Highway toll station traffic flow spatiotemporal prediction method fusing multi-source information

Abstract

The application relates to the field of intelligent traffic systems and data processing technology, and discloses a highway toll station traffic flow space-time prediction method fusing multiple source information, which comprises the following steps: based on toll station flow and vehicle origin-destination point data, constructing node flow feature tensor and basic edge feature tensor containing time delay attribute; using real-time meteorological data to calculate dynamic impedance factor to correct the basic edge feature, and generating meteorological enhanced dynamic edge feature tensor; constructing a space-time graph attention network, explicitly introducing dynamic edge features to calculate attention coefficients, and predicting short-time flow; based on the toll station operation efficiency score, constructing a negative correlation weight and an efficiency weighted loss function, and optimizing model parameters; generating control instructions and feeding back actual traffic data to update the operation efficiency score, so that a closed-loop application is realized. Through meteorological sensing dynamic enhancement of edge features and efficiency feedback weighted loss function, high-precision closed-loop prediction of key congestion nodes in a complex environment is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent transportation systems and data processing technology, specifically a method for spatiotemporal prediction of traffic flow at highway toll stations that integrates multi-source information. Background Technology

[0002] In highway operation and management, toll stations, as key nodes in the road network, directly impact the overall service level of the network due to their traffic efficiency. With increasing traffic volume and frequent extreme weather events, toll station areas are prone to becoming traffic bottlenecks. Accurate short-term traffic flow forecasting is a prerequisite for dynamic allocation of lane resources, route guidance, and congestion mitigation. However, highway traffic flow exhibits high nonlinearity and spatiotemporal dynamics, and is highly susceptible to complex influences from meteorological conditions and road network topology. Therefore, how to integrate multi-source information to achieve high-precision toll station-level traffic flow forecasting has become a critical issue urgently needing to be addressed in the field of intelligent transportation.

[0003] Existing traffic flow prediction methods mainly include traditional statistical models and deep learning-based methods. Traditional statistical models, such as the historical average method and the ARIMA model, primarily rely on linear extrapolation from single time series data. In recent years, deep learning-based methods have gradually become mainstream, especially spatiotemporal graph neural network models that combine graph neural networks with temporal convolutional networks or long short-term memory networks. These methods typically treat toll stations as graph nodes, use adjacency matrices to describe the road network topology, aggregate spatial features through graph convolution operations, and combine them with sequence models to extract temporal features, thereby achieving the prediction of road network traffic flow.

[0004] While existing spatiotemporal graph neural network-based technologies have improved prediction accuracy to some extent, several shortcomings remain. First, when constructing the graph structure, most existing methods define static adjacency matrices based solely on physical connections or use simple distance weights, neglecting the dynamic impact of edge attributes such as road segment travel time and traffic flow dispersion on inter-node interactions. In particular, they lack explicit modeling of how real-time weather factors (such as visibility and road slipperiness) alter road segment impedance, resulting in the model's inability to perceive dynamic changes in road network topology weights under adverse weather conditions, leading to a significant drop in prediction accuracy. Second, existing model training typically employs loss functions with equal weights (such as MSE or MAE), treating all toll station nodes equally. This causes the model to tend to optimize data from the majority of smooth-flowing stations, while prediction errors for congested nodes with low operational efficiency and genuine need for management are often masked by averaging, making it difficult to accurately capture the evolution of traffic bottlenecks. Furthermore, existing technologies are mostly open-loop predictions, lacking a feedback mechanism to apply prediction results to actual operational efficiency after management intervention, resulting in the model's inability to adapt to changes in traffic flow distribution after the intervention of management strategies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a spatiotemporal prediction method for highway toll station traffic flow that integrates multi-source information. This method solves the problems of existing highway toll station traffic flow prediction methods failing to effectively integrate the dynamic impact of meteorological conditions on road network topology and lacking specificity in predicting inefficient congestion nodes.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a spatiotemporal prediction method for highway toll station traffic flow that integrates multi-source information, the method comprising the following steps:

[0007] Step S1: Perform multi-source feature extraction and tensor construction: Based on toll station flow records and vehicle origin and destination data, construct node feature tensors representing node flow attributes and basic edge feature tensors representing road network topology.

[0008] Step S2: Perform dynamic edge feature enhancement based on meteorological perception: Calculate the dynamic impedance correction factor that characterizes the traffic obstruction effect using real-time meteorological monitoring data, and use the dynamic impedance correction factor to correct the time delay attribute in the basic edge feature tensor to generate a dynamic edge feature tensor containing real-time environmental impact information.

[0009] Step S3: Construct a spatiotemporal graph attention network model that integrates edge information and make predictions: Construct a spatiotemporal convolution module that includes gated convolutional units and graph attention units that integrate edge information, wherein the graph attention units explicitly introduce the dynamic edge feature tensor into the calculation process of the attention coefficients, and output the short-term traffic flow prediction results of each toll station after processing by the spatiotemporal graph attention network model.

