Multi-task learning-based multi-mode trip flow collaborative prediction method, system and device

Abstract

The invention discloses a multi-mode travel flow collaborative prediction method, system and device based on multi-task learning. The method comprises the following steps: acquiring related data of multiple travel modes in a to-be-tested area; performing preprocessing and time-space alignment on related data of various travel modes, extracting flow similar characteristics, built environment characteristics and time sequence characteristics of OD flows based on an OD flow matrix, and extracting high-dimensional characteristics of built environment and mobile flows based on the built environment characteristics and the time sequence characteristics; constructing a spatial adjacency graph based on a spatial adjacency relation, constructing a traffic feature graph based on traffic similar features, constructing an environment feature graph based on extracted high-dimensional features, and further constructing a view of each travel mode; constructing a traffic prediction pre-training model, and training the traffic prediction pre-training model to obtain a traffic prediction model; and inputting the OD flow to be measured into the flow prediction model for prediction to obtain a prediction result at the next moment. The multi-traffic OD flow in the subsequent time period is predicted based on the historical OD flow and the built environment data, and the result precision is high.

Description

Technical Field

[0001] The present invention relates to the field of big data technology, and in particular to a method, system and device for collaborative prediction of multi-mode travel flow based on multi-task learning. Background Art

[0002] With the diversification of urban travel modes and the increasing complexity of travel needs, urban residents have more travel options. Compared with a single travel mode, residents are more inclined to adopt a travel mode that combines multiple travel modes, such as taxi and subway intermodal transport, bus and shared bicycle intermodal transport, etc. This multimodal transportation mode brings complex interactions. There is a close correlation between various travel modes. Travel prediction based on a single travel mode alone cannot fully reveal the dynamic changes of the urban transportation system and the mutual influence between different travel modes. At present, most research methods focus on the prediction of single transportation travel demand, and rarely consider the interaction between different travel modes. Therefore, it is particularly necessary to design a framework that can collaboratively predict residents' multimodal transportation travel flow, which will help accurately predict residents' multimodal transportation travel demand and more truly reflect the diversity of residents' travel behavior.

[0003] At present, the prediction methods for urban residents' travel demand are mainly based on the regional scale, focusing on analyzing the inflow and outflow of population in the region, but unable to predict the direction of residents' mobile flow. In contrast, OD (Origin-Destination) flow prediction focuses on the starting and ending points of residents' travel, and can achieve the prediction of directed flow, thereby more accurately capturing the dynamic changes of residents' travel, carefully describing residents' travel behavior, identifying travel hotspots and high-frequency travel routes in the city, so as to meet the precise needs of urban and traffic management.

[0004] In recent years, deep learning models have been widely used in the field of travel demand forecasting because of their ability to capture nonlinear spatiotemporal dependency characteristics. Current forecasting models mainly capture the spatiotemporal dependency characteristics of a single travel mode based on historical data, while the spatiotemporal dependency between different travel modes and their synergistic effects are relatively less considered. In order to achieve accurate prediction of multimodal traffic, it is necessary to build modules for intra-modal and inter-modal learning respectively, capturing the spatiotemporal dependency characteristics of each travel mode and the synergistic effects between different travel modes. At the same time, OD (origin-destination) forecasting still faces many challenges. On the one hand, the data dimension of OD forecasting far exceeds that of regional forecasting, and it shows significant sparsity in time and space. On the other hand, deep learning models usually involve a large number of parameters and require careful adjustment and evaluation. Therefore, how to solve the problem of data sparsity and how to adjust parameters to improve model accuracy have become the core issues of OD demand forecasting.

[0005] In the current methods, the prediction of urban residents' travel demand mainly relies on historical travel data, and deep learning is used as the main research method to seek more accurate models and methods at the regional and OD (origin-destination) scales. Despite this, there are still several problems with the current methods: first, the current methods mainly focus on the prediction of undirected flow at the regional scale, while there are fewer predictions of directed flow at the OD scale; second, most of the current methods only consider the spatiotemporal dependence of a single travel mode itself, while there are few methods for multi-modal travel OD prediction modeling. This method mainly faces two major technical difficulties: data sparsity and model complexity. Solving these two problems has become a difficulty in the current development of technology. In addition, the prediction of urban residents' travel demand has not yet been deeply associated with geographical environmental characteristics. Although existing methods use geographical environmental characteristics to analyze residents' travel patterns and behaviors to a certain extent, there is still limited research on deeply integrating geographical environmental characteristics into predictive modeling. In these studies, geographical environmental characteristics are usually input into the model together with historical data, but are not deeply embedded in the model to capture the complex dependency between urban residents' travel demand and the geographical environment. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a method, system and device for collaborative prediction of multi-mode travel flow based on multi-task learning.

[0007] In order to solve the above technical problems, the present invention is solved by the following technical solutions: A multi-mode travel flow collaborative prediction method based on multi-task learning includes the following steps: Obtain relevant data of various travel modes in the area to be tested, wherein the relevant data include historical start and end point data, point of interest data and road network data. The historical start and end point data include various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data are city-related infrastructure data, service data and geographic information data, and the road network data are the road network density within each spatial unit; The area to be tested is divided into unit grids, the relevant data of various travel modes are preprocessed and aligned in time and space, and the OD flow matrix is ​​extracted. Based on the OD flow matrix, the flow similarity characteristics of the OD flow, the built environment characteristics and the time series characteristics are extracted; Based on the spatial proximity relationship of OD flow, a spatial adjacency graph is constructed, based on the similarity characteristics of traffic, a traffic feature graph is constructed, and based on the built environment characteristics and time series characteristics, high-dimensional features of the built environment and mobile flow are extracted and a built environment feature graph is constructed; Construct each travel mode view through spatial adjacency graph, flow characteristic graph and built environment characteristic graph; Constructing a traffic prediction pre-training model, the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The OD flow to be measured is input into the traffic prediction model for prediction, and the prediction result for the next moment under the current application scenario is obtained.

