Traffic speed prediction method based on dynamic fundamental diagram and spatial-temporal characteristics

Abstract

The invention discloses a traffic speed prediction method based on a dynamic fundamental diagram and spatial-temporal characteristics, and relates to the technical field of intelligent traffic systems, and the method comprises the steps: selecting a space region and a time range, obtaining structural data through a traffic monitoring system, and dividing a training set and a test set; a neural network model based on graph information is constructed, and the neural network model comprises a customized NF traffic flow model calibration module used for obtaining traffic flow parameters of different months and spatial positions, a spatial feature extraction module and a time feature extraction module used for extracting spatial and temporal features of traffic flow speed and flow data; and constructing a composite loss function, training the neural network model based on the graph information, and completing prediction of the traffic speed. Therefore, by adopting the traffic speed prediction method based on the dynamic basic diagram and the spatial-temporal characteristics, the result is ensured to conform to the traffic engineering theory, so that the traffic speed prediction problem under the spatial-temporal heterogeneous condition is solved, and meanwhile, the prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation systems, and in particular to a traffic speed prediction method based on dynamic fundamental diagrams and spatio-temporal characteristics. Background Art

[0002] Traffic speed is an important indicator characterizing traffic conditions. Accurate prediction of traffic speed not only enables traffic management agencies to effectively formulate strategies to alleviate congestion, but also provides drivers with up-to-date information to make informed route decisions, thereby improving the overall efficiency and safety of the traffic system.

[0003] Currently, traffic speed can be collected through various types of sensors, and cross-validation is crucial for maintaining the consistency and reliability of data from different measurement sources. Therefore, it is necessary to use data highly correlated with traffic speed to improve the estimation in the prediction task. Traffic flow is a key parameter closely related to speed when characterizing traffic conditions, and a large number of studies have demonstrated the strong correlation between traffic flow and speed. This interaction can be effectively illustrated by the fundamental diagram of traffic flow. The fundamental diagram describes the interrelationships among three basic traffic indicators: speed, density, and flow, and summarizes the conservation principle, where flow is always the product of density and speed. Although many studies have used traffic flow as an input feature in traffic speed prediction models, research on using the fundamental diagram relationship as prior physical information for predicting traffic speed is still relatively limited.

[0004] In addition, the static nature of the fundamental diagram poses certain challenges to capturing the dynamic characteristics of traffic conditions. Traditional fundamental diagram models have significant spatio-temporal heterogeneity and are affected by multiple factors such as traffic dynamics, seasonal fluctuations, holiday arrangements, weather conditions, and geographical features. Static models are difficult to adapt to dynamic changes. Therefore, in the traffic speed prediction task, it is not only necessary to adjust the fundamental diagram relationship at different locations, but also to consider monthly variability, which is a huge challenge. Summary of the Invention

[0005] The purpose of the present invention is to provide a traffic speed prediction method based on dynamic fundamental diagrams and spatio-temporal characteristics, which can ensure that the results conform to traffic engineering theory, solve the traffic speed prediction problem under spatio-temporal heterogeneous conditions, and improve the prediction accuracy at the same time.

[0006] To achieve the above purpose, the present invention provides a traffic speed prediction method based on dynamic fundamental diagrams and spatio-temporal characteristics, including the following steps:

[0007] S1. Select a spatial region and a time range, and obtain structured data through a traffic monitoring system, including time stamps, traffic flow, traffic speed, and sensor location information, and divide the obtained structured data into a training set and a test set;

[0008] S2. Construct a neural network model based on graph information, including an NF traffic flow model calibration module, a spatial feature extraction module, and a temporal feature extraction module;

[0009] Among them, the NF traffic flow model calibration module customizes corresponding congestion density, free flow speed, and traffic flow characteristic parameters for different months and spatial locations to obtain a calibrated NF traffic flow model to adapt to the actual situation of traffic flow;

[0010] The spatial feature extraction module uses a self-organizing mapping neural network, combines DTW similarity measurement, constructs hyperedges through clustering analysis, and uses a hypergraph neural network to extract spatial correlation;

[0011] The temporal feature extraction module uses a gated recurrent unit to extract temporal dynamic features;

[0012] S3. Weight the loss between the predicted speed and the observed value and the loss of the calibrated NF traffic flow model to construct a composite loss function, and use the training set to train the neural network model based on graph information, thereby completing the prediction of traffic speed.