[0010] Step S4: Perform loss function optimization based on operational efficiency feedback: Introduce the toll station operational efficiency evaluation results, construct station weights that are negatively correlated with operational efficiency scores, and construct an efficiency-weighted loss function based on the station weights to optimize model parameters;

[0011] Step S5: Output and Closed-Loop Application of Prediction Results: Generate traffic control instructions based on the short-term traffic flow prediction results and the real-time meteorological monitoring data, and feed back the actual traffic data after executing the traffic control instructions to update the running efficiency score required for the next round of model training.

[0012] Preferably, the construction of the basic edge feature tensor in step S1 specifically includes:

[0013] Based on vehicle records within a historical period, a three-dimensional feature vector is calculated for any two toll station nodes that have a traffic connection.

[0014] The first dimension of the three-dimensional feature vector is the mean travel time, which is used to characterize the physical time delay characteristics of the road segment;

[0015] The second dimension of the three-dimensional feature vector is the standard deviation of travel time, which is used to characterize the dispersion and uncertainty of traffic arrival.

[0016] The third feature of the three-dimensional feature vector is the transfer flow, which is used to characterize the traffic association strength between nodes;

[0017] The first, second, and third features are standardized and then combined to obtain the basic edge feature tensor.

[0018] Preferably, the calculation of the dynamic impedance correction factor in step S2 specifically includes:

[0019] Obtain real-time visibility data and road surface slippage coefficient data at the current time step, and construct a meteorological feature vector;

[0020] The meteorological feature vector is input into a preset nonlinear mapping function, which includes a trainable weight matrix, a bias vector, and a modified linear unit activation function.

[0021] The dynamic impedance correction factor vector is calculated, and the values ​​of the elements in the dynamic impedance correction factor vector increase with the degree of severe weather, based on the baseline state.

[0022] Preferably, the generation of the dynamic edge feature tensor in step S2 specifically includes:

[0023] The impedance factor of the corresponding starting node in the dynamic impedance correction factor vector is multiplied with the mean travel time in the basic edge feature tensor to obtain the meteorological enhanced travel time mean.

[0024] The mean of the weather-enhanced travel time is vector-concatenated with the standard deviation of the travel time in the basic edge feature tensor, which retains the original statistical properties, and the transfer flow.

[0025] The dynamic edge feature tensor is generated based on the concatenated vector and used as the edge attribute input of the spatiotemporal graph attention network model.

[0026] Preferably, the spatiotemporal convolution module in step S3 adopts a stacked structure in which time dimension units, spatial dimension units and time dimension units are connected in sequence;

[0027] The time dimension unit is a gated convolutional unit, which uses a gated linear unit containing residual connections to perform one-dimensional convolution operations on the input features. The gated linear unit extracts features through the main information flow convolutional kernel and controls the retention ratio of information by combining the gated convolutional kernel with the activation function.

[0028] Preferably, step S3, which involves using a graph attention unit that integrates edge information to extract spatial features, specifically includes:

[0029] For the target node and its neighboring nodes, the node features are linearly mapped using a node feature transformation matrix;

[0030] The dynamic edge feature vectors connecting the target node and its neighboring nodes are linearly mapped using an edge feature transformation matrix;

[0031] The mapped target node features, neighbor node features, and dynamic edge feature vectors are concatenated, and the non-normalized attention score is calculated using the attention parameter vector and activation function.

[0032] The non-normalized attention score is normalized to obtain attention coefficients, and the attention coefficients are used to weight and aggregate the features of the neighboring nodes to obtain spatial update features containing time-varying environmental information.

[0033] Preferably, the construction of site weights in step S4 specifically includes:

[0034] Using the data envelopment analysis method, with the number of lanes open at toll stations, average service time, and queue length as input indicators, and the number of vehicles passing through a cross section per unit time as output indicator, the operational efficiency score of each toll station node is calculated.

[0035] A reciprocal mapping relationship is established between the operating efficiency score and the site weight, such that the toll station node with the lower operating efficiency score corresponds to the larger site weight. A constant term is introduced in the calculation process to prevent numerical overflow.

[0036] Preferably, step S4, which involves constructing the efficiency-weighted loss function and performing optimization, specifically includes:

[0037] Construct an efficiency-weighted mean absolute error as the target loss function;

[0038] When calculating the absolute error between the predicted traffic value and the actual traffic value, the error term of each toll station node is weighted using the site weights.

[0039] The efficiency-weighted loss function is minimized using an adaptive optimization algorithm to drive model parameter updates, making the model training process tend to reduce the prediction residuals of high-weight nodes.

[0040] Preferably, generating traffic control instructions in step S5 specifically includes:

[0041] When the predicted traffic flow at the toll station node exceeds the preset threshold or the real-time meteorological monitoring data indicates a high-risk state, a dynamic lane resource allocation instruction and a meteorological enhanced guidance information instruction are generated.

[0042] The lane resource dynamic allocation command is used to set the open combination mode of electronic non-stop toll lanes, manual toll lanes and mixed lanes in the next time granularity;

[0043] The meteorological enhanced guidance information command is used to generate strategy information that includes dynamic speed limit suggestions and path guidance by combining meteorological sensing data.

