Novel traffic jam prediction method for unmanned vehicle road network based on software defined network

Abstract

The invention belongs to the field of unmanned driving, and provides a novel unmanned vehicle road network traffic jam prediction method based on a software defined network, which comprises the following steps: step 1, constructing a traffic parameter prediction network; step 2, constructing a traffic parameter congestion detection network; and step 3, designing a traffic jam prediction algorithm to realize prediction of a jam state. The efficient traffic jam prediction method provided by the invention can effectively reduce the impact and influence of the unmanned vehicle on wisdom and green traffic, promotes the healthy and rapid development of unmanned driving, and has important significance and application value.

Description

Technical Field

[0001] The present invention relates to the field of unmanned driving. Background Art

[0002] Traffic data is a type of data with inherent spatiotemporal correlations. Existing traffic prediction research has mainly focused on methods based on graph convolutional networks (GCNs). Current research has explored traffic flow, average speed, density, and average queue length prediction, which can help prevent and alleviate traffic congestion.

[0003] With the acceleration of urbanization and the emergence of driverless vehicles, urban traffic congestion is becoming increasingly serious. Traffic congestion not only disrupts people's daily lives but also has adverse effects on the environment. Accurately assessing traffic congestion is a prerequisite for eliminating it. Current traffic congestion detection methods have achieved some success, but they exhibit certain limitations, which are reflected in prioritizing real-time detection rather than accurately predicting the spatiotemporal scale of traffic congestion. The following problems exist:

[0004] (1) The impact of unmanned vehicles on traffic is not considered: Current research on traffic congestion mainly focuses on highways or urban scenes where manned vehicles are located. With the development of unmanned driving technology, the mixed traffic flow formed by unmanned vehicles and manned vehicles has brought new challenges to traffic congestion management. Existing methods fail to consider the impact of unmanned vehicles on manned vehicles and road environment, making it difficult to meet the management needs of mixed unmanned and manned traffic, and hindering the development of new integrated intelligent transportation systems.

[0005] (2) Failure to consider the evolution of congestion in terms of temporal and spatial scale: Current urban traffic congestion detection methods focus on single-point detection, that is, determining whether a road section or intersection is congested, but lack analysis of the congestion relationship between multiple intersections. In addition, current congestion detection methods do not consider the evolution of congestion in terms of temporal and spatial scale within a region, such as the location of congestion bottlenecks and the diffusion of congestion.

[0006] (3) Failure to consider the imbalance between congested and non-congested samples: Current traffic prediction methods mainly focus on using two deep learning methods, GNN+RNN or CNN+RNN, to predict time series data. Although GCN-based methods can extract traffic data features by combining spatiotemporal invariant information, they suffer from slow training speeds and the limitation of fixed length of time series analysis sequences, making it difficult to capture the correlation between time series of different lengths. In addition, during the training process of deep learning methods, the imbalance between congested and non-congested samples can easily lead to model overfitting, and existing studies have not paid attention to this phenomenon. Summary of the Invention

[0007] In order to solve the above problems, the present invention proposes a new traffic congestion prediction method for unmanned vehicle-road networks based on software-defined networks.

[0008] Technical solution of the present invention:

[0009] A new traffic congestion prediction method for autonomous vehicle-road networks based on software-defined networks includes the following steps:

[0010] Step 1. Build a traffic parameter prediction network;

[0011] Step 2. Construct a traffic parameter congestion detection network;

[0012] Step 3. Design a traffic congestion prediction algorithm to predict the congestion status.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] The present invention discloses an efficient traffic congestion prediction method, which is used to solve the impact of low-speed unmanned vehicles on human driving and road environment in open scenarios in terms of traffic congestion, traffic management, traffic safety and user travel. It can effectively reduce the impact and influence of unmanned vehicles on smart and green transportation, promote the healthy and rapid development of unmanned driving, and has important significance and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A new traffic congestion prediction method architecture for autonomous vehicle-road networks based on software-defined networking

