A method for predicting travel time by fusing high-order intention semantics and deep attention mechanism

Abstract

The application discloses a travel time prediction method fusing high-order intention semantics and a deep attention mechanism, and relates to the technical field of intelligent traffic and data mining. The method comprises the following steps: acquiring multi-source heterogeneous traffic data of a city road network and constructing a multi-dimensional space-time feature sequence; performing long tail deviation elimination processing on the feature sequence to obtain balanced feature space data; inputting the balanced data into a travel intention inference model to obtain high-order intention semantic labels; cascading the high-order intention semantic labels as global prior information and path space features; inputting a multi-dimensional fusion feature matrix into a deep table network (TabNet) to output a prediction result through a sparse sequential attention mechanism. The application effectively solves the problem of extreme long tail error caused by the excessive dependence of traditional models on physical distance, and greatly improves the robustness and accuracy of travel time prediction under complex road networks by introducing high-order travel intention and inverse frequency resampling technology.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and traffic big data processing technology, and in particular to a method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms. Background Technology

[0002] Travel time is a crucial indicator for measuring traffic efficiency. It plays a key role in accurately understanding the operational status of road networks, helping traffic management departments balance traffic flow, alleviating traffic congestion, and improving traffic efficiency. Existing travel time prediction models are mostly based on modeling the basic spatiotemporal trajectories of vehicles and road topology.

[0003] However, traditional travel time prediction models (such as physical distance-based tree models) perform reasonably well under normal road conditions, but have significant shortcomings in complex urban road networks. Existing models rely excessively on "Euclidean physical distance" and basic spatiotemporal features, failing to effectively capture travelers' deeper motivations (such as commuting, business, leisure, etc.). When faced with extreme congestion, holiday travel, or special area scheduling, this "physical distance dependence" easily leads to "long-tail error," manifested as extremely poor prediction ability for a few extreme delay extremes. In addition, massive amounts of ride-hailing and other trajectory data often exhibit severe long-tail distribution imbalance (i.e., weekday and core area samples are absolutely dominant). The lack of high-level semantic labels in the underlying trajectory data causes data bias in downstream prediction models, failing to meet the high-precision, dynamic route planning requirements of modern intelligent transportation systems.

[0004] Therefore, there is an urgent need to design a method and system for travel time prediction that integrates high-order intent semantics and deep attention mechanisms to overcome the excessive dependence of traditional models on physical distance and effectively handle the long-tail distribution problem of traffic data. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by providing a travel time prediction method that integrates high-order intent semantics and deep attention mechanisms. This solves the problem that existing technologies suffer from significant long-tail errors in travel time prediction due to neglecting traveler intent and data imbalance under actual road conditions.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms is provided, comprising the following steps: S1: Acquire multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence; the multi-source heterogeneous traffic data includes trajectory sequence data and spatial distribution data of points of interest. S2: Perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain balanced feature space data; S3: Input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels of multiple categories corresponding to the underlying physical features; S4: The higher-order intention semantic tags are used as global prior information and cascaded and fused with path space features to construct a multi-dimensional fusion feature matrix; S5: Input the multidimensional fusion feature matrix into a preset deep attention prediction network, dynamically assign weights to the fusion features through a feature-level attention mechanism, and output the travel time prediction result.

[0007] Preferably, in step S1, the multi-source heterogeneous traffic data includes massive vehicle trajectory data and urban point of interest spatial distribution data; and a multi-dimensional spatiotemporal feature sequence is constructed based on the multi-source heterogeneous traffic data, specifically including: extracting the kinematic features of the trajectory sequence data, constructing spatial topological association information, and combining the point of interest spatial distribution data to construct a four-dimensional spatiotemporal feature bucket containing distance, time, space and calendar attributes.

[0008] Preferably, in step S2, the long-tail bias elimination process is performed on the multidimensional spatiotemporal feature sequence, including the following steps: The sample capacity distribution of each multidimensional spatiotemporal feature sequence is statistically analyzed; the multidimensional spatiotemporal feature sequence is balanced by applying an inverse frequency resampling strategy to reduce redundant samples at the head and improve the sample coverage of sparse class features.