[0013] In a possible implementation, in step S2, the NF traffic flow model calibration module includes calibrating the congestion density, free flow speed, and traffic flow characteristic parameters in the NF traffic flow model by minimizing the error of the NF traffic flow model for different months and spatial locations, and embedding the calibrated parameters into the neural network;

[0014] Among them, the expression for minimizing the error of the NF traffic flow model is:

[0015]

[0016] In the formula, MSE NF is the error of the NF traffic flow model, L represents the set of road segments, t represents the time interval, |L| represents the number of links, T (m) represents the total duration of the speed observation time series in the m-th month, are respectively the observed traffic flow and observed speed of road segment l at time interval t in the m-th month, is the congestion density of road segment l in the m-th month, represents the free flow speed of road segment l in the m-th month, λ (m) (l) represents the traffic flow characteristic parameter of road segment l in the m-th month.

[0017] In a possible implementation, in step S2, the spatial feature extraction module includes training a self-organizing mapping neural network using speed observation values and flow observation values respectively to obtain a set of weights for each mapping neuron, and then combining the DTW distance to classify the links between speed observation values or flow observation values projected onto the same mapping neuron into the same category. Hyperedges are connected within the same category of links, and an adjacency matrix of the hyperedges is constructed. The expression is as follows:

[0018]

[0019] In the formula, H(i,j) represents the element in the i-th row and j-th column of the hyperedge adjacency matrix H, d ij represents the DTW distance from node i to hyperedge j, and D j represents the average DTW distance of the hyperedges.

[0020] In a possible implementation, in step S3, the expression of the composite loss function is:

[0021] Loss = α × MSE v + β × MSE FD ;

[0022] where

[0023]

[0024] in the formula, Loss is the composite loss function, MSE v is the mean square error between the predicted speed and the observed speed, MSE FD is the loss of the calibrated NF traffic flow model, α and β are the weights of the corresponding losses, v lt is the predicted speed of section l within time period t, are the observed flow and observed speed of section l at time interval t of the m-th month respectively, L represents the set of sections, T (m) represents the total duration of the speed observation time series of the m-th month, λ (m) (l) are the congestion density, free flow speed, and traffic flow characteristic parameters of section l of the m-th month respectively.

[0025] In a possible implementation, before calculating the mean square error MSE v between the predicted speed and the observed speed, it is necessary to perform min-max normalization on the predicted speed v lt and the observed speed so that the value range is in the interval [-1, 1]; when calculating the loss of the calibrated NF traffic flow model, inverse normalization is required to ensure that the numerical scales of different links are consistent.

[0026] In a possible implementation, in step S3, the training process includes using a hypergraph neural network to extract the spatial features of the traffic speed and flow data of the fully connected layer, using a gated recurrent unit to capture the temporal features, and then obtaining the spatio-temporal feature representation that fuses the traffic speed and flow through projection by the fully connected layer, so as to obtain the predicted speed of the traffic flow.

[0027] Therefore, the present invention adopts the above traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features, and has the following technical effects:

[0028] (1) By calibrating the NF traffic flow model through data driving, the present invention can capture the differences in monthly and regional traffic characteristics, has strong dynamic adaptability, and at the same time introduces physical information loss to ensure that the prediction results conform to traffic flow theory and avoid the "black box" problem of pure data-driven models.

[0029] (2) By fusing the physical model and the data-driven method, the present invention can reduce the prediction error of the model, which is better than traditional models, and effectively captures complex spatio-temporal dependencies by combining the hypergraph neural network and the gated recurrent unit, and is applicable to large-scale traffic networks.

[0030] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0031] Figure 1 is a schematic diagram of the overall framework of the FDINN model in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features;

[0032] Figure 2 is a schematic diagram of a small traffic network in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features;

[0033] Figure 3 is a representation diagram of hyperedges and hypergraph adjacency matrices in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features;

[0034] Figure 4 is the NF model calibration diagram in February in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features, where a is the calibration curve of speed and density, b is the calibration curve of flow and density, and c is the calibration curve of speed and flow;

[0035] Figure 5 is the training result of the FDINN model in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features, where a is the Loss loss, b is the MSE v loss, c is the MSE FD loss;