[0008] As an implementable method, the area to be tested is divided into unit grids, the relevant data of multiple travel modes are processed in time and space, and the OD flow matrix is ​​extracted, including the following steps: The area to be measured is represented as a rectangle and divided into H×W spatial units along longitude and latitude. Each spatial unit has the same size and is encoded to N in sequence. The starting point and end point of each travel mode are located in N spatial units, and there are Kind of OD flow; Select M time periods in the historical time period, extract the flow of each OD flow in each time period in turn, that is, extract the number of OD pairs that fall into the first spatial unit and the end point falls into the second spatial unit, and align multiple travel modes in time and space, then the data matrix dimension is expressed as That is the OD flow matrix of M time periods.

[0009] As an implementable method, constructing a spatial adjacency graph based on spatial proximity relationship and constructing a traffic feature graph based on traffic similarity features include the following steps: The OD flows are regarded as nodes, and the adjacent relationship between two OD flows is used as the edge of the graph structure to construct a spatial adjacency graph, which is expressed as: ; The flow characteristic diagram is expressed as: , It means that each OD flow is connected with several OD flows that rank top in the OD flow ranking, which is used to capture the mobility characteristics of the OD flow; Among them, according to the proximity relationship of the OD flow, if the two OD flows are in a neighboring relationship, the weight is 1, otherwise the weight is 0. The proximity relationship of the OD flow is determined according to the adjacency relationship between the starting point and the end point. For a given OD flow, there are three types of spatial neighbors, namely, OD flows that are adjacent to the starting point and end point of the current OD flow respectively, OD flows that are adjacent to the starting point of the OD flow and have the same end point, and OD flows that have the same starting point as the OD flow and are adjacent to the end point.

[0010] As an implementable method, the method of extracting high-dimensional features of the built environment and the mobile flow based on the built environment features and the time series features and constructing a built environment feature map includes the following steps: The point of interest data and road network data are processed, and the point of interest data and road network data of various travel modes in each spatial unit are counted, so as to obtain the built environment characteristics of each spatial unit; The built environment features related to the starting and ending points of the OD flow are extracted, and the historical time period of the OD flow is evenly divided into several intervals. The OD flow of each travel mode in each time period is summarized to obtain the time series features; Based on the graph convolutional network, the built environment features and time series features are processed respectively. The graph convolutional network sets multiple convolutional layers and an output layer. The dimension of the convolutional layer is set to 64 or 128, and the dimension of the output layer is set to 1. The activation function uses Relu, and the model is optimized using the MSE loss function and Adam optimizer to obtain a high-dimensional feature matrix of the built environment mobile flow. , and the dimension is expressed as ; Based on the high-dimensional feature matrix, the inner product of the high-dimensional feature matrix is ​​calculated by the relationship matrix calculation formula. The similarity of the high-dimensional features is expressed by the inner product. The relationship matrix calculation formula is: , represents the inner product of a high-dimensional feature matrix and has dimensions , Represents a high-dimensional feature matrix; Based on the high-dimensional feature matrix and similarity, the built environment feature map is constructed and expressed as: ; in, express Kind of OD flow, The dimension is The characteristic matrix of is the inner product, Represents the feature matrix of a node.

[0011] As an implementable method, each travel mode view is represented as follows:

[0012] in, represents the nth travel mode view, They represent the spatial adjacency graph, the traffic characteristic graph, and the built environment characteristic graph respectively.

[0013] As an implementable method, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network, including: The spatiotemporal dependency feature extraction module includes a multi-layer temporal graph convolutional network, wherein the temporal graph convolutional network includes a temporal convolutional network and a graph convolutional network, wherein the temporal convolutional network captures temporal dependency features, and the graph convolutional network captures spatial dependency features; The temporal convolutional network includes a causal dilated convolutional structure and a gated recurrent unit. The causal dilated convolutional structure is used to learn the dynamic changes of the OD flow in each travel mode view to capture the time-dependent features, and the gated recurrent unit is used to capture the time-dependent features again. Constructing filters in the Fourier domain of the graph convolutional network to capture spatial dependency features between nodes; The causal dilated convolution structure is a one-dimensional convolution, expressed as: ; The gated recursive unit is expressed as: ; The graph convolutional network is represented as: ; in, represents the features after dilated causal convolution, represents the time-dependent feature, Indicates size The core, represents the dilation factor that controls the jump distance, represents the step length, denote convolution filters, represents the sigmoid function, They represent the coefficient of deviation, represents the time-dependent features captured again, represents the convolution kernel, express The symmetric Laplacian matrix of represents the identity matrix, express The diagonal matrix of Represents spatially dependent features.

[0014] As an implementable method, the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism, including the following steps: Obtain the spatiotemporal dependency features of the same graph in each travel mode, concatenate the spatiotemporal dependency features in their respective dimensions, and perform dimension transformation to form a merged tensor; Through the attention mechanism, the merged tensor is mapped into query, key and value vectors through the convolutional layer to capture the interactive dependencies between the same type of graphs of multiple travel modes and obtain interactive dependency features; The attention mechanism weights are expressed as follows:

[0015] in, represents the interactive dependency feature, Represents the dimension of the feature, They represent query vector, key vector and value vector respectively.

[0016] As an implementable method, the fusion feature is expressed as follows:

[0017] in, represents the learnable parameter matrix, represents the Hadamard product, They represent interactive dependency features respectively.

[0018] As an implementable method, the loss function is obtained by: Constructing a first loss function and a second loss function respectively, and obtaining a total loss function based on the first loss function and the second loss function; The first loss function is expressed as follows:

[0019] in, represents the unobstructed OD flow, represents the predicted value, Indicates the actual value, mask indicates that the matrix file value is 0 or 1, indicating whether the corresponding OD flow is masked; The second loss function is expressed as follows:

[0020] The loss function is expressed as follows: ; in, Represent the importance weights corresponding to the first loss function and the second loss function respectively.