[0044] Preferably, step S5, which involves outputting the prediction results and applying them in a closed loop, further includes:

[0045] The lane resource dynamic allocation command is executed through the free-flow tolling system, and the weather-enhanced guidance information command is issued through roadside variable information signs and roadside units;

[0046] The actual queue length, actual average service time, and actual cross-sectional traffic volume in the next time window after the collection command is executed;

[0047] The operational efficiency score is recalculated using the collected actual data, and the updated operational efficiency score is used as the input for the next round of model training or online updates.

[0048] This invention provides a spatiotemporal prediction method for traffic flow at highway toll stations that integrates multi-source information. It has the following beneficial effects:

[0049] 1. This invention introduces a dynamic edge feature enhancement mechanism based on meteorological perception, using real-time visibility and road surface slippage coefficient to calculate a dynamic impedance correction factor, thereby correcting the travel time attribute in the basic edge features in real time. This enables the road network topology to dynamically reflect the obstruction effect of severe weather on vehicle traffic, solving the problem that traditional static graph models cannot perceive environmental changes, and improving the robustness and accuracy of traffic flow prediction under complex weather conditions.

[0050] 2. This invention constructs a spatiotemporal graph attention network model that integrates edge information, explicitly incorporating dynamic edge feature tensors containing meteorological enhancement information into the calculation process of attention coefficients. Unlike traditional methods that only focus on node connectivity, this mechanism can comprehensively consider road segment delay, traffic flow dispersion, and weather impact to dynamically allocate neighborhood aggregation weights, thereby uncovering more complex spatiotemporal dependencies in the road network and achieving accurate capture of the spatiotemporal evolution patterns of traffic flow.

[0051] 3. This invention establishes an inverse mapping relationship between the toll station's operational efficiency score and the model training weights by performing loss function optimization based on operational efficiency feedback, thus constructing an efficiency-weighted loss function. This mechanism forces the model to automatically focus its attention on key and difficult toll stations with low operational efficiency and prone to congestion during training, avoiding the model from getting stuck in local optima where it accurately predicts only normally smooth toll stations but fails to predict key congestion nodes, thereby improving the prediction performance for traffic bottleneck nodes.

[0052] 4. This invention establishes a closed-loop application mechanism for prediction result output and control, transforming prediction data into dynamic lane allocation and path guidance strategies, and feeding back the actual traffic efficiency after control to the model training stage to update weights. This closed-loop design of prediction, control, feedback, and optimization enables the prediction model to adapt to changes in the flow field after the intervention of traffic management strategies, achieving deep coupling and continuous adaptive evolution between the prediction algorithm and the actual business scenario. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0054] Figure 2 This is a schematic diagram of the multi-source data preprocessing and feature tensor construction process of the present invention;

[0055] Figure 3 This is a schematic diagram illustrating the construction process of the spatiotemporal graph attention network model that integrates edge information according to the present invention.

[0056] Figure 4 This is a schematic diagram of the process for optimizing the loss function based on operational efficiency feedback according to the present invention. Detailed Implementation

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see the appendix Figure 1 , Figure 1 This is a schematic flowchart of a method according to an embodiment of the present invention. The present invention provides a method for spatiotemporal prediction of traffic flow at highway toll stations by fusing multi-source information, the method comprising the following steps:

[0059] S1. Multi-source feature extraction and tensor construction;

[0060] S2. Dynamic edge feature enhancement based on meteorological perception;

[0061] S3. Construct a spatiotemporal graph attention network model that integrates edge information and perform predictions;

[0062] S4. Loss function optimization based on runtime efficiency feedback;

[0063] S5. Prediction result output and closed-loop application.

[0064] In step S1, multi-source feature extraction and tensor construction are performed. This step obtains raw data from highway toll stations and monitoring equipment deployed along the route, and constructs node feature tensors and basic edge feature tensors. Specifically, based on the highway network toll collection records, the inbound and outbound traffic volumes of each toll station across the entire network are statistically analyzed according to a preset time granularity, forming node feature tensors containing historical time-series information. To eliminate dimensional differences, the node feature tensors are standardized.

[0065] Simultaneously, based on historical vehicle origin-destination data, the statistical characteristics of travel time and traffic transfer characteristics between any two toll station nodes are calculated. The extracted basic edge features include the mean travel time characterizing road segment delay characteristics, the standard deviation of travel time characterizing traffic flow dispersion, and the transferred traffic flow characterizing the strength of traffic association between nodes. Based on these features, a basic edge feature tensor is constructed and standardized to serve as benchmark data for describing the road network topology.

[0066] In step S2, dynamic edge feature enhancement based on meteorological sensing is performed. This step uses real-time meteorological monitoring data to calculate a dynamic impedance correction factor to correct the time delay attribute in the basic edge features. Real-time visibility and road surface slippage coefficient are obtained at the current time step, and the dynamic impedance correction factor for the region where each node is located is calculated using a nonlinear mapping function.

[0067] The mean travel time in the basic edge features is weighted and corrected using the calculated impedance correction factor to obtain the meteorological enhanced mean travel time. The corrected mean travel time is then concatenated with the original standard deviation of travel time and the transfer flow to generate a dynamic edge feature tensor containing real-time environmental impact information.