[0036] Figure 6This is the prediction result of the FDINN model in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features. Here, a represents the prediction range of 5 minutes, b represents the prediction range of 10 minutes, c represents the prediction range of 15 minutes, and d represents the prediction range of 20 minutes;

[0037] Figure 7 This is the speed observation and prediction heat map when the prediction range is 20 minutes in the embodiment of the traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features. Here, a is the speed observation heat map at 8:00 am on April 9, 2018, b is the speed prediction heat map at 8:00 am on April 9, 2018, c is the speed observation heat map at 8:00 am on February 10, 2018, and b is the speed prediction heat map at 8:00 am on February 10, 2018. Detailed implementation mode

[0038] The present invention can be more specifically explained through the following embodiments. The purpose of disclosing the present invention is to protect all changes and improvements within the scope of the present invention. The present invention is not limited to the following embodiments.

[0039] Please refer to Figure 1 , the present invention provides a traffic speed prediction method based on the dynamic fundamental diagram and spatio-temporal features, including two stages: one is the calibration stage of the non-linear car-following (NF) traffic flow model, and the other is the speed prediction stage of the traffic flow.

[0040] S1. In the specific implementation process of this embodiment, the traffic speed and flow observation data in Area 7 of the PeMS traffic data of the California highway network are used as the research object. These data contain a large number of data points, which are captured by sensors along the Los Angeles highway at 5-minute intervals in 2018, including time stamps, traffic flow, traffic speed, and sensor locations.

[0041] In this embodiment, 272 sensors on the Southern California highways, namely Highways 5, 10, 405, 210, and 110, are selected, covering a total of 105,106 time slots, which are divided into 12 monthly subsets. The data of three randomly selected observation days each month are used as the test set, and the rest are used as the training set.

[0042] S2. Aiming at the problem that the traditional fundamental diagram is difficult to adapt to dynamic changes, in this embodiment, the dynamically calibrated NF traffic flow model is introduced as physical prior information and embedded into the prediction network framework to achieve personalized parameter adaptation for different months and different road sections. On this basis, this embodiment constructs a neural network model based on fundamental diagram information (FDINN), including an NF traffic flow model calibration module, a spatial feature extraction module, and a temporal feature extraction module.

[0043] By analyzing the data selected in this embodiment, it is found that the congestion density varies greatly between different locations. The congestion density is higher near the city center, which may be because the population and traffic demand in the city center are higher, and serious traffic congestion is more likely to occur. On the contrary, in geographical areas far from the city center, the population and traffic demand are lower, and there is less traffic congestion. The congestion density also varies dynamically in different months. Therefore, this embodiment designs an NF traffic flow model calibration module, which customizes the corresponding congestion density for different months and spatial locations. Free flow speed and the traffic flow characteristic parameter λ (m) (l) to describe the traffic states at different locations in each month, so as to obtain the calibrated NF traffic flow model to adapt to the actual situation of the traffic flow. Please refer to Figure 4 . In this embodiment, the observed data of February, May, August, and November from sensor 715929 on Highway 5 are selected. For different months and spatial locations, the error of the NF traffic flow model is minimized by the least squares method, and the congestion density, free flow speed, and traffic flow characteristic parameter in the NF traffic flow model are calibrated, and the calibrated parameters are embedded into the neural network. This method can calibrate the free flow speed through a limited sample size, which enables accurate estimation of the free flow speed even in the case of missing partial observed data of a section of the road.

[0044] Among them, the expression for minimizing the error of the NF traffic flow model is:

[0045]

[0046] In the formula, MSE NF is the error of the NF traffic flow model, L represents the set of road sections, t represents the time interval, |L| represents the number of links, T (m) represents the total duration of the speed observation time series in the mth month, are respectively the observed traffic flow and observed speed of road section l at time interval t in the mth month, is the congestion density of road section l in the mth month, represents the free flow speed of road section l in the mth month, and λ (m) (l) represents the traffic flow characteristic parameter of road section l at time m.