[0009] Preferably, in step S3, the travel intention inference model is an ensemble learning classification model that obtains high-order intention semantic labels for multiple categories corresponding to the underlying physical features, specifically including: S31: Based on the improved gravity model and the spatial distribution data of the points of interest, calculate the spatial attraction index of the preset point of interest category in the target area; S32: Input the spatial attractiveness index and the equalized feature space data into the ensemble learning classification model, output the probability distribution of discrete semantic categories including production and office, leisure and entertainment, and extract the category with the highest probability as the higher-order intention semantic label.

[0010] Preferably, in step S32, the ensemble learning classification model is iteratively optimized using an objective function, the expression of which is:

[0011] in, This represents the set of parameters that the model needs to learn during the iterative optimization process; This is a multi-class cross-entropy loss function used to measure the deviation between the predicted intent and the true label; Indicates the first The high-level travel intent real label corresponding to each trajectory sample; Indicates the classifier for the first... The high-order travel intention prediction value output from each trajectory sample; Indicates the first The structure of the basic decision tree and the functional mapping of the predicted scores of the leaf nodes; For the first The regularization term of the decision tree, where To control the penalty coefficient for the difficulty of splitting leaf nodes, For the first The total number of leaf nodes in the decision tree. This is the weight decay coefficient of the L2 regularization penalty term. For the first The output weight vector of each leaf node of the decision tree.

[0012] Preferably, the deep attention prediction network in step S5 is a deep table network, and the step of dynamically assigning weights to the fused features through a feature-level attention mechanism to output the travel time prediction result includes the following steps: S51: By using the sparse sequential attention mechanism of the deep table network, the heterogeneity of traffic flow micro-features under different intention semantics is captured; S52: Based on the nonlinear mapping relationship between the higher-order intention semantic tags and the micro-road network features, output the travel time regression prediction value.

[0013] Preferably, in step S51, the deep table network actively filters high-order intent semantics through sequential attention masks, and the mask generation formula is as follows:

[0014] in, For the first The feature mask of the step, Indicates the previous decision step, i.e., the first step. The prior scale factor passed through each step is updated to record the frequency of use of each feature in historical steps, preventing the overuse of a single feature. express In the first decision step, a feature processing network function for a specific mapping is executed; This refers to the feature information obtained after the previous step of processing; Furthermore, nonlinear mapping is performed using gated linear units, and the calculation process is as follows:

[0015] in, This represents the nonlinear mapping result vector output by the gated linear unit after performing feature transformation on the input; This represents the local feature vector input to the gated linear unit; These represent the learnable weight matrices used for feature linear transformation and gating signal generation in the neural network structure, respectively. , These respectively represent the weight matrices mentioned above. The corresponding learnable bias term vector; express Activation function.

[0016] Preferably, the training process of the deep attention prediction network employs a joint loss function for parameter optimization, the expression of which is:

[0017] in, Let be the joint loss function of the deep attention prediction network during training; the first term is the mean squared error loss term, and the second term is the sparse entropy regularization loss term. This represents the total number of trajectory samples in the current training batch. This is the index number of a single sample in the training batch; For the model to the first The predicted travel time for each sample output; For the first The actual travel time corresponding to each sample; To control the sparsity of model features, an entropy regularization penalty coefficient is selected. The total number of decision steps set for a deep table network; This represents the iteration number of the decision step; This represents the total number of dimensions of the input feature matrix; The index number for the feature dimension; In the first In the decision step, the allocation to the first Attention mask weights for dimensional features; The term "tiny positive number" was introduced to prevent undefined mathematical operations in logarithmic operations.

[0018] According to another aspect of the present invention, a travel time prediction system integrating high-order intent semantics and deep attention mechanisms is provided, employing the above-mentioned method, comprising: The feature construction module is used to acquire multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence; The equalization sampling module is used to perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain equalized feature space data. The semantic inference module is used to input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels; The feature fusion module is used to cascade and fuse the high-order intent semantic tags as global prior information with path space features to construct a multi-dimensional fusion feature matrix. The deep prediction module is used to input the multidimensional fused feature matrix into a preset deep attention prediction network and output the travel time prediction result through a feature-level attention mechanism.

[0019] According to another aspect of the present invention, a computer device includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms.

[0020] In summary, compared with the prior art, the trip time prediction method fused with high-order intent semantics and deep attention mechanism provided by the present invention has the following beneficial effects: 1. Innovation at the Feature Engineering Level: This invention proposes a multi-dimensional spatiotemporal feature equalization processing method based on inverse frequency resampling, which effectively eliminates the long-tail distribution imbalance of multi-source trajectory data. While maintaining the macroscopic structure of the spatial region, it significantly reduces homogeneous short-distance redundant orders and proactively improves the coverage of relatively scarce semantic categories such as "holidays," achieving both "macroscopic fidelity" and "microscopic purification" of the data, providing a high-quality foundation for downstream tasks.