[0016] Figure 2 Traffic congestion prediction algorithm flow chart

[0017] Figure 3 Quantization error of congestion detection module STSOM

[0018] Figure 4 Visualization of congestion detection distribution in one week on PeMSD8 and Jiading datasets

[0019] Figure 5 Comparison of TCP with state-of-the-art methods on PeMSD8 and Jiading datasets

[0020] Figure 6 Loss curves during training and validation

[0021] Figure 7 Visualization results of congestion scenarios at two nodes

[0022] Figure 8 Flowchart of the invention DETAILED DESCRIPTION

[0023] The "Design and Implementation Method for Spatiotemporal Data Analysis of Unmanned Vehicle-Road Networks Based on Software-Defined Networks" (applicant: Tongji University, patent application number: 202410724909.5) applied by inventors Cheng Jiujun and others on June 5, 2024 proposed a traffic parameter data set constructed by an unmanned vehicle-road network based on a software-defined network. On this basis, the present invention further proposed a new traffic congestion prediction method for unmanned vehicle-road networks based on software-defined networks.

[0024] The present invention uses a traffic parameter prediction network and a congestion detection network to achieve traffic congestion prediction. The Transformer-based traffic parameter prediction network obtains future traffic parameters through feature extraction and time series prediction, and uses an asymmetric loss function to solve the sample imbalance problem. The congestion detection network based on spatiotemporal self-organizing mapping determines the congestion state of traffic parameters through spatial detection and time series matching, and solves the problem of unequal length of time series. Through the alternating iterative training process of the two networks, the traffic congestion prediction algorithm integrates traffic parameter prediction and congestion detection. Experimental results show that on the unmanned vehicle road network dataset based on software-defined networking, the method of the present invention has high accuracy and rationality in congestion detection and prediction.

[0025] The technical solution provided by this application will be further described below in conjunction with specific implementations and accompanying drawings. The advantages and features of this application will become more apparent with the following description.

[0026] The present invention specifically comprises the following steps:

[0027] Step 1. Build a traffic parameter prediction network;

[0028] Step 2. Build a traffic congestion detection network;

[0029] Step 3. Design a traffic congestion prediction algorithm to predict the congestion status.

[0030] Details are as follows:

[0031] The present invention introduces a traffic congestion prediction method (TCP) deployed on a data computing node of an unmanned vehicle-road network based on a software-defined network, such as Figure 1 As shown in the figure, it combines a Transformer-based traffic parameter prediction network with a spatiotemporal self-organizing map (STSOM)-based traffic congestion detection network. The two networks perform traffic congestion prediction in an iterative crossover manner. Furthermore, the present invention proposes an asymmetric loss function for the software-defined network-based autonomous vehicle-road network to address the imbalance in the number of congested and non-congested samples during the prediction process.

[0032] Step 1: Build a traffic parameter prediction network

[0033] The present invention adopts a traffic parameter prediction network to capture the periodic spatiotemporal characteristics of traffic parameters and predict future traffic parameters.

[0034] The traffic parameter prediction network includes a convolution module, a feature splicing layer, a Transformer module, and a traffic parameter splicing layer, which are connected in sequence.

[0035] The predicted traffic parameters can be formalized as:

[0036]

[0037] in represents the loss function, Ψ(·) represents the proposed traffic parameter prediction network, X t Represents the traffic parameters of all road nodes at time t, where T is the input time length.

[0038] (1) Convolutional module

[0039] The convolution module is used to extract the primary spatiotemporal features of traffic parameters.

[0040] The convolution module consists of several standard three-dimensional convolutional layers. Its input dimensions are determined by the spatial and temporal dimensions of the traffic parameters, and the number of layers and output dimensions are determined by the number of predicted time windows. The corresponding example of the present invention uses three convolutional layers with 128, 64, and 32 convolution kernels, respectively, resulting in an output dimension of 32. This is then fed into a fully connected layer with 276 neurons to output primary spatiotemporal features, with a dimension of 276×1.

[0041] The primary spatiotemporal features output by the convolution module first pass through a feature splicing layer to splice the primary spatiotemporal features at different times to obtain the primary spatiotemporal features of all input traffic parameters, which are used as the input of the Transformer module.