[0021] A multi-task learning-based multi-mode travel flow collaborative prediction system, including a data acquisition module, a processing and extraction module, a feature extraction module, a view construction module, a model training module and a prediction and reasoning module; The data acquisition module is used to acquire relevant data of various travel modes in the area to be tested, wherein the relevant data includes historical start and end point data, point of interest data and road network data, the historical start and end point data includes various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data is city-related infrastructure data, service data and geographic information data, and the road network data is the road network density within each spatial unit; The processing and extraction module is used to divide the area to be tested into unit grids, pre-process and align the relevant data of multiple travel modes in time and space, extract the OD flow matrix, and based on the OD flow matrix, extract the flow similarity characteristics of the OD flow and the built environment characteristics and time series characteristics; The feature extraction module constructs a spatial adjacency graph based on the spatial proximity relationship of the OD flow, constructs a flow feature graph based on the flow similarity features, extracts high-dimensional features of the built environment and the mobile flow based on the built environment features and the time series features, and constructs a built environment feature graph; The view construction module constructs each travel mode view through a spatial adjacency graph, a traffic characteristic graph, and a built environment characteristic graph; The model training module is used to construct a traffic prediction pre-training model, and the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module. The spatiotemporal dependency feature extraction module receives each travel mode view and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The prediction and reasoning module is used to input the OD flow to be measured into the flow prediction model for prediction, and obtain the prediction result of the next moment under the application scenario at the current moment.

[0022] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following method is implemented: Obtain relevant data of various travel modes in the area to be tested, wherein the relevant data include historical start and end point data, point of interest data and road network data. The historical start and end point data include various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data are city-related infrastructure data, service data and geographic information data, and the road network data are the road network density within each spatial unit; The area to be tested is divided into unit grids, the relevant data of various travel modes are preprocessed and aligned in time and space, and the OD flow matrix is ​​extracted. Based on the OD flow matrix, the flow similarity characteristics of the OD flow, the built environment characteristics and the time series characteristics are extracted; Based on the spatial proximity relationship of OD flow, a spatial adjacency graph is constructed, based on the similarity characteristics of traffic, a traffic feature graph is constructed, and based on the built environment characteristics and time series characteristics, high-dimensional features of the built environment and mobile flow are extracted and a built environment feature graph is constructed; Construct each travel mode view through spatial adjacency graph, flow characteristic graph and built environment characteristic graph; Constructing a traffic prediction pre-training model, the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The OD flow to be measured is input into the traffic prediction model for prediction, and the prediction result for the next moment under the current application scenario is obtained.

[0023] A multi-mode travel flow collaborative prediction device based on multi-task learning includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, the following method is implemented: Obtain relevant data of various travel modes in the area to be tested, wherein the relevant data include historical start and end point data, point of interest data and road network data. The historical start and end point data include various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data are city-related infrastructure data, service data and geographic information data, and the road network data are the road network density within each spatial unit; The area to be tested is divided into unit grids, the relevant data of various travel modes are preprocessed and aligned in time and space, and the OD flow matrix is ​​extracted. Based on the OD flow matrix, the flow similarity characteristics of the OD flow, the built environment characteristics and the time series characteristics are extracted; Based on the spatial proximity relationship of OD flow, a spatial adjacency graph is constructed, based on the similarity characteristics of traffic, a traffic feature graph is constructed, and based on the built environment characteristics and time series characteristics, high-dimensional features of the built environment and mobile flow are extracted and a built environment feature graph is constructed; Construct each travel mode view through spatial adjacency graph, flow characteristic graph and built environment characteristic graph; Constructing a traffic prediction pre-training model, the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The OD flow to be measured is input into the traffic prediction model for prediction, and the prediction result for the next moment under the current application scenario is obtained.

[0024] The present invention has significant technical effects due to the adoption of the above technical solution: This paper integrates multi-source geographic data and deep learning methods, and considers the interactive dependence and synergistic influence of built environment characteristics and multi-modal transportation, to construct a high-precision urban resident transportation travel demand prediction model at the OD scale to support scientific decision-making in multi-modal transportation collaborative management and urban planning; Specifically, this paper focuses on multi-modal traffic prediction at the OD scale. Based on the representation learning method, the temporal dynamic characteristics of the built environment characteristics and the crowd movement flow are extracted, and multi-view representations are constructed with the spatial correlation and flow characteristics of the movement flow. Based on the idea of ​​multi-task learning, the temporal graph convolutional network (TGCN) and the self-attention mechanism (Self-Attention) are used to capture the spatiotemporal dependency characteristics and the interactive dependency between multi-modal traffic, and an auxiliary loss function is designed to enhance the prediction. The present invention uses a masked loss function to reduce the impact of data sparsity on prediction. Through the above framework, the present invention can achieve collaborative prediction of multiple traffic flows; based on historical OD flows and built environment data, multiple traffic OD flows in subsequent periods are predicted. Compared with existing methods, the model is relatively simple, the computing power requirement is smaller, and the result accuracy is high. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] 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 or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0026] Figure 1 It is a schematic flow diagram of the method of the present invention; Figure 2 It is a schematic diagram of the overall structure of the system of the present invention; Figure 3 is a schematic diagram of a method for constructing a built environment characteristic map of the present invention; Figure 4 It is a schematic diagram of the traffic prediction pre-training model of the present invention. DETAILED DESCRIPTION

[0027] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0028] Embodiment 1: A multi-mode travel flow collaborative prediction method based on multi-task learning, such as Figure 1 As shown, the following steps are included: S100, obtaining relevant data of various travel modes in the area to be tested, wherein the relevant data includes historical start and end point data, point of interest data and road network data, the historical start and end point data includes various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data is city-related infrastructure data, service data and geographic information data, and the road network data is the road network density within each spatial unit; S200, dividing the area to be measured into unit grids, preprocessing and aligning relevant data of multiple travel modes in time and space, extracting the OD flow matrix, and based on the OD flow matrix, extracting the flow similarity characteristics of the OD flow and the built environment characteristics and time series characteristics; S300, constructing a spatial adjacency graph based on the spatial proximity relationship of the OD flow, constructing a traffic feature graph based on traffic similarity features, extracting high-dimensional features of the built environment and mobile flow based on built environment features and time series features and constructing a built environment feature graph; S400, construct each travel mode view through the spatial adjacency graph, the flow characteristic graph and the built environment characteristic graph; S500, constructing a traffic prediction pre-training model, the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; S600: Input the OD flow to be measured into the traffic prediction model for prediction, and obtain the prediction result for the next moment under the application scenario at the current moment.

[0029] In order to capture the interactive dependencies between different travel modes, the idea of ​​multi-view learning is adopted to decompose the complex interactive pattern into multiple independent simple patterns, capture the spatiotemporal dependencies of each view respectively, and then fuse the multi-view features. This method can reduce the burden of capturing complex nonlinear dependencies, reduce the complexity of the model, and make the generated model more interpretable.

