in the area to be predicted is collected and the training set is obtained after standardized processing; S2: construct a deep learning model based on prior causal transfer, wherein the deep learning model includes a causal debiasing module, a causal graph learner, a prior causal transfer module and a fusion prediction module; In the causal graph learner, a time embedding matrix is ​​constructed based on the time position information, and then a dynamic feature matrix is ​​constructed for the region to be predicted using the time embedding matrix, and finally a causal transfer matrix is ​​constructed based on the dynamic feature matrix; In the prior causal transfer module, based on the causal transfer matrix, the distance graph and the interaction graph are fused as prior knowledge, and the graph convolutional network is used to construct the causal transfer process and generate enhanced spatiotemporal causal features; S3: Use the training set in S1 to complete the training of the deep learning model; S4: Apply the trained deep learning model to the expected urban travel demand data prediction; The standardization process in S1 is: divide the area to be predicted into N For each sub-region, the urban travel demand data for each sub-region at each time period is collected; The causal debiasing module divides all N The sub-regions were divided into p The parallel encoders are used to extract the spatiotemporal causal features of each type of region in different time periods, and then the spatiotemporal causal features of different time periods are obtained by merging them. The process of constructing the causal transfer matrix by the causal graph learner includes: (1) Construct a time embedding matrix based on the time position information and construct a projection matrix; (2) Construct the dynamic feature matrix of all regions: in For the t The dynamic feature matrix of all regions in the time period, For the t -1 Dynamically enhanced feature matrix for all regions in the time period, For the t The spatiotemporal causal features extracted by time periods, For the t -1 period enhances the spatiotemporal causal characteristics, and are projection matrices, , , , is a learnable parameter, For the t Time period and location information, For the t -1 time period time location information; (3) Based on and Generate from Time period to Causal transfer matrix of time period ; The causal transfer process in the prior causal transfer module is as follows: (1) The causal transmission process is established as follows: in, Indicates t The final output features of the period causal transfer module, represents the ReLU activation function, l Indicates l Layer causal transmission, L represents the total number of causal transmission layers, Indicates t The initial value of the period causal transfer process, Indicates t Time period The output features of the layer, Indicates The output features of the layer, Indicates Layer network parameter weights, Indicates Layer network bias; (2) Incorporate the distance graph and interaction graph as prior knowledge into the causal transfer process to generate the first t Time segment enhances spatiotemporal causal characteristics : in, represents the causal transfer process of the distance graph, Represents the causal transfer process of the interaction graph, represents the adjacency matrix of the inter-region distance graph, Indicates t The interaction matrix of the time zone where the time period is located; Said The way to obtain is as follows: (1) Divide a week into K time zones, and count the interactive traffic between regions in each time zone , Indicates time zone k Neiyudi i Departure and arrival at the area j Interaction traffic in the class area, k =1,2,3… K ; (2) Yes After normalization, the normalized interaction traffic between regions in each time zone is obtained: (3) Constructing an interaction matrix for each time zone , No. t The interaction matrix of the time zone in which the time period is located is ; The fusion prediction module takes the output of multiple series-connected enhanced spatiotemporal causal feature extraction modules as input, and uses stacked fully connected layers and Relu activation functions to generate urban travel demand data prediction results: in, For the t +1 period forecasted urban travel demand data, , , and are all learnable parameters. Enhanced spatiotemporal causal features extracted for the last period.

2. The urban travel demand forecasting method based on a priori causal transfer according to claim 1 is characterized in that: The urban travel demand includes: taxi demand, bus load, shared bicycle demand or subway passenger flow.