[0047] The spatial feature extraction module constructs a hypergraph network for the geographical location of the sensor, and uses a self-organizing mapping neural network, combined with the DTW similarity metric, to define hyperedges through clustering analysis to capture complex regional associations. Please refer to Figure 2, taking the small traffic network as an example, this network consists of 5 links, and the constructed hyperedges can connect multiple nodes simultaneously. Considering the high possibility of strong correlation between sensor data on links sharing similar characteristics, in this embodiment, clustering techniques are selected to cluster traffic speed and flow observations respectively. The links between speed observations or flow observations projected onto the same mapping neuron are grouped into the same class, and hyperedges are connected within the same class of links. Please refer to Figure 3 , the small traffic network contains 5 links and 3 hyperedges. Hyperedge e1 contains links 1, 4, and 5, hyperedge e2 connects links 1, 3, and 4, and hyperedge e3 connects links 1, 2, and 5. These hyperedges connect links with common characteristics, thereby capturing the spatial dependence between them. In addition, in this embodiment, a hyperedge adjacency matrix is also constructed to represent the relationship between hyperedges and links, denoted as where |L| represents the number of links and M represents the number of hyperedges. Figure 2 or Figure 3 The dimension of the hyperedge adjacency matrix corresponding to the small traffic network in

[0048] is 5×3, and it only contains 0 and non-zero elements (non-zero elements are represented by 1 in the figure): if there is no connection between a hyperedge and a link, the corresponding element is 0; if there is a connection, the element is a non-zero value.

[0049]

[0050] In the formula, H(i,j) represents the element in the i-th row and j-th column of the hyperedge adjacency matrix H, and d ij represents the DTW distance from node i to hyperedge j, and D j represents the average DTW distance of the hyperedge.

[0051] According to the hypergraph network constructed above, a hypergraph neural network (HGNN) is used to capture the spatial correlation of traffic flow and speed.

[0052] In addition, in this embodiment, a time feature extraction module is also designed to capture the time correlation of speed and traffic flow data using a gated recurrent unit (GRU). Specifically, the historical time series observations of traffic speed and flow are used as inputs. The update gate controls the degree to which past memories are retained, and the reset gate determines how much previous memory will be ignored at the current time step, thereby obtaining time dynamic features.

[0053] After the above HGNN and GRU neural networks learn in their respective hidden layers, the spatio-temporal feature representations of traffic speed and flow are obtained by fusing the extracted features, and then these features are projected through a fully connected layer to achieve the final prediction result.

[0054] In the training process of FDINN proposed in this embodiment, a composite loss function is designed, which consists of two parts: the first part quantifies the difference between the predicted speed and the observed result, and the second part results from applying the calibrated NF traffic flow model, specifically as follows:

[0055] Loss = α × MSE v + β × MSE FD ;

[0056] Where,

[0057]

[0058] In the formula, Loss is the composite loss function, MSE v is the mean square error between the predicted speed and the observed speed, MSE FD is the loss of the calibrated NF traffic flow model, α and β are the weights of the corresponding losses, v lt is the predicted speed of section l within time period t, are the observed flow and observed speed of section l at time interval t of the m-th month respectively, L represents the set of sections, T (m) represents the total duration of the speed observation time series of the m-th month, λ (m) (l) are the congestion density, free flow speed, and traffic flow characteristic parameters of section l in the m-th month respectively.

[0059] Considering the huge numerical difference between traffic speed and flow observations, as well as the computational requirements of MSE FD , before calculating the value of MSE v , in this embodiment, min-max normalization is first applied to v lt and to scale them into the interval [-1, 1]. For the calculation of MSE FD , the anti-normalization process needs to be performed on v lt first. This method ensures that the MSE FD loss values calculated for different links maintain a comparable numerical scale. On this basis, the calculation of MSE FD is transformed into:

[0060]

[0061] Based on the above method, in this embodiment, the FDINN model is implemented using the deep learning framework PyTorch, which can use the observed data in the past hour to predict the traffic speed in the next 20 minutes. Among them, the time step is set to 12, the prediction range is set to 4, the weights α and β are respectively set to 1 and 0.4, the learning rate is optimized to 0.001, and the batch configuration is 128. In addition, the number of neurons in the hidden layer of HGNN is 256, 256, 256, 256, and the number of neurons in the hidden layer of GRU is 512, 512, 512. The fully connected neural network layer consists of 1024 neurons in the first two layers and 1088 neurons in the last layer.

[0062] Please refer to Figure 5 , the FDINN model achieved fast convergence and stabilized near its optimal state within approximately 500 iterations. The fundamental diagram part shows an overall downward trend, and occasional variations may be due to different batches of samples selected in each training epoch. The continuous reduction of errors in the training set and test set emphasizes the robust generalization performance of the FDINN model for unknown data.