[0030] In a specific embodiment, the method of dividing the area to be tested into unit grids, performing spatiotemporal alignment processing on the relevant data of multiple travel modes, and extracting the OD flow matrix includes the following steps: The area to be measured is represented as a rectangle and divided into H×W spatial units along longitude and latitude. Each spatial unit has the same size and is encoded to N in sequence. The starting point and end point of each travel mode are located in N spatial units, and there are Kind of OD flow; Select M time periods in the historical time period, extract the flow of each OD flow in each time period in turn, that is, extract the number of OD pairs that fall into the first spatial unit and the end point falls into the second spatial unit, and align multiple travel modes in time and space, then the data matrix dimension is expressed as That is the OD flow matrix of M time periods.

[0031] If we regard OD flows as nodes and the adjacent relationships between two OD flows as edges of the graph structure, we can construct the graph structure as ,in, represents a node set, Represents the dimension, The dimension is The characteristic matrix of is the characteristics of the edge, Represents the feature matrix of a node, expressed as , represents the historical travel demand of the two travel modes for each node in the graph structure. The travel prediction problem is to build and train a reasonable model given a graph or multiple graphs G and their feature matrix X, which can predict the characteristics of the next time period under the two travel modes.

[0032] The method of constructing a spatial adjacency graph based on the spatial proximity relationship and constructing a traffic feature graph based on traffic similarity features comprises the following steps: The OD flows are regarded as nodes, and the adjacent relationship between two OD flows is used as the edge of the graph structure to construct a spatial adjacency graph, which is expressed as: ; The flow characteristic diagram is expressed as: , It means that each OD flow is connected with several OD flows that rank top in the OD flow ranking, which is used to capture the mobility characteristics of the OD flow; Among them, according to the proximity relationship of the OD flow, if the two OD flows are in a neighboring relationship, the weight is 1, otherwise the weight is 0. The proximity relationship of the OD flow is determined according to the adjacency relationship between the starting point and the end point. For a given OD flow, there are three types of spatial neighbors, namely, OD flows that are adjacent to the starting point and end point of the current OD flow respectively, OD flows that are adjacent to the starting point of the OD flow and have the same end point, and OD flows that have the same starting point as the OD flow and are adjacent to the end point.

[0033] High-dimensional features are a necessary condition for making predictions more accurate. Therefore, the built environment feature map based on high-dimensional features is constructed as follows: Figure 3 As shown, the following steps are included: Process the POI data and road network data, count the POI data road network density in each spatial unit, and then obtain the built environment characteristics of each spatial unit; The built environment features related to the starting and ending points of the OD flow are extracted, and the historical time period of the OD flow is evenly divided into several intervals. The OD flow of each travel mode in each time period is summarized to obtain the time series features; Based on the graph convolutional network, the built environment features and time series features are processed respectively. The graph convolutional network sets multiple convolutional layers and an output layer. The dimension of the convolutional layer is set to 64 or 128, and the dimension of the output layer is set to 1. The activation function uses Relu, and the model is optimized using the MSE loss function and Adam optimizer to obtain a high-dimensional feature matrix of the built environment mobile flow. , and the dimension is expressed as ; Based on the high-dimensional feature matrix, the inner product of the high-dimensional feature matrix is ​​calculated by the relationship matrix calculation formula. The similarity of the high-dimensional features is expressed by the inner product. The relationship matrix calculation formula is: , represents the inner product of a high-dimensional feature matrix and has dimensions , Represents a high-dimensional feature matrix; Based on the high-dimensional feature matrix and similarity, the built environment feature map is constructed and expressed as: ; in, express Kind of OD flow, The dimension is The characteristic matrix of is the inner product, Represents the feature matrix of a node.

[0034] The final view of each travel mode is as follows:

[0035] in, represents the nth travel mode view, They represent the spatial adjacency graph, the traffic characteristic graph, and the built environment characteristic graph respectively.

[0036] In one embodiment, the spatiotemporal dependency feature extraction module receives each travel mode view and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network, including: The spatiotemporal dependency feature extraction module includes a multi-layer temporal graph convolutional network, wherein the temporal graph convolutional network includes a temporal convolutional network and a graph convolutional network, wherein the temporal convolutional network captures temporal dependency features, and the graph convolutional network captures spatial dependency features; The temporal convolutional network includes a causal dilated convolutional structure and a gated recurrent unit. The causal dilated convolutional structure is used to learn the dynamic changes of the OD flow in each travel mode view to capture the time-dependent features, and the gated recurrent unit is used to capture the time-dependent features again. Constructing filters in the Fourier domain of the graph convolutional network to capture spatial dependency features between nodes; The causal dilated convolution structure is a one-dimensional convolution, expressed as: ; The gated recursive unit is expressed as: ; The graph convolutional network is represented as: ; in, represents the features after dilated causal convolution, represents the time-dependent feature, Indicates size The core, represents the dilation factor that controls the jump distance, represents the step length, denote convolution filters, represents the sigmoid function, They represent the coefficient of deviation, represents the time-dependent features captured again, represents the convolution kernel, express The symmetric Laplacian matrix of represents the identity matrix, express The diagonal matrix of Represents spatially dependent features.

[0037] Finally, the spatiotemporal dependency features corresponding to each travel mode view are extracted through the spatiotemporal dependency feature extraction module, and the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through the self-attention mechanism, including the following steps: Get each travel mode view, concatenate each travel mode view in its own dimension and transform the dimension to form a merged tensor; Through the attention mechanism, the merged tensor is mapped into query, key and value vectors through the convolutional layer to capture the interactive dependencies between the same type of graphs of multiple travel modes and obtain interactive dependency features; The attention mechanism weights are expressed as follows:

[0038] in, represents the interactive dependency feature, Represents the dimension of the feature, They represent query vector, key vector and value vector respectively.

[0039] The final fusion features are expressed as follows:

[0040] in, represents the learnable parameter matrix, represents the Hadamard product, They represent interactive dependency features respectively.