[0042] (2) Transformer module

[0043] The Transformer module is used to predict future traffic parameters based on existing traffic parameters.

[0044] The Transformer module consists of an encoder layer and a decoder layer, and each layer includes several encoding nodes and decoding nodes.

[0045] The input dimension of the encoder layer is not less than the output dimension of the decoder layer, and the output dimension of the decoder layer is determined by the predicted time window. The input dimension of the encoder layer and the output dimension of the decoder layer of the corresponding embodiment of the present invention are 276 and 12 respectively.

[0046] The Transformer module outputs a sequence of future traffic parameters, which, after passing through a traffic parameter concatenation layer, outputs the predicted future traffic parameters.

[0047] (3) Asymmetric loss function

[0048] There is a significant imbalance in the distribution of congested and non-congested samples in the autonomous vehicle road network based on software-defined networking. During peak hours, the number of non-congested road nodes exceeds the number of congested road nodes, and this disproportion is exacerbated on weekends. To address this imbalance, this paper proposes a new asymmetric loss function The formula is as follows:

[0049]

[0050] where z e and z c represents the number of non-congested and congested road nodes respectively, and Z is the total number of road nodes, where Z = z e +z c ,y i and y i ′ represent the output value and true value of the traffic parameter prediction network respectively.

[0051] The number of congested and non-congested samples required for calculating the asymmetric loss function is obtained from the traffic parameter congestion detection network in step 2.

[0052] Step 2: Build a traffic congestion detection network

[0053] Traffic status data has inherent spatiotemporal characteristics. In order to preserve the spatial topology, the present invention uses an unsupervised self-organizing map (SOM) clustering method based on a dataset constructed from a software-defined networked unmanned vehicle-road network. However, traditional SOMs are limited to capturing two-dimensional spatial features, and 3D convolutions have difficulty learning correlations between unequal time series. In order to better utilize the spatiotemporal characteristics of new traffic data, the present invention improves the traditional SOM by matching congestion time series of different lengths by expanding the time dimension based on dynamic programming. The spatiotemporal self-organizing map (STSOM) of the software-defined networked unmanned vehicle-road network proposed in the present invention takes traffic parameters as input, learns the periodic spatiotemporal pattern of traffic congestion, and outputs the traffic congestion status, that is, the spatiotemporal scale of congestion.

[0054] The traffic parameter congestion prediction network of the present invention is used to detect whether a traffic parameter belongs to a congestion or non-congestion category.

[0055] The traffic parameter congestion prediction network includes a spatial SOM layer, a temporal SOM layer and a competition layer, which are connected in sequence.

[0056] The detected congestion state can be formalized as:

[0057]

[0058] in is the input traffic parameter, φ(·) represents the traffic parameter congestion detection network, C represents the congestion category matrix, T represents the input time length, and M represents the input space length.

[0059] (1) Spatial SOM layer

[0060] Spatial SOM (S-SOM) is used to detect congestion categories of traffic parameters in the spatial dimension.

[0061] The spatial SOM consists of a layer of network with dimension M×3, representing three traffic parameters of M road nodes, including average speed, traffic flow and occupancy rate.

[0062] At time t, for the traffic parameters X of M road nodes t , the center g of the congestion cluster satisfies:

[0063]

[0064] where w g Represents the weight of g, weight w i is randomly initialized and iteratively updated as:

[0065] w i,k+1 =w i,k +η(k)·(x t -w i,k )·ρ g,i (k) (4)

[0066] where k, η(k), and ρ g,i (k) are the learning rates of the current iteration step, the kth step, and the related neighboring kernels. The learning rate η(k) is iteratively updated as:

[0067]

[0068] Related adjacent kernel ρ g,i The calculation formula for (k) is:

[0069]

[0070] The radius δ of the adjacent core g is updated as:

[0071]

[0072] Where K is the maximum iteration step size.

[0073] (2) Temporal SOM layer

[0074] Temporal SOM (T-SOM) is used to detect congestion categories of traffic parameters in the time dimension.