[0021] 2. Innovation at the algorithm architecture level: For the first time, offline travel purpose is injected into the prediction model as a high-order intention semantic label, breaking the traditional prediction's excessive reliance on Euclidean physical space.

[0022] 3. Significant improvement in prediction performance: By leveraging TabNet's sparse sequential attention mechanism to dynamically assign weights to high-order semantics, compared to traditional ensemble tree models, this invention can effectively absorb and suppress long-tailed outliers caused by extreme congestion or special events, significantly reducing prediction error (MAE) and improving the prediction robustness and reliability of intelligent transportation systems. Attached Figure Description

[0023] Figure 1 The flowchart illustrates a trip time prediction method that integrates high-order intent semantics and deep attention mechanisms, as provided in an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the fusion model module in an embodiment of the present invention.

[0025] Figure 3 This is a global feature attention weight distribution diagram of the deep table network (TabNet) in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0027] The data input in this embodiment of the invention consists of 6,584,397 orders from approximately 17,000 ride-hailing vehicles in a central urban area from October 1, 2016 to November 30, 2016. The total number of cumulative trajectory sampling points for the orders exceeds 2 billion, and the original sampled ride-hailing vehicles account for about 70% of the daily active ride-hailing vehicles, which has extremely high sampling representativeness.

[0028] Please see Figure 1 This invention provides a method for predicting travel time that integrates high-order intent semantics and deep attention mechanisms, specifically including the following steps: Step S1: Obtain multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence.

[0029] Among them, multi-source heterogeneous traffic data includes massive vehicle trajectory data (composed of departure time, arrival time, origin and destination latitude and longitude) and spatial distribution data of city points of interest (POI).

[0030] Furthermore, kinematic features of the trajectory data are extracted, and the origin and destination points are mapped to the urban core area and suburbs using a Geographic Information System (GIS) polygon matching algorithm. Simultaneously, calendar attributes (such as weekdays and holidays) and time-slice attributes (such as morning peak, evening peak, off-peak, and nighttime) of the orders are extracted. This constructs a four-dimensional spatiotemporal feature bucket containing distance, time, space, and calendar attributes, forming a multi-dimensional spatiotemporal feature sequence.

[0031] Step S2: Perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain balanced feature space data; Because weekday and core area travel constitutes the vast majority of traffic under normal conditions, the data exhibits a severe long-tail distribution. This embodiment addresses this by statistically analyzing the sample size of each multidimensional spatiotemporal feature bucket and applying an inverse frequency resampling strategy to equalize the data.

[0032] Among these features, the processing significantly reduced the highly homogeneous, redundant short-distance orders within the core area, while retaining cross-regional business travel and long-distance commuting samples, keeping the core area sample proportion at approximately 52% and maintaining "macro-fidelity." At the same time, it strongly boosted the holiday samples, which accounted for only about 36% in the original distribution, to nearly 50%, achieving "micro-purification" of semantic coverage and thoroughly correcting the imbalance in the natural data.

[0033] Step S3: Input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels of multiple categories corresponding to the underlying physical features; This step aims to inversely transform the underlying physical coordinates into a person's higher-order travel intention. The travel intention inference model is an ensemble learning classification model that obtains higher-order intention semantic labels for multiple categories corresponding to the underlying physical features, specifically including: S31: Based on the improved gravity model and the spatial distribution data of the points of interest, calculate the spatial attraction index of the preset point of interest category in the target area; S32: Input the spatial attractiveness index and the equalized feature space data into the ensemble learning classification model, output the probability distribution of discrete semantic categories, and extract the category with the highest probability as the higher-order intention semantic label.

[0034] Specifically, based on the improved gravity model, and combined with the POI entropy and context density of the target area, the spatial attractiveness index of different POI categories (such as office buildings, shopping malls, and residences) is calculated.

[0035] Subsequently, the spatial attractiveness index and dwell time features are input into a pre-trained gradient boosting tree (XGBoost) ensemble learning classifier, which outputs the probability distribution of discrete semantic categories (such as seven travel purposes including production and office, leisure and entertainment, etc.) and extracts the category with the highest probability as the global "high-order intention semantic label" of the trajectory.