[0041] During the training process, the loss function is used to adjust the optimal parameters of the model. The loss function in the present invention is obtained by the following method: Constructing a first loss function and a second loss function respectively, and obtaining a total loss function based on the first loss function and the second loss function; The first loss function is expressed as follows:

[0042] in, represents the unobstructed OD flow, represents the predicted value, Indicates the actual value, mask indicates that the matrix file value is 0 or 1, indicating whether the corresponding OD flow is blocked; if the value is blocked, no back propagation will be performed. This method can significantly improve the prediction performance of unblocked OD. Due to the sparsity of OD data, ODs with smaller flow can be blocked. In this example, ODs with a flow of 0 are blocked; The second loss function is used to enhance the identification of interactive dependencies between modes. The cosine distance is calculated for the graphs of the same category of the two travel modes, which is expressed as follows:

[0043] The loss function is expressed as follows: ;in, Represent the importance weights corresponding to the first loss function and the second loss function respectively.

[0044] Finally, the model testing module tests and evaluates the trained model. The mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE) are used as evaluation models to test and evaluate the prediction model. The prediction output module can call the trained model and obtain the prediction results of the OD flow by inputting historical data.

[0045] See below for a specific embodiment: Collect relevant data of the study area in the historical time period, such as historical OD data, point of interest data, and road network data; Historical OD data records the starting position and destination of an individual or object's movement in space, as well as the time when the movement occurred. This type of data includes data on various travel modes such as taxis, subways, and shared bicycles. For data in the form of trajectories, further processing is required to extract OD data.

[0046] Points of interest data, or POI data, can be obtained from open data sets. For example, multiple types of POI data of the study area can be obtained from the Amap platform, including companies, shopping services, transportation facilities, science and education and cultural services, accommodation services, tourist attractions, cultural media, leisure and entertainment services, medical care services, sports and leisure services, and residential areas. The data of various types of points of interest in each spatial unit of the study area can be selectively aggregated and counted. Road network data is obtained from open data platforms such as Amap and OpenStreetMap, and the road network density in each spatial unit is counted.

[0047] The present invention is to predict the travel demand between each starting point and destination pair in multiple travel modes based on the data of historical time step T. After dividing the study area into spatial units, the historical OD data of multiple travel modes are spatiotemporally aligned to extract the OD flow matrix. For example, in this example, the study area is divided into 1 km*1 km grids, which are coded from 1 to N in sequence. The starting point and destination of each travel mode are located in the spatial unit, and N grids are used as the starting point and the destination respectively, so it can be obtained From the time perspective, set the time interval of the historical time period to 30 minutes, and obtain a total of M time periods. For a given time period and any two spatial units and , statistical time period Inside, starting point falls into And the end point falls into The number of start-end pairs (i.e., OD pairs) is calculated, and the flow of each OD flow in each time period is extracted in turn. When extracting, it should be ensured that multiple travel modes are aligned in time and space. Then the dimensions of the data matrix of a certain travel mode are , that is, the OD flow matrix of M time periods is extracted.

[0048] If the OD flows are taken as nodes of the graph structure and the associations between the OD flows are taken as edges of the graph structure, the graph structure is constructed. The graph structure is expressed as , where V represents a node set and the dimension is ; A represents the dimension The feature matrix of is the feature matrix describing the edge; X is the feature matrix of the node, expressed as , represents the historical travel demand of the two travel modes for each node in graph G.

[0049] The travel prediction problem is to learn a function that can predict the characteristics of the graph in the next time period under two travel modes given a graph or multiple graphs G and their feature matrix X.

[0050] In one embodiment, the data of taxi and shared bicycle travel modes are taken as an example, and are divided into taxi view and shared bicycle view according to the travel mode. In each travel mode view, the spatial proximity relationship, flow relationship and built environment relationship between OD flows are considered, so three relationship graphs are constructed, namely spatial adjacency graph, flow characteristic graph and built environment characteristic graph.

[0051] A spatial adjacency graph is constructed based on the proximity relationship of OD flows , if two OD flows are adjacent, the weight is 1, and if they are not adjacent, the weight is 0. Here, the neighbor relationship of the OD flow is determined according to the adjacency relationship between the starting point and the end point. For a given OD flow, there are three types of spatial neighbors, including OD flows that are adjacent to the starting point and end point of the current OD flow, OD flows that are adjacent to the starting point of the current OD flow and have the same end point, and OD flows that have the same starting point as the current OD flow and are adjacent to the end point.

[0052] Constructing traffic feature graph , It is represented by connecting each OD flow with several OD flows with the highest flow rate, which is used to capture the mobility characteristics of the OD flow.

[0053] Constructing a built environment characterization map , based on representation learning, high-dimensional features of mobile flows in built environments are extracted to calculate the similarity between OD flows.

[0054] Here we focus on constructing the built environment feature map. First, the POI data and road network data are processed, and the various types of historical point of interest data and historical road network data in each spatial unit are counted. Here, taking POI divided into three categories as an example, a total of four built environment features are obtained for each spatial unit, namely, three types of POI density and road network density. For a given OD flow, the built environment features of the starting point and the end point are extracted, a total of 8 features. At the same time, the historical time period of the OD flow is divided into 12 intervals according to the time of the day, each interval is 2 hours, namely 00:00-02:00, ..., 22:00-24:00. The traffic flow of various travel modes at the starting point and the end point in each time period is summarized, and a total of 24 time series features are obtained for the two travel modes.

[0055] On this basis, we construct a graph , where V is a node, representing Kind of OD flow; The dimension is Here, the adjacency matrix is ​​constructed using the spatial proximity relationship of the flow; X is the feature matrix of the node, and the input feature matrix is ​​the extracted built environment features and time series features, a total of 32 features.

[0056] The feature map is processed using a graph convolutional network. Multiple convolutional layers and an output layer can be set. The dimension of the convolutional layer is set to 64 or 128, the dimension of the output layer is set to 1, the activation function uses Relu, and the model is optimized using the MSE loss function and the Adam optimizer. The extracted built environment-mobile flow high-dimensional feature matrix F is obtained, and its dimension is Then, for the extracted high-dimensional feature matrix F, the relationship matrix is ​​calculated to express the similarity of the features. The calculation formula of the relationship matrix R is: , that is, calculating the inner product of the high-dimensional features of the built environment of the OD flow and the mobile flow, whose dimension is ; Then the built environment feature map is constructed, which is expressed as: ; in, express Kind of OD flow, The dimension is The characteristic matrix of is the inner product, The feature matrix representing the node is the built environment features and time series features.