[0075] The temporal SOM consists of a single layer of networks with dimensions T×M, representing the spatial congestion status of M road nodes at T moments in time. Because the congestion sequences generated by the S-SOM for each road node in a software-defined network-based autonomous vehicle road network vary in length, the proposed T-SOM uses a dynamic programming algorithm to map similar congestion sequences, thereby matching the temporal relationships between congestion sequences at adjacent road nodes.

[0076] Optimal mapping relationship satisfy:

[0077]

[0078] in Represents the mapping relationship between each two sequences and the mapping relationship between two nodes The calculation formula is:

[0079]

[0080] in The following constraints are met:

[0081]

[0082] This ensures a continuous time mapping relation R; Indicates congestion category and The similarity between them is calculated as follows:

[0083]

[0084] (3) Competition layer

[0085] The competition layer is used to integrate the congestion categories detected by the spatial SOM layer and the temporal SOM layer;

[0086] The competition layer consists of a network with the dimension of T×M×|g|, where |g| represents the number of traffic congestion categories;

[0087] At time t, the congestion category c of the mth node in the S-SOM is expressed as:

[0088]

[0089] in Represents node b at time t m congestion category;

[0090] The competition layer combines temporal similarity and spatial correlation. The congestion category of the mth node in the clustering matrix at time t is updated as:

[0091]

[0092] Defining the congestion state of the autonomous vehicle road network based on software-defined networking:

[0093]

[0094] Where g represents the cluster center, i.e., the congestion bottleneck, and ρ g,m (t) is the congested area where the bottleneck g is located at time t, λ(g,ρ g,m (t)) represents the congested nodes in the bottleneck g area, Indicates the congestion duration.

[0095] Theorem 1: The congestion state of the software-defined network-based autonomous vehicle-road network is positively correlated with the spatiotemporal scale of traffic congestion.

[0096] Proof: Congestion State The derivative with respect to time t is expressed as:

[0097]

[0098] The first part represents the spatial variation of the congestion state, and the second part represents the temporal variation of the congestion state.

[0099] According to the definition of clustering result C, the degree of spatial and temporal variation can be expressed as:

[0100]

[0101] As the spatiotemporal scale of congestion increases, therefore,

[0102]

[0103] Similarly, when the spatiotemporal scale of congestion decreases,

[0104]

[0105] therefore, It is positively correlated with t, indicating that the congestion status reflects the spatiotemporal changes in the scale of traffic congestion.

[0106] Step 3: Design a traffic congestion prediction algorithm to predict the congestion status of the autonomous vehicle road network.

[0107] In the traffic congestion prediction algorithm for the autonomous vehicle-road network, the traffic parameter prediction network constructed in step 1 and the traffic parameter congestion detection network constructed in step 2 are trained alternately and iteratively, and finally the future congestion state of the autonomous vehicle-road network is predicted.

[0108] First, the network parameters of the traffic parameter prediction network and the traffic parameter congestion detection network are initialized respectively;

[0109] Secondly, the input traffic parameters are first passed through the traffic parameter congestion detection network to obtain the corresponding congestion status. Specifically, spatial congestion is detected according to formulas (3)-(7), congestion time series are matched according to formulas (8)-(11), and spatiotemporal congestion categories are integrated according to formulas (12)-(13), and the congestion status is output;

[0110] Then, the asymmetric loss function value is calculated according to formula (2), and the future traffic parameters are predicted according to formula (1);

[0111] Finally, with future traffic parameters as input, the traffic parameter congestion detection network detects the future congestion state, and the two networks alternately iterate to achieve traffic congestion prediction.

[0112] The traffic congestion prediction algorithm is shown in Algorithm 1, and the specific flow chart is as follows Figure 2 Said, described as follows:

[0113] 1) STSOM for existing traffic parameters Perform clustering (lines 1-8).

[0114] 2) The traffic prediction network calculates the loss function based on the clustering results and predicts the future traffic parameters of the autonomous vehicle road network based on the software-defined network (Line 9).