[0068] In step S3, a spatiotemporal graph attention network model incorporating edge information is constructed and used for prediction. A spatiotemporal convolutional module containing gated convolutional units and graph attention units incorporating edge information is constructed to extract the spatiotemporal features of traffic flow. In the time dimension, gated linear units are used to perform one-dimensional convolution operations on the input time series to capture long-term and short-term time dependencies.

[0069] In the spatial dimension, a graph attention mechanism that integrates edge information is employed. The dynamic edge feature tensor generated in step S2 is explicitly incorporated into the calculation of the attention coefficients. Using the node feature transformation matrix and the edge feature transformation matrix, the node features and dynamic edge features are mapped to the same dimension and concatenated. Based on the concatenated vector, the attention score between a node and its neighboring nodes is calculated and normalized to obtain the attention coefficients.

[0070] This attention coefficient is used to weight and aggregate the features of neighboring nodes, thereby enabling spatial feature extraction based on real-time traffic conditions and environmental factors. After processing by a multi-layer spatiotemporal convolutional module, the short-term traffic flow prediction results for each toll station are output through a fully connected layer.

[0071] In step S4, loss function optimization based on operational efficiency feedback is performed. This step incorporates the operational efficiency evaluation results of toll stations to weight the loss function used in model training. Operational efficiency scores for each toll station are obtained, and station weights are constructed that are inversely proportional to these scores. Toll stations with lower operational efficiency and a higher risk of congestion are assigned higher weights, while those with higher operational efficiency are assigned lower weights.

[0072] An efficiency-weighted average absolute error loss function is constructed. When calculating the error between the predicted and actual values, the error term is weighted using the aforementioned site weights. By minimizing this weighted loss function to update the model parameters, the model automatically focuses on traffic prediction for key and challenging sites during training, improving the prediction accuracy for frequently congested nodes.

[0073] In step S5, the prediction results are output and applied in a closed loop. The predicted short-term traffic flow results are output to the traffic control platform. Based on the prediction results and real-time meteorological data, the traffic control platform comprehensively analyzes and generates a dynamic lane allocation strategy and route guidance information. The free-flow tolling system dynamically switches lane functions, and guidance information is disseminated through roadside variable message signs.

[0074] Simultaneously, actual traffic data after the implementation of the control strategy is collected to update the operational efficiency score for the next time window, and the updated score is fed back to step S4 as the weight input for the next round of model training, forming a closed loop of prediction, control and optimization.

[0075] Please see the appendix Figure 2 , Figure 2 This is a schematic diagram of a multi-source data preprocessing and feature tensor construction process according to an embodiment of the present invention. In a specific embodiment of the present invention, the multi-source feature extraction and tensor construction process in step S1 specifically includes a node feature tensor construction step S101 and a basic edge feature tensor extraction step S102.

[0076] In step S101, a node feature tensor is constructed. Based on the transaction records of the highway network toll collection system, the timestamps of vehicles passing through toll stations, toll station numbers, and entrance / exit status data are obtained. The time axis is divided into continuous preset time granularities. For each time granularity, the statistics for each toll station node in the entire road network are calculated separately. The inbound and outbound traffic flows. The time step of the input model is set to... The total number of toll station nodes is The number of feature channels for each node is (In this embodiment, (These correspond to inbound and outbound traffic characteristics, respectively). Based on the above statistical data, a dimension is constructed as follows: Node feature tensor .

[0077] To eliminate the differences in traffic volume between different toll stations, the node feature tensor... Standardization is performed. In this embodiment, the standard score standardization method is used, that is, the original traffic statistics are mapped to values ​​that conform to a standard normal distribution using the mean and standard deviation of historical traffic data, thereby completing the construction of the node feature tensor.

[0078] In step S102, the basic edge feature tensor is extracted. Based on the origin and destination data of historical vehicles, deep spatiotemporal transfer features between toll station nodes are mined, and weighted graph structure edge features are constructed. The road network graph structure is defined. ,in For the set of toll station nodes, Let be the set of edges. For any two toll station nodes that are connected by traffic, ... and Based on all slave nodes within the historical period Drive in and from the node For the records of departing vehicles, calculate the following three-dimensional feature vectors:

[0079] The first feature is the mean travel time. It refers to the vehicles moving from the upstream node within a historical cycle. Reaching downstream nodes The arithmetic mean of the required travel time is used to characterize the physical time delay characteristics of a road segment.

[0080] The second feature is the standard deviation of travel time. It refers to the vehicles moving from the upstream node within a historical cycle. Reaching downstream nodes The statistical standard deviation of travel time is used to characterize the dispersion and uncertainty of traffic arrival.

[0081] The third dimension is OD transfer flow. It is a node within a historical cycle. Transfer to node The total number of vehicles is used to characterize the strength of traffic associations between nodes.

[0082] Based on the above three-dimensional features, a dimension of is constructed. Basic edge feature tensor Among them, corresponding to node pairs Basic edge feature vector Represented as:

[0083] ;

[0084] In the formula, For toll station nodes Pointing to toll station node The basic edge feature vector; For vehicles passing through toll station nodes within the historical period Arrival at the toll station node The average travel time is used to characterize the physical time delay characteristics; For vehicles passing through toll station nodes within the historical period Arrival at the toll station node The standard deviation of travel time is used to characterize the dispersion of traffic flow. For the historical period from the toll station node Transfer to toll station node The total number of vehicles is used to characterize the strength of traffic association.