[0057] To perform multi-view learning, construct a taxi view and shared bike view , each view includes three types of graphs: spatial adjacency graph, traffic feature graph and built environment feature graph, then: ; .

[0058] Although the three types of graphs have different edges, the nodes are the same, and the feature X of the travel mode belongs to the node. The six graphs are input into the constructed traffic prediction pre-training model to extract features and fit the model.

[0059] The traffic prediction pre-training model or model training module includes an intra-mode learning module, a spatiotemporal dependency feature extraction module, an inter-mode learning module, an interactive dependency feature extraction module, and a feature fusion module. The entire traffic prediction pre-training model can be found in the attached Figure 4 As shown in the figure, the intra-mode learning module is mainly composed of TGCN, which combines the temporal convolutional network (TCN) and the graph convolutional network (GCN). TCN includes convolutional neural network (CNN) and gated recurrent unit (GRU). GCN is used to capture spatial dependency features from the graph, and TCN is mainly used to learn the dynamic changes of OD flow to capture temporal dependency features. Specifically, TCN uses a causal dilated convolution structure to capture the temporal dependency features of the OD flow. The convolution is a one-dimensional convolution that slides the input by skipping a certain step size. For each graph , is the time series feature of view v. Under step size s, the dilated causal convolution can be expressed as ,in, represents the features obtained after dilated causal convolution, is a kernel of size L, and d is the dilation factor that controls the jump distance. Gated recurrent unit (GRU) is used to further capture time-dependent features and can be expressed as ,in and is the convolution filter, represents the sigmoid function, b and c are the bias coefficients. GCN constructs a filter in the Fourier domain and acts on the nodes of the graph, capturing the spatial features between nodes through the first-order neighborhood. Graph convolution can be expressed as ,in is the convolution kernel, is the symmetric Laplacian matrix of G, I is the unit matrix, and D is the diagonal matrix of A. By stacking multiple layers of TGCN, the spatial dependency features of each graph can be effectively captured. The number of stacked layers can be determined according to the actual application scenario. The example is set to 4 layers. In addition, the in-mode learning module involves the setting of the temporal convolution kernel, the spatial convolution kernel size, and the number of convolution channels. The example sets the temporal and spatial convolution kernel sizes to 3 and the number of convolution channels to 128. Overall, the six graphs are placed in TGCN respectively, and six spatiotemporal feature matrices of the three graphs under the two travel modes are extracted. Each spatiotemporal feature matrix represents the spatiotemporal dependency features of the corresponding graph itself in the historical data. Refer to the attached Figure 4As shown in the figure, the spatial adjacency graph, flow characteristic graph and built environment characteristic graph constructed based on the taxi OD flow will be processed by the in-pattern learning module, namely the spatiotemporal dependency feature extraction module, to extract the spatiotemporal dependency features corresponding to the spatial adjacency graph, flow characteristic graph and built environment characteristic graph. Similarly, the spatial adjacency graph, flow characteristic graph and built environment characteristic graph based on the shared bicycle OD flow are processed in the same way, and then the spatiotemporal dependency features corresponding to the spatial adjacency graph, flow characteristic graph and built environment characteristic graph are extracted.

[0060] The inter-modal learning module is used to fuse the spatiotemporal dependency features extracted from multiple graphs to capture the interactive dependencies between travel modes or traffic modes. The inter-modal learning module mainly relies on the self-attention mechanism (Self-Attention) to map the input data to query (Q), key (K), and value (V) vectors through the convolution layer, which can generate more efficient feature expressions in the spatial dimension, thereby achieving more accurate feature capture and information fusion. The input data is a tensor containing multiple view features. The multi-view features are concatenated and transformed in dimension to adapt to subsequent convolution operations. If there are more than one view features, they will be concatenated along the last dimension to form a merged tensor. The formula for calculating the attention weight is ,in, is the dimension of the feature, Q, K, V are the query vector, key vector and value vector respectively, which are generated by merging the tensor through a 1D convolution layer, and then the attention weight is calculated through the softmax function and multiplied with V to obtain the attention value. The final output is mapped through a linear layer, and the ReLU activation function and Dropout are used to prevent overfitting. In the example, the spatial adjacency graphs of the taxi and shared bicycle views are connected, the traffic feature graphs of the two are connected, and the built environment feature graphs of the two are connected. A total of three sets of self-attention mechanism operations are performed to capture the interactive dependencies between the same graphs of the two travel modes, such as the interactive dependencies between the spatiotemporal dependency features extracted from the taxi OD flow and the spatiotemporal dependency features extracted from the shared bicycle OD flow, and then the interactive dependency features between the same graphs of the two travel modes are obtained. The three types of features , , Of course, if there are multiple travel modes, it is sufficient to capture the interactive dependency features between the same graphs of two travel modes.

[0061] The feature fusion module uses a fusion method based on a parameter matrix to fuse the three types of features. The formula is: ,in , and is the learnable parameter matrix, and ⊙ represents the Hadamard product. In order to achieve effective model training, two loss functions are used to constrain the model training process. The main loss function uses the masked mean square error function (Masked MSE), and the formula is ,in represents the unmasked OD number, is the predicted value, is the actual value, mask is a matrix file with values ​​of 0 and 1, indicating whether the corresponding OD flow is masked. If the value is masked, no backpropagation will be performed. This method can significantly improve the prediction performance of unmasked OD. Due to the sparsity of OD data, ODs with smaller flow can be masked. In this example, ODs with a flow of 0 are masked. The auxiliary loss function is used to enhance the identification of interactive dependencies between modes. The cosine distance is calculated for the graphs of the same category of the two travel modes, and the formula is as follows: .

[0062]

[0063] The total loss function is ,in and Used to control the importance of the two losses.

[0064] The model testing module tests and evaluates the trained model. The mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE) are used as evaluation models to test and evaluate the prediction model. The prediction output module can call the trained model and obtain the prediction results of the OD flow by inputting historical data.

[0065] In view of the above problems, the present invention designs a collaborative prediction method for residents' multimodal transportation travel flow based on multi-task learning. The present invention integrates multi-source geographic data and deep learning methods, and considers the interactive dependence and collaborative influence of built environment characteristics and multimodal transportation, and constructs a high-precision urban residents' transportation travel demand prediction model at the OD scale to support the scientific decision-making of multimodal transportation collaborative management and urban planning.