[0115] 3) STSOM detects future traffic parameters of autonomous vehicle-road networks based on software-defined networks Traffic congestion status (lines 10-16)

[0116] 4) Aggregate the congestion categories C and calculate the congestion status (line 17).

[0117] 5) Output the congestion status of the software-defined network-based autonomous vehicle-road network (line 18).

[0118]

[0119] Simulation experiment and result analysis

[0120] (1) Evaluation indicators

[0121] The following four widely used metrics can be used to evaluate traffic congestion prediction methods. MAE is a popular evaluation metric that quantifies the average prediction error across all samples. RMSE focuses on the impact of the maximum error, while MAPE considers the ratio of the error to the true value. R² provides a comprehensive assessment of model quality, with higher values ​​indicating better prediction performance. The specific calculation method is as follows:

[0122] (1) MAE is used to measure the mean absolute error between the predicted result and the true value, which is defined as:

[0123]

[0124] where y={y1,y2,...,y n} is the true value, y′={y1′,y2′,…,y n ′} is the predicted value, and n is the number of samples.

[0125] (2) RMSE measures the error scale of the prediction results by assigning greater weights to larger errors and is defined as:

[0126]

[0127] (3) MAPE is used to measure the mean absolute percentage error of each sample and is defined as:

[0128]

[0129] (4) R2 is used to measure the predictive ability of the model and is defined as:

[0130]

[0131] in is the average value.

[0132] (2) Analysis of experimental results

[0133] (1) Traffic parameter congestion detection results

[0134] The effects of different cluster numbers and training cycles on the STSOM performance are compared by quantizing the error.

[0135] 1) Quantization error results

[0136] In order to obtain the optimal traffic parameter congestion detection module for software-defined network-based autonomous vehicle-road networks, the quantization errors of different clustering and training cycles were evaluated, as shown in Table 1. The quantization error measures the difference between the input traffic parameters and the optimal congestion detection clustering model. The results show that when the number of cycles is 700, the quantization error of 4 clusters on the PeMSD8 dataset is the smallest, while the optimal number of cycles for 2 clusters is 800. For the Jiading dataset, the optimal number of cycles for 2 clusters and 4 clusters is 800 and 700, respectively. The quantization error trends of the datasets are shown in Figure 1. Figure 3 As shown (the quantization error trends of the two datasets are shown in (a) and (b) respectively), the best model with the smallest error is applied to TCP to calculate the asymmetric loss function.

[0137] Table 1

[0138]

[0139] 2) Visualization of test results

[0140] In order to intuitively evaluate the performance of the proposed traffic parameter congestion detection method for software-defined network-based autonomous vehicle-road networks, the congestion distribution within a week is visualized to explore specific urban traffic patterns. Figure 4 (a) and (b) illustrate distinct patterns, including morning and evening peaks, as well as a distinction between weekdays and weekends in the detection results. The ratio of congested to non-congested hours is approximately 3:2. In both datasets, Friday evenings consistently exhibit the highest levels of congestion, while Sundays experience low activity throughout the day. Notably, the proportion of congested nodes peaks at 40% and 26% for both datasets, respectively, further confirming the previously hypothesized imbalance between congested and non-congested traffic states. These empirical observations underscore the effectiveness of our congestion detection approach.

[0141] (2) Traffic parameter prediction results

[0142] Table 2 shows the comparison of the proposed TCP with other traffic prediction methods on two datasets. A comprehensive comparison is performed with several state-of-the-art GCN-based methods, namely, MSTGCN [1] 、STSGCN [2] ,ASTGCN [1] 、AGCRN [3] 、DGCN [4] , and a CNN+RNN-based method similar to the prediction network architecture of the present invention, namely ResLSTM [5]. It can be observed from the results that TCP outperforms other methods in four indicators. Compared with ResLSTM, it improves by 28%, 0.25%, 7.59% and 0.62% in MAE, RMSE, MAPE and R2 respectively. In addition, TCP shows significant improvement compared with the GCN-based method, which is mainly attributed to the fact that the method of the present invention uses STSOM to effectively detect congested and non-congested states without changing the spatiotemporal dependencies. Although all methods perform slightly worse on the Jiading dataset due to the large geographical scope and the large number of road nodes, TCP still achieves satisfactory results in the unmanned driving demonstration area. The performance of TCP with and without an asymmetric loss function (calculated by the STSOM results) is compared. The combination of the traffic parameter congestion detection network of the software-defined network-based unmanned vehicle-road network and the proposed asymmetric loss function greatly improves the prediction performance. The ablation experiments conducted confirm the effectiveness of the traffic parameter congestion detection method.