[0085] For node pairs without direct traffic connections, their feature vectors are set to zero. To make the basic edge feature tensor numerically comparable to the node feature tensor, each feature dimension in the basic edge feature tensor is standardized using standard scores, that is, the feature values ​​are mapped using the mean and standard deviation of all edges in the entire road network. This basic edge feature tensor is used for dynamic correction in subsequent steps using meteorological data.

[0086] In one specific embodiment of the present invention, the dynamic edge feature enhancement process based on meteorological perception in step S2 specifically includes a meteorological impedance factor calculation step S201, a time delay feature dynamic correction step S202, and a dynamic edge feature tensor generation step S203.

[0087] In step S201, the meteorological impedance factor is calculated. The current time step is obtained from the integrated meteorological monitoring equipment deployed along the highway. Real-time environmental perception data, including visibility. and road surface slip coefficient A nonlinear mapping function is constructed to transform real-time environmental perception data into quantified values ​​of traffic impedance. A dynamic impedance correction factor vector is defined. The dimension of this vector is The calculation formula is as follows:

[0088] ;

[0089] In the formula, For the current time step The dynamic impedance correction factor vector of the entire road network, with dimension . ( (Total number of toll station nodes). For dimension The all-1 vector represents the reference impedance state under conditions without adverse weather conditions; To correct the activation function of the linear unit, so as to ensure that the calculated impedance increment is non-negative; This is a trainable weight matrix for the meteorological sensing mapping layer, used to map meteorological features to impedance space, with dimensions of [missing information]. ; The current time step after normalization The meteorological feature vector, containing visibility and road surface slippage coefficient, has a dimension of ; The bias vector of the meteorological sensing mapping layer has a dimension of ; This represents matrix multiplication.

[0090] In step S202, dynamic correction of the time delay characteristics is performed. The dynamic impedance correction factor calculated in step S201 is used. The basic edge feature tensor extracted in step S102 Average travel time Make corrections. For each edge in the road network, use the starting node of that edge. Corresponding impedance factor The average travel time was adjusted to obtain the average travel time for weather-enhanced conditions. The calculation formula is as follows:

[0091] ;

[0092] In the formula, The average historical travel time in the basic edge features; For vectors The middle corresponds to the node The amount; For the current time step The toll station node after weather enhancement correction Toll station node The average travel time; This indicates a scalar multiplication operation.

[0093] In step S203, a dynamic edge feature tensor is generated. The weather-enhanced travel time mean calculated in step S202 is then used. The standard deviation of the travel time in the basic edge feature tensor extracted in step S102 and OD transfer flow Perform feature stitching. For each edge in the road network... Its dynamic edge feature vector Represented as:

[0094] ;

[0095] In the formula, For the current time step Exit toll station node Pointing to toll station node The dynamic edge feature vector; This is a vector concatenation operation function used to combine features of different dimensions into a new high-dimensional vector; For the current time step The average travel time after weather enhancement correction; The standard deviation of travel time in the basic edge features is used to maintain the original statistical properties. The OD transfer flow in the basic edge features is used to maintain the original statistical properties.

[0096] The current time step is generated by combining the dynamic edge feature vectors calculated from all edges of the entire road network. Dynamic edge feature tensor The dimension of this tensor is This dynamic edge feature tensor serves as graph structure data containing real-time environmental awareness information, and is used as input to the subsequent spatiotemporal graph attention network model.

[0097] Please see the appendix Figure 3 , Figure 3 This is a schematic diagram of the construction process of a spatiotemporal graph attention network model that incorporates edge information according to an embodiment of the present invention. In a specific embodiment of the present invention, the process of constructing a spatiotemporal graph attention network model that incorporates edge information and making predictions in step S3 specifically includes a spatiotemporal convolution module construction step S301, a temporal feature extraction step S302, a spatial feature extraction step S303, and an output layer prediction step S304.

[0098] In step S301, a spatiotemporal convolutional module is constructed. This module adopts a time-space-time stacked structure. Each spatiotemporal convolutional module consists of a front-end gated convolutional unit, a middle graph attention unit that fuses side information, and a rear-end gated convolutional unit connected sequentially. The front-end gated convolutional unit is used to map the input data to a high-dimensional latent space and capture short-term temporal dependencies. The graph attention unit that fuses side information is used to aggregate neighborhood information affected by meteorological factors in the feature space. The rear-end gated convolutional unit is used for feature fusion and dimensionality adjustment.

[0099] In step S302, gated convolutional units are used to extract time features. For the traffic flow time series data of each toll station node, a one-dimensional convolution operation is performed using gated linear units with residual connections. Let the input feature tensor be... Output feature tensor The calculation formula is as follows:

[0100] ;

[0101] In the formula, The feature tensor output by the gated convolutional unit; The feature tensor input to the gated convolutional unit; This is a one-dimensional convolution operation along the time dimension; The trainable weight matrix of the main information flow convolution kernel is used to extract linear time features; The trainable bias vector of the main information flow; The Hadamard product is the element-wise multiplication of a matrix or vector. This is the Sigmoid activation function, used to map the output value to the (0, 1) interval to control the proportion of information retained; This is the trainable weight matrix of the convolutional kernel for the gating mechanism, used to determine the gating coefficients; This is the trainable bias vector for the gating mechanism.