[0063] In addition, to verify the prediction performance of the FDINN model, in this embodiment, the speed and flow observation data in July are selected, and the prediction time ranges are set to 5 minutes, 10 minutes, 15 minutes, and 20 minutes to display the prediction results. Please refer to Figure 6 , the horizontal axis represents the observed speed values, and the vertical axis corresponds to the predicted speed values. The predicted scatter points under different prediction time ranges are evenly distributed around the line y = x, that is, the predicted values are close to the observed values, indicating that the FDINN model can accurately capture the dynamic changes of traffic states and accurately predict speed changes. Through the comparative analysis of the speed observation and prediction heatmaps, whether on weekdays or weekends, the color intensity in the predicted heatmap is very matched with the color intensity in the observed heatmap, indicating that the predicted speed and the observed speed are highly consistent. Please refer to Figure 7 , in this embodiment, April 9, 2018 (weekday) and February 10, 2018 (weekend) are selected, and the FDINN model is used to predict the traffic speed at 8:00 am during the morning rush hour. Figure 7 In the area marked by the red circle in , the color in the central area of c is darker than the color in the central area of a, that is, the speed on weekends is higher than that on weekdays, indicating that while the FDINN model can achieve accurate traffic speed prediction, it can also capture the differences in traffic flow characteristics in the same geographical area in different months.

[0064] Embodiment 2

[0065] The traffic speed prediction method based on dynamic base map and spatio-temporal features provided by the present invention also conducts sensitivity analysis by adjusting the values of α and β to explore the impact of different loss weight coefficients on the performance of the FDINN model. In this embodiment, the root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) are used to measure the difference between the predicted speed value and the observed value, as shown in Table 1.

[0066] Table 1 Performance of the FDINN model under different weight coefficient settings

[0067] (α，β) RMSE MAE MAPE 1.0，0.0 6.03 3.72 8.13％ 1.0，0.2 5.87 3.62 7.82％ 1.0，0.4 5.56 3.51 7.52％ 1.0，0.6 5.60 3.42 7.61％ 1.0，0.8 6.32 3.89 8.50％ 1.0，1.0 7.06 4.27 9.46％

[0068] As can be seen from Table 1, when the FDINN model does not include the base map loss function (β = 0), the error of the model is greater than that of the FDINN model (for example, when β = 0.4), which indicates that the introduction of the NF base map model not only enhances the interpretability of the model but also reduces the prediction error. When the weight of the loss function reaches α = 1, the performance of the model begins to decline. This shows that the inclusion of the base map model promotes the model within a certain range, but if the direct loss between the data themselves is sacrificed and the base map relationship is overly emphasized, the performance of the model may decline. Therefore, when using the FDINN model for traffic speed prediction, it is necessary to balance the amount of prior information introduced and select appropriate weights to achieve the optimal performance of the model.

[0069] In addition, this embodiment also compares the FDINN model with several benchmark models to evaluate the prediction ability, including GRU, HGNN, LSTM, GCN, and T-GCN, as shown in Table 2.

[0070] Table 2 Comparison of prediction errors between the FDINN model and other benchmark models

[0071] Model RMSE MAE MAPE GRU 7.41 5.05 10.17％ HGNN 8.92 6.34 13.37％ LSTM 6.08 3.97 8.15％ GCN 9.58 6.80 12.86％ T-GCN 5.75 3.98 8.12％ FDINN 5.56 3.51 7.52％

[0072] As can be seen from Table 2, the FDINN model is superior to all other benchmark models in terms of the selected performance metrics. Among these models, the GRU and LSTM neural networks are good at capturing the temporal aspects of traffic observation data and achieve a MAPE within 11%. However, these models are insufficient in capturing spatial correlations. The HGNN benchmark model uses the K-means clustering algorithm to construct its hyperedges and does not fully consider the temporal dynamics existing in traffic speed data, resulting in relatively high prediction errors. The T-GCN model stands out as an excellent spatio-temporal feature prediction model. It proficiently utilizes the temporal and spatial data features and is superior to all other benchmark models. However, it does not provide insights into the base map relationship between speed and flow.