[0143] Table 2

[0144]

[0145]

[0146] Table 3 shows the 12-step prediction results of TCP on both datasets. The MAE decreases slightly with the increase of the prediction period because the correlation between existing traffic parameters and future traffic parameters decreases with time. Similar trends are observed in the other three indicators over the 12 prediction steps. Although the results of the Jiading dataset are slightly worse than those of PeMSD8, the overall trend remains consistent. It is worth noting that when the prediction period is extended to 60 minutes, TCP shows satisfactory performance on both datasets, and the performance gap observed between the 5-minute and 60-minute prediction periods is the smallest among all methods.

[0147] Table 3

[0148]

[0149]

[0150] Figure 5 The traffic prediction methods on PeMSD8 and Jiading datasets are compared intuitively, and the four indicators of seven methods in 12 prediction windows are shown respectively. Figure 5 As shown in (a), except for STSGCN and TCP, the MAE indicators of all methods increase with the extension of prediction time. Figure 5 (b) shows the RMSE results, indicating that except for STSGCN, ResLSTM, TCP (without ) and TCP, the RMSE of all methods increases significantly with the extension of the prediction window. Figure 5 In (c), except for STSGCN, DGCN and ResLSTM, the MAPE of all methods increases significantly. TCP and The TCP performance of is relatively stable at first, and then increases slightly from the 10th prediction window. Figure 5 In (d), except for TCP, the R2 of all methods decreases with the length of the prediction period. In summary, TCP outperforms the state-of-the-art methods and its performance changes less with the extension of the prediction period, indicating that it is effective in both short-term and long-term predictions and has better stability on PeMSD8.

[0151] On the Jiading dataset, MAE( Figure 5 (e))、RMSE( Figure 5 (f)) and MAPE( Figure 5 (g)) shows an overall upward trend, and TCP performs better than other methods in all three indicators. At the same time, as the prediction window increases, the indicator changes less, showing the best stability. Regarding the R2 indicator ( Figure 5 (h)), all methods show a decreasing trend with increasing prediction time. Although TCP performs moderately well in the first three prediction windows, it outperforms all other methods at longer prediction times, demonstrating superior overall performance. Performance comparisons across two datasets demonstrate that the proposed TCP method demonstrates superior effectiveness and stability compared to existing methods, particularly GCN-based methods. Performance varies little with increasing prediction time windows, demonstrating that this method has excellent modeling capabilities for the spatiotemporal dynamics of traffic in software-defined network-based autonomous vehicle-road networks, leading to its superior performance in long-term predictions.

[0152] For the traffic prediction module, the training loss and validation loss curves are as follows Figure 6 As shown in the figure. During training, the training and validation losses of the PeMSD8 dataset fluctuated within the first 10 epochs and quickly converged to around 2.5 and 3, respectively, after 20 epochs. Similarly, for the Jiading dataset, the training and validation losses converged to 4.25 and 3.5 after approximately 25 and 20 epochs, respectively. After 45 epochs, both loss curves remained stable. Overall, the proposed traffic prediction network for software-defined network-based autonomous vehicle-road network showed good convergence on both datasets, with both losses dropping to the minimum level and remaining stable within 25 epochs, indicating that the traffic prediction module has strong robustness.