[0036] In step S32, the ensemble learning classifier adopts an objective function. The expression for iterative optimization is as follows:

[0037] in, This represents the set of parameters that the model needs to learn during the iterative optimization process; This is a multi-class cross-entropy loss function used to measure the deviation between the predicted intent and the true label; Indicates the first The high-level travel intent real label corresponding to each trajectory sample; Indicates the classifier for the first... The high-order travel intention prediction value output from each trajectory sample; Indicates the first The structure of the basic decision tree and the functional mapping of the predicted scores of the leaf nodes; For the first The regularization term of the decision tree, where To control the penalty coefficient for the difficulty of splitting leaf nodes, For the first The total number of leaf nodes in the decision tree. This is the weight decay coefficient of the L2 regularization penalty term. For the first The output weight vector of each leaf node in the decision tree; by limiting the number of leaf nodes and weighting coefficients This ensures that the model has strong generalization ability when extracting higher-order semantics such as "office" and "leisure".

[0038] Step S4: Use the higher-order intent semantic tags as global prior information and cascade them with path space features to construct a multi-dimensional fusion feature matrix; Specifically, the higher-order intent priors obtained in step S3 are concatenated and spliced ​​with the underlying spatial features such as the starting point coordinates, ending point coordinates, and physical distance of the road network in the channel dimension to construct a multi-dimensional fusion feature matrix that contains both micro-topological details and macro-motivational interpretability.

[0039] Step S5: Input the multidimensional fusion feature matrix into a preset deep attention prediction network, dynamically assign weights to the fusion features through a feature-level attention mechanism, and output the travel time prediction result.

[0040] Specifically, the fused feature matrix is ​​fed into a deep TabNet network configured with a sparse sequential attention mechanism. This network dynamically assigns attention weights to different input features through internal decision steps. For samples experiencing extreme congestion or special holidays, the network adaptively amplifies the weights of "higher-order intent semantic labels" and "calendar attributes," while mitigating the misleading influence of "physical distance." Finally, a regression prediction of travel time is output based on a non-linear mapping relationship.

[0041] In step S5, the deep table network actively filters high-order intent semantics through a sequential attention mask (Attentive Transformer), and the mask generation formula is as follows:

[0042] in, For the first The feature mask of the step, Indicates the previous decision step (the first step) (Step) Update the prior scale term passed in the previous steps to record the frequency of use of each feature in the historical steps and prevent a single feature from being overused. express In the first decision step, a feature processing network function for a specific mapping is executed (usually containing fully connected operations). This represents the feature information processed in the previous step. Using the sparsemax activation function, the model forces the weights of irrelevant physical features to zero, while granting higher activation gain to the "higher-order intent semantic labels." Subsequently, a nonlinear mapping is performed using a gated linear unit (GLU), the calculation process of which is as follows:

[0043] in, This represents the nonlinear mapping result vector output by the gated linear unit after performing a feature transformation on the input; This represents the local feature vector input to the gated linear unit; These represent the learnable weight matrices used for feature linear transformation and gating signal generation in the neural network structure, respectively. , These respectively represent the weight matrices mentioned above. The corresponding learnable bias vector; This represents the Sigmoid activation function, which acts as a gating valve to map data to the (0, 1) interval, thereby dynamically controlling the information flow ratio of each feature channel. This structure ensures that the model can accurately capture the strong nonlinear correlation between intent features and abnormal travel times when processing long-tail congestion data.

[0044] The improved training process and loss function design of the model in this invention are as follows: To ensure that the model can absorb high-order semantics without losing micro-level road network features, this embodiment designs a joint loss function. It is composed of a weighted average of regression loss and sparse regularization loss:

[0045] in, Let be the joint loss function of the deep attention prediction network during training; the first term is the mean squared error loss term, and the second term is the sparse entropy regularization loss term. This represents the total number of trajectory samples in the current training batch. This is the index number of a single sample in the training batch; For the model to the first The predicted travel time for each sample output; For the first The actual travel time corresponding to each sample; To control the sparsity of model features, an entropy regularization penalty coefficient is selected. The total number of decision steps set for a deep table network; This represents the iteration number of the decision step; This represents the total number of dimensions of the input feature matrix; The index number for the feature dimension; In the first In the decision step, the allocation to the first Attention mask weights for dimensional features; The term "tiny positive number" was introduced to prevent undefined mathematical operations in logarithmic operations.