[0102] In step S303, spatial features are extracted using graph attention units that fuse edge information. The dynamic edge feature tensor generated in step S203 is then processed. The process of explicitly introducing attention weights is described. For any time step... Set nodes The feature vector is Its neighboring nodes The feature vector is Connecting nodes and The dynamic edge feature vector is Compute nodes with neighboring nodes The nonnormalized attention score between The calculation formula is as follows:

[0103] ;

[0104] In the formula, For nodes Its neighboring nodes The nonnormalized attention score between them; This is a modified linear unit activation function with leakage, used to introduce nonlinearity and prevent neuron death; This is a trainable parameter vector for the attention mechanism, used to map high-dimensional features to scalar scores; This is a transpose operation; This is a vector concatenation operation; This is a trainable weight matrix used to perform linear transformations on node features; For the current time step target node The input feature vector; For the current time step Neighbor nodes The input feature vector; This is a trainable weight matrix used to perform linear transformations on edge features to match the node feature dimensions; For the current time step Connecting nodes With nodes The dynamic edge feature vector; The index of the target node; This is the index of the neighboring node.

[0105] Based on nonnormalized attention score Calculate the normalized attention coefficient :

[0106] ;

[0107] In the formula, The normalized attention coefficients represent the neighboring nodes. For the target node Spatial influence weight; For the natural constant An exponential function with base 0; For nodes With any node in its neighborhood The nonnormalized attention score between them; For nodes The set of first-order neighbor nodes; For nodes A set of itself represents a self-loop connection that includes the node itself when calculating attention; The node index in the neighborhood set; This is the union operator for sets.

[0108] Using attention coefficient Weighted aggregation of node features within the neighborhood is performed to calculate the node Spatial update features :

[0109]

[0110] In the formula, For time step Time node The output feature vector after spatial feature aggregation; To modify the activation function of the linear unit; This is a trainable weight matrix used to perform linear transformations on node features.

[0111] In step S304, output layer prediction is performed. After processing by multiple spatiotemporal convolutional modules, the temporal dimension is aggregated through a gated convolutional layer with a time span equal to the sequence length. Subsequently, a fully connected layer is used to map the feature space back to the target physical space. Let the input features of the fully connected layer be... The weight matrix is The bias vector is Calculate the final prediction result. :

[0112] ;

[0113] In the formula, This is the tensor for the final output short-term traffic flow prediction values; The input feature tensor of the fully connected layer; This is the matrix multiplication operator; This is the trainable weight matrix for the fully connected layer; This is the bias vector for the fully connected layer.

[0114] Please see the appendix Figure 4 , Figure 4 This is a schematic diagram of a loss function optimization process based on operational efficiency feedback according to an embodiment of the present invention. In a specific embodiment of the present invention, the loss function optimization process based on operational efficiency feedback in step S4 specifically includes an operational efficiency score calculation step S401, a site weight construction step S402, and an efficiency-weighted loss function calculation and optimization step S403.

[0115] In step S401, the operational efficiency score of each toll station node is calculated. An independent highway toll station operational efficiency evaluation model is used to quantitatively evaluate the traffic service level of each toll station node within the previous time window. In this embodiment, the data envelopment analysis method is used. The number of lanes open at the toll station, the average service time, and the queue length are selected as input indicators, and the number of vehicles passing through the cross-section per unit time is selected as the output indicator.

[0116] By calculating the relative effectiveness of the decision-making units, the results for each toll station node are obtained. Operational efficiency score The score The value range is (0, 1]. The closer the value is to 1, the higher the input-output ratio of the toll station, and the more efficient the operation is in a non-congested state; the lower the value, the more congested or inefficient the toll station is in a congested state.

[0117] In step S402, site weights are constructed. An inverse mapping relationship is established between operational efficiency scores and model training weights, so that nodes with lower operational efficiency receive higher attention during model training. For tollbooth nodes... Calculate its site weight The calculation formula is as follows:

[0118] ;

[0119] In the formula, For the first The site weight of each toll station node; The running efficiency score calculated in step S401; To prevent constants with a denominator of zero; This serves as the index identifier for tollbooth nodes. Through this calculation, the weight of congested stations is amplified, while the weight of uncongested stations is relatively smaller.

[0120] In step S403, the efficiency-weighted loss function is calculated and optimization is performed. An efficiency-weighted mean absolute error is constructed as the target loss function, replacing the traditional equal-weighted loss function. Let the batch size for model training be... The predicted target time step is The total number of toll station nodes is For the first The sample, the first The prediction step, the first For each toll station, the model predicts the traffic flow as follows: The corresponding actual traffic value Efficiency-weighted loss function The calculation formula is as follows:

[0121]

[0122] In the formula, The calculated efficiency-weighted loss function value; This is the set of all trainable weights and bias parameters in the spatiotemporal graph attention network model; This refers to the batch size of samples used during model training. This represents the total number of toll station nodes in the entire road network. The total number of predicted target time steps output by the model; This is the index variable for the samples in the current training batch, with a value ranging from 1 to... ; The index variable for predicting time steps, with values ​​ranging from 1 to... ; This is the index variable for the toll station node, with a value range of 1 to... ; For the first The site weight of each toll station node; For the first The sample at the th The prediction time step is for the first Predicted traffic volume for each toll station node; For the first The sample at the th The prediction time step is for the first The actual traffic volume of each toll station node; This is the absolute value operator.