[0153] (3) Traffic congestion prediction results and analysis

[0154] Figure 7The congestion process at two bottleneck nodes in the Jiading road network is shown. The congestion at node A gradually increases over time and spreads to the adjacent node B. Therefore, node A becomes the bottleneck node in the congested area and the epicenter of the congestion. It is clear from the diagram that the congestion originates at node A and eventually extends to five adjacent roads. Notably, the traffic parameter predictions for nodes A and B closely match the actual values, accurately depicting the spatiotemporal extent of future congestion. The proposed TCP can accurately predict the spatiotemporal scale of congestion and precisely locate the bottleneck road segments within the congestion. This capability is of great value to transportation systems and can help address persistent congestion at its source.

[0155] The above description is only a description of the preferred embodiments of the present application and does not limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical content should be regarded as equivalent valid embodiments and fall within the scope of protection of the technical solution of the present application.

[0156] References:

[0157] [1]Guo S, Lin Y, Feng N, et al.Attention based spatial-temporal graphconvolutional networks for traffic flow forecasting[C] / / Proceedings of the AAAI conference on artificial intelligence.2019,33(01):922-929.

[0158] [2]Song C,Lin Y,Guo S,et al.Spatial-temporal synchronous graphconvolutional networks:Anew framework for spatial-temporal network dataforecasting[C] / / Proceedings of the AAAI conference on artificialintelligence.2020,34(01):914-921.

[0159] [3]Bai L,Yao L,Li C,et al.Adaptive graph convolutional recurrentnetwork for traffic forecasting[J].Advances in neural information processingsystems,2020,33:17804-17815.

[0160] [4]Guo K,Hu Y,Qian Z,et al.Dynamic graph convolution network fortraffic forecasting based on latent network of Laplace matrix estimation[J].IEEE Transactions on Intelligent Transportation Systems,2020,23(2):1009-1018.

[0161] [5]Zhang J,Chen F,Cui Z,et al.Deep learning architecture for short-term passenger flow forecasting in urban rail transit[J].IEEE Transactions onIntelligent Transportation Systems,2020,22(11):7004-7014.

Claims

1. A new traffic congestion prediction method for driverless vehicle-road network based on software-defined network, characterized in that: The steps include: Step 1. Construct a traffic parameter prediction network; Step 2. Construct a traffic parameter congestion detection network; Step 3. Design a traffic congestion prediction algorithm to predict the congestion status.

2. The novel traffic congestion prediction method for unmanned vehicle-road network based on software defined network as claimed in claim 1, characterized in that: Step 1: The traffic parameter prediction network is used to capture the periodic spatiotemporal characteristics of traffic parameters and predict future traffic parameters; The traffic parameter prediction network includes a convolution module, a feature splicing layer, a Transformer module, and a traffic parameter splicing layer, which are connected in sequence; Predicted traffic parameters Formally: in represents the loss function, Ψ(·) represents the proposed traffic parameter prediction network, X t Represents the traffic parameters of all road nodes at time t, where T is the input time length; Convolutional module: The convolution module is used to extract primary spatiotemporal features of traffic parameters; The convolution module consists of several standard three-dimensional convolutional layers, whose input dimensions are determined by the spatial and temporal dimensions of the traffic parameters, and the number of layers and output dimensions are determined by the number of predicted time windows; The primary spatiotemporal features output by the convolution module first pass through a feature concatenation layer to concatenate the primary spatiotemporal features at different times to obtain the primary spatiotemporal features of all input traffic parameters, which are then used as the input of the Transformer module. Transformer module: The Transformer module is used to predict future traffic parameters based on existing traffic parameters; The Transformer module consists of an encoder layer and a decoder layer, each layer includes a number of encoding nodes and decoding nodes; The input dimension of the encoder layer is not less than the output dimension of the decoder layer, and the output dimension of the decoder layer is determined by the predicted time window; The Transformer module outputs a sequence of future traffic parameters, which, after passing through a traffic parameter concatenation layer, outputs the predicted future traffic parameters. Asymmetric loss function: Asymmetric loss function The formula is as follows: where z e and z c represents the number of non-congested and congested road nodes respectively, and Z is the total number of road nodes, where Z = z e +z c ,y i and i ′ respectively represent the output value and true value of the traffic parameter prediction network; The number of congested and non-congested samples required for calculating the asymmetric loss function is obtained from the traffic parameter congestion detection network in step 2.