[0066] Specifically, the present invention focuses on multimodal traffic prediction at the OD scale. Based on the representation learning method, the dynamic characteristics of the built environment characteristics and the crowd movement flow are extracted, and multi-views are constructed for representation respectively with the spatial correlation and flow characteristics of the movement flow. Based on the idea of ​​multi-task learning, the temporal graph convolutional network (TGCN) and the self-attention mechanism (Self-Attention) are used to capture the spatiotemporal dependency characteristics and the interactive dependency between multimodal traffic, and an auxiliary loss function is designed to enhance the prediction. In particular, the present invention uses the masked loss function to reduce the impact of data sparsity on the prediction. Through the above framework, the present invention can realize the collaborative prediction of multiple traffic travel flows; based on the historical OD flow and the built environment data, the multi-traffic OD flow in the subsequent time period is predicted. Compared with the existing methods, the model is relatively simple, the computing power requirement is smaller, and the result accuracy is high.

[0067] Embodiment 2: A multi-mode travel flow collaborative prediction system based on multi-task learning, such as Figure 2 As shown, it includes a data acquisition module 100, a processing and extraction module 200, a feature extraction module 300, a view construction module 400, a model training module 500 and a prediction and reasoning module 600; The data acquisition module 100 is used to acquire relevant data of various travel modes in the area to be tested, wherein the relevant data includes historical start and end point data, point of interest data and road network data, the historical start and end point data includes various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data is data such as city-related infrastructure, services, geographic information, etc., and the road network data is the road network density within each spatial unit; The processing and extraction module 200 is used to divide the area to be tested into unit grids, pre-process the relevant data of multiple travel modes and align them in time and space, and extract the OD flow matrix; The feature extraction module 300 is used to extract the flow similarity features and built environment features and time series features of the OD flow based on the OD flow matrix, and extract the high-dimensional features of the built environment and the mobile flow based on the built environment features and time series features; The view construction module 400 constructs a spatial adjacency graph based on spatial proximity relationships, constructs a flow characteristic graph based on flow similarity features, and constructs a built environment characteristic graph based on high-dimensional features, and constructs each travel mode view through the spatial adjacency graph, the flow characteristic graph, and the built environment characteristic graph; The model training module 500 is used to construct a traffic prediction pre-training model, which includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module. The spatiotemporal dependency feature extraction module receives each travel mode view and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The prediction and reasoning module 600 is used to input the OD flow to be measured into the flow prediction model for prediction, and obtain the prediction result of the next moment under the application scenario at the current moment.

[0068] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also belong to the scope of the present invention.

[0069] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0070] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0071] The present invention is described with reference to the flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0072] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0073] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0074] It should be noted that: The "one embodiment" or "embodiment" mentioned in the specification means that the specific features, structures or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases "one embodiment" or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0075] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different. Any equivalent or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. The technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A multi-mode travel flow collaborative prediction method based on multi-task learning, characterized in that: The following steps are involved: Obtain relevant data of various travel modes in the area to be tested, wherein the relevant data include historical start and end point data, point of interest data and road network data. The historical start and end point data include various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data are city-related infrastructure data, service data and geographic information data, and the road network data are the road network density within each spatial unit; The area to be tested is divided into unit grids, the relevant data of various travel modes are preprocessed and aligned in time and space, and the OD flow matrix is ​​extracted. Based on the OD flow matrix, the flow similarity characteristics of the OD flow, the built environment characteristics and the time series characteristics are extracted; Based on the spatial proximity relationship of OD flow, a spatial adjacency graph is constructed, based on the similarity characteristics of traffic, a traffic feature graph is constructed, and based on the built environment characteristics and time series characteristics, high-dimensional features of the built environment and mobile flow are extracted and a built environment feature graph is constructed; Construct each travel mode view through spatial adjacency graph, flow characteristic graph and built environment characteristic graph; Constructing a traffic prediction pre-training model, the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module, the spatiotemporal dependency feature extraction module receives each travel mode view, and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The OD flow to be measured is input into the traffic prediction model for prediction, and the prediction result for the next moment under the current application scenario is obtained.

2. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1 is characterized in that: The method of dividing the area to be tested into unit grids, performing spatiotemporal alignment processing on the relevant data of multiple travel modes, and extracting the OD flow matrix includes the following steps: The area to be measured is represented as a rectangle and divided into H×W spatial units along longitude and latitude. Each spatial unit has the same size and is encoded to N in sequence. The starting point and end point of each travel mode are located in N spatial units, and there are Kind of OD flow; Select M time periods in the historical time period, extract the flow of each OD flow in each time period in turn, that is, extract the number of OD pairs that fall into the first spatial unit and the end point falls into the second spatial unit, and align multiple travel modes in time and space, then the data matrix dimension is expressed as That is the OD flow matrix of M time periods.

3. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1 is characterized in that: The method of constructing a spatial adjacency graph based on the spatial proximity relationship and constructing a traffic feature graph based on traffic similarity features comprises the following steps: The OD flows are regarded as nodes, and the adjacent relationship between two OD flows is used as the edge of the graph structure to construct a spatial adjacency graph, which is expressed as: ; The flow characteristic diagram is expressed as: , It means that each OD flow is connected with several OD flows that rank top in the OD flow ranking, which is used to capture the mobility characteristics of the OD flow; Among them, according to the proximity relationship of the OD flow, if the two OD flows are in a neighboring relationship, the weight is 1, otherwise the weight is 0. The proximity relationship of the OD flow is determined according to the adjacency relationship between the starting point and the end point. For a given OD flow, there are three types of spatial neighbors, namely, OD flows that are adjacent to the starting point and end point of the current OD flow respectively, OD flows that are adjacent to the starting point of the OD flow and have the same end point, and OD flows that have the same starting point as the OD flow and are adjacent to the end point.

4. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1 is characterized in that: The method of extracting high-dimensional features of the built environment and the mobile flow based on the built environment features and the time series features and constructing a built environment feature map includes the following steps: The point of interest data and road network data are processed, and the point of interest data and road network data of various travel modes in each spatial unit are counted, so as to obtain the built environment characteristics of each spatial unit; The built environment features related to the starting and ending points of the OD flow are extracted, and the historical time period of the OD flow is evenly divided into several intervals. The OD flow of each travel mode in each time period is summarized to obtain the time series features; Based on the graph convolutional network, the built environment features and time series features are processed respectively. The graph convolutional network sets multiple convolutional layers and an output layer. The dimension of the convolutional layer is set to 64 or 128, and the dimension of the output layer is set to 1. The activation function uses Relu, and the model is optimized using the MSE loss function and Adam optimizer to obtain a high-dimensional feature matrix of the built environment mobile flow. , and the dimension is expressed as ; Based on the high-dimensional feature matrix, the inner product of the high-dimensional feature matrix is ​​calculated by the relationship matrix calculation formula. The similarity of the high-dimensional features is expressed by the inner product. The relationship matrix calculation formula is: , represents the inner product of a high-dimensional feature matrix and has the dimension , Represents a high-dimensional feature matrix; Based on the high-dimensional feature matrix and similarity, the built environment feature map is constructed and expressed as: ; in, express Kind of OD flow, The dimension is The characteristic matrix of is the inner product, Represents the feature matrix of a node.

5. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1 is characterized in that: Each travel mode view is represented as follows: in, represents the nth travel mode view, They represent the spatial adjacency graph, the traffic characteristic graph, and the built environment characteristic graph respectively.

6. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1, characterized in that: The spatiotemporal dependency feature extraction module receives each travel mode view and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network, including: The spatiotemporal dependency feature extraction module includes a multi-layer temporal graph convolutional network, wherein the temporal graph convolutional network includes a temporal convolutional network and a graph convolutional network, wherein the temporal convolutional network captures temporal dependency features, and the graph convolutional network captures spatial dependency features; The temporal convolutional network includes a causal dilated convolutional structure and a gated recurrent unit. The causal dilated convolutional structure is used to learn the dynamic changes of the OD flow in each travel mode view to capture the time-dependent features, and the gated recurrent unit is used to capture the time-dependent features again. Constructing filters in the Fourier domain of the graph convolutional network to capture spatial dependency features between nodes; The causal dilated convolution structure is a one-dimensional convolution, expressed as: ; The gated recursive unit is expressed as: ; The graph convolutional network is represented as: ; in, represents the features after dilated causal convolution, represents the time-dependent feature, Indicates size The core, represents the dilation factor that controls the jump distance, represents the step length, denote convolution filters, represents the sigmoid function, They represent the coefficient of deviation, represents the time-dependent features captured again, represents the convolution kernel, express The symmetric Laplacian matrix of represents the identity matrix, express The diagonal matrix of Represents spatially dependent features.

7. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1, characterized in that: The interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism, including the following steps: Obtain the spatiotemporal dependency features of the same graph in each travel mode, concatenate the spatiotemporal dependency features in their respective dimensions, and perform dimension transformation to form a merged tensor; Through the attention mechanism, the merged tensor is mapped into query, key and value vectors through the convolutional layer to capture the interactive dependencies between the same type of graphs of multiple travel modes and obtain interactive dependency features; The attention mechanism weights are expressed as follows: in, represents the interactive dependency feature, represents the dimension of the feature, Represent the query vector, key vector and value vector respectively.

8. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1, characterized in that: The fusion features are expressed as follows: in, represents the learnable parameter matrix, represents the Hadamard product, They represent interactive dependency features respectively.

9. The multi-task learning-based multi-mode travel flow collaborative prediction method according to claim 1, characterized in that: The loss function is obtained in the following way: Constructing a first loss function and a second loss function respectively, and obtaining a total loss function based on the first loss function and the second loss function; The first loss function is expressed as follows: in, represents the unobstructed OD flow, represents the predicted value, Indicates the actual value, mask indicates that the matrix file value is 0 or 1, indicating whether the corresponding OD flow is masked; The second loss function is expressed as follows: The loss function is expressed as follows: ; in, Represent the importance weights corresponding to the first loss function and the second loss function respectively.

10. A multi-mode travel flow collaborative prediction system based on multi-task learning, characterized in that: It includes data acquisition module, processing and extraction module, feature extraction module, view construction module, model training module and prediction and reasoning module; The data acquisition module is used to acquire relevant data of various travel modes in the area to be tested, wherein the relevant data includes historical start and end point data, point of interest data and road network data, the historical start and end point data includes various travel modes, movement occurrence time data and movement start and end point location data; the point of interest data is city-related infrastructure data, service data and geographic information data, and the road network data is the road network density within each spatial unit; The processing and extraction module is used to divide the area to be tested into unit grids, pre-process and align the relevant data of multiple travel modes in time and space, extract the OD flow matrix, and based on the OD flow matrix, extract the flow similarity characteristics of the OD flow and the built environment characteristics and time series characteristics; The feature extraction module constructs a spatial adjacency graph based on the spatial proximity relationship of the OD flow, constructs a flow feature graph based on the flow similarity features, extracts high-dimensional features of the built environment and the mobile flow based on the built environment features and the time series features, and constructs a built environment feature graph; The view construction module constructs each travel mode view through a spatial adjacency graph, a traffic characteristic graph, and a built environment characteristic graph; The model training module is used to construct a traffic prediction pre-training model, and the traffic prediction pre-training model includes a spatiotemporal dependency feature extraction module, an interactive dependency feature extraction module and a feature fusion module. The spatiotemporal dependency feature extraction module receives each travel mode view and extracts the spatiotemporal dependency features corresponding to each travel mode view through a temporal graph convolutional network; the interactive dependency feature extraction module fuses the spatiotemporal dependency features of the same graph in each travel mode and captures the interactive dependency features through a self-attention mechanism; the feature fusion module fuses the spatiotemporal dependency features and the interactive dependency features to obtain a fused feature; the traffic prediction pre-training model is trained based on the traffic feature graph of the historical time period, and the prediction is enhanced using a loss function to obtain a traffic prediction model; The prediction and reasoning module is used to input the OD flow to be measured into the flow prediction model for prediction, and obtain the prediction result of the next moment under the application scenario at the current moment.

11. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 9 is implemented.

12. A multi-mode travel flow collaborative prediction device based on multi-task learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 9 is implemented.

Images