[0046] The training process for this model is described below: Pre-training phase: First, the classifier in step S3 is pre-trained using labeled OD survey data. After freezing the parameters, the intent is labeled on the full trajectories. Warm start phase: Input the concatenated feature matrix into TabNet and initialize the parameters using the Adam optimizer; Dynamic adjustment phase: A learning rate decay strategy is introduced. If the validation set loss no longer decreases after every 50 epochs, the learning rate is reduced to 0.5 times the original value. Convergence criterion: When the joint loss function The rate of change over 10 consecutive epochs is less than When the time comes, stop training and save the model weights.

[0047] Experimental evaluation results: The predicted travel times output by the model based on this invention are compared with the actual values ​​to analyze its reliability. Evaluation metrics used include mean absolute error (MAE). Experiments show that, through a sampling strategy of "macroscopic fidelity and microscopic refinement" and the introduction of higher-order intent semantics, the TabNet model successfully overcomes the prediction bottleneck of traditional tree models on long-tailed data. In comparative experiments, the prediction accuracy of the method of this invention is significantly improved, with its mean absolute error (MAE) decreasing by approximately 31.9% compared to the traditional basic model. Please also refer to... Figure 3 The importance of the features output by this model objectively proves that the "Travel Purpose" feature has received extremely high global attention weight, which greatly enhances the interpretability and robustness of the model's predictions.

[0048] Please see Figure 2This embodiment also provides a trip time prediction system that integrates high-order intent semantics and deep attention mechanisms, employing the above-mentioned method, which includes: The feature construction module is used to acquire multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence; The equalization sampling module is used to perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain equalized feature space data. The semantic inference module is used to input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels; The feature fusion module is used to cascade and fuse the high-order intent semantic tags as global prior information with path space features to construct a multi-dimensional fusion feature matrix. The deep prediction module is used to input the multidimensional fused feature matrix into a preset deep attention prediction network and output the travel time prediction result through a feature-level attention mechanism.

[0049] This embodiment also provides a computer device, including a memory and a processor. The memory stores a program, and when the program is executed by the processor, it causes the processor to perform the steps of a traffic travel time optimization method.

[0050] The above description, in conjunction with specific preferred embodiments, provides a more detailed explanation of the present invention. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such deductions or substitutions should be considered to fall within the scope of protection of the present invention.

Claims

1. A method for predicting travel time that integrates high-order intent semantics and deep attention mechanisms, characterized in that, include: S1: Acquire multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence; the multi-source heterogeneous traffic data includes trajectory sequence data and spatial distribution data of points of interest. S2: Perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain balanced feature space data; S3: Input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels of multiple categories corresponding to the underlying physical features; S4: The higher-order intent semantic tags are used as global prior information and cascaded and fused with path space features to construct a multi-dimensional fusion feature matrix; S5: Input the multidimensional fusion feature matrix into a preset deep attention prediction network, dynamically assign weights to the fusion features through a feature-level attention mechanism, and output the travel time prediction result.

2. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 1, characterized in that, In step S1, the multi-source heterogeneous traffic data includes massive vehicle trajectory data and urban point of interest spatial distribution data; and a multi-dimensional spatiotemporal feature sequence is constructed based on the multi-source heterogeneous traffic data, specifically including: extracting the kinematic features of the trajectory sequence data, constructing spatial topological association information, and combining the point of interest spatial distribution data to construct a four-dimensional spatiotemporal feature bucket containing distance, time, space and calendar attributes.

3. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 1, characterized in that, In step S2, the long-tail bias elimination process is performed on the multidimensional spatiotemporal feature sequence, including the following steps: The sample capacity distribution of each multidimensional spatiotemporal feature sequence is statistically analyzed; the multidimensional spatiotemporal feature sequence is balanced by applying an inverse frequency resampling strategy to reduce redundant samples at the head and improve the sample coverage of sparse class features.

4. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 1, characterized in that, In step S3, the travel intention inference model is an ensemble learning classification model that obtains high-order intention semantic labels for multiple categories corresponding to the underlying physical features, specifically including: S31: Based on the improved gravity model and the spatial distribution data of the points of interest, calculate the spatial attraction index of the preset point of interest category in the target area; S32: Input the spatial attractiveness index and the equalized feature space data into the ensemble learning classification model, output the probability distribution of discrete semantic categories, and extract the category with the highest probability as the higher-order intention semantic label.

5. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms according to claim 4, characterized in that, In step S32, the ensemble learning classification model is iteratively optimized using an objective function, the expression of which is: in, This represents the set of parameters that the model needs to learn during the iterative optimization process; This is a multi-class cross-entropy loss function used to measure the deviation between the predicted intent and the true label; Indicates the first The high-level travel intent real label corresponding to each trajectory sample; Indicates the classifier for the first... The high-order travel intention prediction value output from each trajectory sample; Indicates the first The structure of the basic decision tree and the functional mapping of the predicted scores of the leaf nodes; For the first The regularization term of the decision tree, where To control the penalty coefficient for the difficulty of splitting leaf nodes, For the first The total number of leaf nodes in the decision tree. This is the weight decay coefficient of the L2 regularization penalty term. For the first The output weight vector of each leaf node of the decision tree.

6. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 1, characterized in that, The deep attention prediction network mentioned in step S5 is a deep table network. The step of dynamically assigning weights to the fused features through a feature-level attention mechanism and outputting the travel time prediction result includes the following steps: S51: By using the sparse sequential attention mechanism of the deep table network, the heterogeneity of traffic flow micro-features under different intention semantics is captured; S52: Based on the nonlinear mapping relationship between the higher-order intention semantic tags and the micro-road network features, output the travel time regression prediction value.

7. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 6, characterized in that, In step S51, the deep table network actively filters high-order intent semantics through sequential attention masks, and the mask generation formula is as follows: in, For the first The feature mask of the step, Indicates the previous decision step, i.e., the first step. The prior scale factor passed through each step is updated to record the frequency of use of each feature in historical steps, preventing the overuse of a single feature. express In the first decision step, a feature processing network function for a specific mapping is executed; This refers to the feature information obtained after the previous step of processing; Furthermore, nonlinear mapping is performed using gated linear units, and the calculation process is as follows: in, This represents the nonlinear mapping result vector output by the gated linear unit after performing feature transformation on the input; This represents the local feature vector input to the gated linear unit; These represent the learnable weight matrices used for feature linear transformation and gating signal generation in the neural network structure, respectively. , These respectively represent the weight matrices mentioned above. The corresponding learnable bias term vector; This represents the Sigmoid activation function.

8. The method for travel time prediction that integrates high-order intent semantics and deep attention mechanisms as described in claim 6, characterized in that, The training process of the deep attention prediction network uses a joint loss function for parameter optimization. The expression of the joint loss function is as follows: in, Let be the joint loss function of the deep attention prediction network during training; the first term is the mean squared error loss term, and the second term is the sparse entropy regularization loss term. This represents the total number of trajectory samples in the current training batch. This is the index number of a single sample in the training batch; For the model to the first The predicted travel time for each sample output; For the first The actual travel time corresponding to each sample; To control the sparsity of model features, an entropy regularization penalty coefficient is selected. The total number of decision steps set for a deep table network; This represents the iteration number of the decision step; This represents the total number of dimensions of the input feature matrix; The index number of the feature dimension; In the first In the decision step, the decision is assigned to the first... Attention mask weights for dimensional features; The term "tiny positive number" is introduced to prevent undefined mathematical operations in logarithmic operations.

9. A travel time prediction system integrating high-order intent semantics and deep attention mechanisms, comprising the method described in any one of claims 1-8, characterized in that, include: The feature construction module is used to acquire multi-source heterogeneous traffic data of the urban road network and construct a multi-dimensional spatiotemporal feature sequence; The equalization sampling module is used to perform long-tail bias elimination processing on the multidimensional spatiotemporal feature sequence to obtain equalized feature space data. The semantic inference module is used to input the equalized feature space data into the pre-trained travel intention inference model to obtain high-order intention semantic labels; The feature fusion module is used to cascade and fuse the high-order intent semantic tags as global prior information with path space features to construct a multi-dimensional fusion feature matrix. The deep prediction module is used to input the multidimensional fused feature matrix into a preset deep attention prediction network and output the travel time prediction result through a feature-level attention mechanism.

10. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a program that, when executed by the processor, causes the processor to perform the steps of a trip time prediction method that integrates high-order intent semantics and deep attention mechanisms as described in any one of claims 1 to 8.

Images