3. The novel traffic congestion prediction method for unmanned vehicle-road network based on software defined network as claimed in claim 1, characterized in that: Step 2: The traffic parameter congestion prediction network is used to detect whether the traffic parameter belongs to the congestion or non-congestion category; The traffic parameter congestion prediction network includes a spatial SOM layer, a temporal SOM layer and a competition layer, which are connected in sequence; The detected congestion state is formalized as: in is the input traffic parameter, φ(·) represents the traffic parameter congestion detection network, C represents the congestion category matrix, T represents the input time length, and M represents the input space length; Spatial SOM layer: Spatial SOM, namely S-SOM, is used to detect the congestion category of traffic parameters in the spatial dimension; The spatial SOM consists of a layer of networks with a dimension of M × 3, representing three traffic parameters of M road nodes, including average speed, traffic flow, and occupancy rate; At time t, for the traffic parameters X of M road nodes t , the center g of the congestion cluster satisfies: where w g represents the weight of g, weight w i is randomly initialized and iteratively updated as: where k, η(k), and ρ g,i (k) are the learning rates of the current iteration step, the kth step, and the related adjacent kernels; the learning rate η(k) is iteratively updated as: The relevant neighbor kernel ρ g,i The calculation formula of (k) is: The radius δ of the adjacent core g is updated as: Where K is the maximum iteration step size; Temporal SOM layer: Temporal SOM or T-SOM is used to detect the congestion category of traffic parameters in the time dimension; The temporal SOM consists of a layer of networks with a dimension of T×M, which represents the spatial congestion status of M road nodes at T moments. T-SOM uses a dynamic programming algorithm to map similar congestion sequences and match the temporal relationship between congestion sequences of adjacent road nodes. Optimal mapping relationship satisfy: stmax(M,T)≤|R|≤M+T (8) in Represents the mapping relationship between each two sequences and the mapping relationship between two nodes The calculation formula is: in The following constraints are met: This ensures a continuous time mapping relationship R; Indicates congestion category and The similarity between them is calculated as: Competition layer: The competition layer is used to integrate the congestion categories detected by the spatial SOM layer and the temporal SOM layer; The competition layer consists of a network with a dimension of T×M×|g|, where |g| represents the number of traffic congestion categories; At time t, the congestion category c of the mth node of S-SOM is expressed as: in represents node b at time t m congestion category; The competition layer combines temporal similarity and spatial correlation, and the congestion category of the mth node in the clustering matrix at time t is updated as:

4. The novel traffic congestion prediction method for unmanned vehicle-road network based on software defined network as claimed in claim 1, characterized in that: Step 3: In the traffic congestion prediction algorithm, the traffic parameter prediction network constructed in step 1 and the traffic parameter congestion detection network constructed in step 2 are trained alternately and iteratively, and finally the future congestion state of the driverless vehicle road network is predicted; specifically: First, the network parameters of the traffic parameter prediction network and the traffic parameter congestion detection network are initialized respectively; Secondly, the input traffic parameters are first passed through the traffic parameter congestion detection network to obtain the corresponding congestion status, including: detecting spatial congestion, matching congestion time series, integrating spatiotemporal congestion categories, and outputting congestion status; Then, the asymmetric loss function value is calculated to predict future traffic parameters; Finally, taking future traffic parameters as input, the traffic parameter congestion detection network detects the future congestion state, and the two networks alternately iterate to realize traffic congestion prediction.

5. The novel traffic congestion prediction method for unmanned vehicle-road network based on software defined network as claimed in claim 4, characterized in that: The traffic congestion prediction algorithm is specifically described as follows: 1) STSOM for existing traffic parameters Perform clustering; 2) The traffic prediction network calculates the loss function based on the clustering results and predicts the future traffic parameters of the driverless vehicle-road network based on the software-defined network 3) STSOM detects future traffic parameters of driverless vehicle-road networks based on software-defined networks Traffic congestion status; 4) Aggregate congestion category C and calculate congestion status; 5) Output congestion status.

Images