[0073] Therefore, the traffic speed prediction method based on dynamic fundamental diagrams and spatio-temporal features adopted by the present invention not only has the lowest prediction error, but also combines the understanding of fundamental diagrams to ensure that the prediction is more consistent with basic traffic physics, thereby improving the reliability of the prediction results.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A traffic speed prediction method based on dynamic basic maps and spatio-temporal features, characterized in that, It includes the following steps: S1. Select a spatial region and a time range, obtain structured data through a traffic monitoring system, including timestamps, traffic flow, traffic speed, and sensor location information, and divide the obtained structured data into a training set and a test set; S2. Construct a neural network model based on graph information, including an NF traffic flow model calibration module, a spatial feature extraction module, and a temporal feature extraction module; Among them, the NF traffic flow model calibration module customizes corresponding congestion density, free flow speed, and traffic flow characteristic parameters for different months and spatial locations to obtain a calibrated NF traffic flow model to adapt to the actual situation of traffic flow; The spatial feature extraction module uses a self-organizing map neural network, combines the DTW similarity metric, constructs hyperedges through cluster analysis, and uses a hypergraph neural network to extract spatial correlations; The temporal feature extraction module uses a gated recurrent unit to extract temporal dynamic features; S3. Weight the loss between the predicted speed and the observed value and the loss of the calibrated NF traffic flow model to construct a composite loss function, and use the training set to train the neural network model based on graph information, thereby completing the prediction of traffic speed.

2. The traffic speed prediction method based on dynamic base map and spatio-temporal features according to claim 1, wherein In step S2, the NF traffic flow model calibration module includes calibrating the congestion density, free flow speed, and traffic flow characteristic parameters in the NF traffic flow model by minimizing the error of the NF traffic flow model for different months and spatial locations, and embedding the calibrated parameters into the neural network; Among them, the expression for minimizing the error of the NF traffic flow model is: where, MSE NF is the error of the NF traffic flow model, L represents the set of road segments, t represents the time interval, |L| represents the number of links, and T (m) represents the total duration of the speed observation time series for the m-th month, are respectively the observed traffic flow and the observed speed of road segment l at time interval t in the m-th month, is the congestion density of road segment l in the m-th month, represents the free flow speed of road segment l in the m-th month, and λ (m) (l) represents the traffic flow characteristic parameter of road segment l in the m-th month.

3. The traffic speed prediction method based on dynamic base map and spatio-temporal features according to claim 1, wherein In step S2, the spatial feature extraction module includes training the self-organizing map neural network using speed observation values and flow observation values respectively to obtain a set of weights for each mapping neuron, and then combining the DTW distance to classify the links between speed observation values or flow observation values projected onto the same mapping neuron into the same class, connecting hyperedges within the same class of links, and constructing an adjacency matrix of the hyperedges. The expression is as follows: Where, H(i, j) represents the element in the i-th row and j-th column of the hyperedge adjacency matrix H, and d ij represents the DTW distance from node i to hyperedge j, and D j represents the average DTW distance of the hyperedges.

4. The traffic speed prediction method based on dynamic base map and spatio-temporal features according to claim 1, wherein In step S3, the expression for the composite loss function is: Loss = α × MSE v + β × MSE FD ; Among them, where Loss is the composite loss function, MSE v is the mean square error between the predicted speed and the observed speed, MSE FD is the loss of the calibrated NF traffic flow model, and α and β are the weights of the corresponding losses, v lt is the predicted speed of section l during period t, are the observed flow and observed speed of section l at time interval t of the m-th month, respectively. L represents the set of sections, and T (m) represents the total duration of the speed observation time series of the m-th month, λ (m) (l) are the congestion density, free flow speed, and traffic flow characteristic parameters of section l in the m-th month, respectively.

5. The traffic speed prediction method based on dynamic base map and spatio-temporal features according to claim 4, wherein Before calculating the mean square error MSE between the predicted speed and the observed speed v it is necessary to perform min-max normalization on the predicted speed v lt and the observed speed so that the range of their values is in the interval [-1, 1]. When calculating the loss of the calibrated NF traffic flow model, inverse normalization is required to ensure that the numerical scales of different links are consistent.

6. The traffic speed prediction method based on dynamic basic map and spatio-temporal features according to claim 1, wherein In step S3, the training process includes using a hypergraph neural network to extract the spatial features of traffic speed and flow data in the fully connected layer, using a gated recurrent unit to capture temporal features, and then obtaining a spatio-temporal feature representation that fuses traffic speed and flow through projection by the fully connected layer, thereby obtaining the predicted speed of traffic flow.

Images