[0123] During model training, this loss function is minimized. This drives parameter updates, causing the model to prioritize optimizing the prediction error of high-weight (i.e., low-efficiency) nodes.

[0124] In one specific embodiment of the present invention, the output and closed-loop application process of the prediction result in step S5 specifically includes a decision instruction generation step S501, a dynamic control execution step S502, and a feedback data transmission step S503.

[0125] In step S501, a decision instruction is generated. The traffic control platform receives the short-term traffic flow forecast output in step S3. (Includes inbound and outbound traffic forecasts for each toll station node), combined with real-time meteorological sensing data (visibility). Road surface wet slip coefficient (and the performance score for the previous time window) A comprehensive assessment is conducted. When the predicted traffic flow at a toll station exceeds the preset congestion threshold, or when real-time weather data indicates a risk of low visibility or slippery road surfaces, the platform generates two types of control instructions:

[0126] The first category is dynamic lane resource allocation instructions. For toll station nodes, the instructions specify the number and combination mode of ETC-only lanes, MTC manual lanes, and mixed lanes to be opened within the next time granularity. For example, if a surge in exit traffic is predicted and current operating efficiency is low, the instructions will set some mixed lanes to switch to ETC-only mode to improve traffic flow.

[0127] The second category is weather-enhanced guidance information instructions. For the relevant road segment, the instructions generate text or graphic information including speed limit suggestions and route guidance. For example, combining weather sensing data, a composite guidance strategy can be generated suggesting a reduced speed limit of X km / h due to predicted high traffic volume and slippery road conditions.

[0128] In step S502, dynamic control is performed. The free-flow toll collection system receives dynamic lane resource allocation instructions and switches lane functions in real time by controlling lane indicator screens, electronic barriers, and toll collection software logic. Variable message signs and roadside units deployed along the route receive weather-enhanced guidance information instructions and issue dynamic speed limits and route guidance information to passing vehicles.

[0129] In step S503, feedback data is transmitted back. In the next time window after executing the aforementioned control instructions, actual traffic data for each toll station node is collected, including actual queue length, actual average service time, and actual cross-sectional traffic volume. This actual data is then input into the operational efficiency evaluation model in step S4 to calculate the latest operational efficiency score. The latest runtime efficiency score is directly used to calculate site weights during the next round of model training or online updates. The basis for this is that by feeding back the actual operational effects after control measures into the weight calculation of the model's loss function, the predictive model achieves adaptive learning of the effectiveness of traffic management strategies.

[0130] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A spatiotemporal prediction method for traffic flow at highway toll stations that integrates multi-source information, characterized in that, The method includes the following steps: Step S1: Perform multi-source feature extraction and tensor construction: Based on toll station flow records and vehicle origin and destination data, construct node feature tensors representing node flow attributes and basic edge feature tensors representing road network topology. Step S2: Perform dynamic edge feature enhancement based on meteorological perception: Calculate the dynamic impedance correction factor that characterizes the traffic obstruction effect using real-time meteorological monitoring data, and use the dynamic impedance correction factor to correct the time delay attribute in the basic edge feature tensor to generate a dynamic edge feature tensor containing real-time environmental impact information. Step S3: Construct a spatiotemporal graph attention network model that integrates edge information and make predictions: Construct a spatiotemporal convolution module that includes gated convolutional units and graph attention units that integrate edge information, wherein the graph attention units explicitly introduce the dynamic edge feature tensor into the calculation process of the attention coefficients, and output the short-term traffic flow prediction results of each toll station after processing by the spatiotemporal graph attention network model. Step S4: Perform loss function optimization based on operational efficiency feedback: Introduce the toll station operational efficiency evaluation results, construct station weights that are negatively correlated with operational efficiency scores, and construct an efficiency-weighted loss function based on the station weights to optimize model parameters; Step S5: Output and Closed-Loop Application of Prediction Results: Generate traffic control instructions based on the short-term traffic flow prediction results and the real-time meteorological monitoring data, and feed back the actual traffic data after executing the traffic control instructions to update the running efficiency score required for the next round of model training.

2. The spatiotemporal prediction method for highway toll station traffic flow integrating multi-source information according to claim 1, characterized in that, The construction of the basic edge feature tensor in step S1 specifically includes: Based on vehicle records within a historical period, a three-dimensional feature vector is calculated for any two toll station nodes that have a traffic connection. The first dimension of the three-dimensional feature vector is the mean travel time, which is used to characterize the physical time delay characteristics of the road segment; The second dimension of the three-dimensional feature vector is the standard deviation of travel time, which is used to characterize the dispersion and uncertainty of traffic arrival. The third feature of the three-dimensional feature vector is the transfer flow, which is used to characterize the traffic association strength between nodes; The first, second, and third features are standardized and then combined to obtain the basic edge feature tensor.

3. The spatiotemporal prediction method for highway toll station traffic flow integrating multi-source information according to claim 2, characterized in that, The calculation of the dynamic impedance correction factor in step S2 specifically includes: Obtain real-time visibility data and road surface slippage coefficient data at the current time step, and construct a meteorological feature vector; The meteorological feature vector is input into a preset nonlinear mapping function, which includes a trainable weight matrix, a bias vector, and a modified linear unit activation function. The dynamic impedance correction factor vector is calculated, and the values ​​of the elements in the dynamic impedance correction factor vector increase with the degree of severe weather, based on the baseline state.

4. The spatiotemporal prediction method for highway toll station traffic flow integrating multi-source information according to claim 3, characterized in that, The specific steps in step S2 to generate the dynamic edge feature tensor include: The impedance factor of the corresponding starting node in the dynamic impedance correction factor vector is multiplied with the mean travel time in the basic edge feature tensor to obtain the meteorological enhanced travel time mean. The mean of the weather-enhanced travel time is vector-concatenated with the standard deviation of the travel time in the basic edge feature tensor, which retains the original statistical properties, and the transfer flow. The dynamic edge feature tensor is generated based on the concatenated vector and used as the edge attribute input of the spatiotemporal graph attention network model.

5. The method for spatiotemporal prediction of highway toll station traffic flow by integrating multi-source information according to claim 1, characterized in that, The spatiotemporal convolution module in step S3 adopts a stacked structure in which time dimension units, spatial dimension units and time dimension units are connected in sequence. The time dimension unit is a gated convolutional unit, which uses a gated linear unit containing residual connections to perform one-dimensional convolution operations on the input features. The gated linear unit extracts features through the main information flow convolutional kernel and controls the retention ratio of information by combining the gated convolutional kernel with the activation function.

6. The method for spatiotemporal prediction of highway toll station traffic flow by fusing multi-source information as described in claim 1, characterized in that, The extraction of spatial features using the graph attention unit that fuses edge information in step S3 specifically includes: For the target node and its neighboring nodes, the node features are linearly mapped using a node feature transformation matrix; The dynamic edge feature vectors connecting the target node and its neighboring nodes are linearly mapped using an edge feature transformation matrix; The mapped target node features, neighbor node features, and dynamic edge feature vectors are concatenated, and the non-normalized attention score is calculated using the attention parameter vector and activation function. The non-normalized attention score is normalized to obtain attention coefficients, and the attention coefficients are used to weight and aggregate the features of the neighboring nodes to obtain spatial update features containing time-varying environmental information.

7. The method for spatiotemporal prediction of highway toll station traffic flow by fusing multi-source information as described in claim 1, characterized in that, The specific steps in step S4 of constructing site weights include: Using the data envelopment analysis method, with the number of lanes open at toll stations, average service time, and queue length as input indicators, and the number of vehicles passing through a cross section per unit time as output indicator, the operational efficiency score of each toll station node is calculated. A reciprocal mapping relationship is established between the operating efficiency score and the site weight, such that the toll station node with the lower operating efficiency score corresponds to the larger site weight. A constant term is introduced in the calculation process to prevent numerical overflow.

8. The method for spatiotemporal prediction of highway toll station traffic flow by integrating multi-source information according to claim 7, characterized in that, The specific steps in step S4, including constructing the efficiency-weighted loss function and performing optimization, include: Construct an efficiency-weighted mean absolute error as the target loss function; When calculating the absolute error between the predicted traffic value and the actual traffic value, the error term of each toll station node is weighted using the site weights. The efficiency-weighted loss function is minimized using an adaptive optimization algorithm to drive model parameter updates, making the model training process tend to reduce the prediction residuals of high-weight nodes.

9. The method for spatiotemporal prediction of highway toll station traffic flow by fusing multi-source information as described in claim 1, characterized in that, The specific steps in step S5 of generating traffic control instructions include: When the predicted traffic flow at the toll station node exceeds the preset threshold or the real-time meteorological monitoring data indicates a high-risk state, a dynamic lane resource allocation instruction and a meteorological enhanced guidance information instruction are generated. The lane resource dynamic allocation command is used to set the open combination mode of electronic non-stop toll lanes, manual toll lanes and mixed lanes in the next time granularity; The meteorological enhanced guidance information command is used to generate strategy information that includes dynamic speed limit suggestions and path guidance by combining meteorological sensing data.

10. The method for spatiotemporal prediction of highway toll station traffic flow by fusing multi-source information according to claim 9, characterized in that, Step S5, which involves outputting the prediction results and applying the closed-loop logic, also includes: The lane resource dynamic allocation command is executed through the free-flow tolling system, and the weather-enhanced guidance information command is issued through roadside variable information signs and roadside units; The actual queue length, actual average service time, and actual cross-sectional traffic volume in the next time window after the collection command is executed; The operational efficiency score is recalculated using the collected actual data, and the updated operational efficiency score is used as the input for the next round of model training or online updates.

Images