A traffic resilience evaluation method based on a graph neural network and comprehensive resilience calculation

Abstract

The application relates to the field of intelligent transportation, and discloses a traffic resilience evaluation method based on a graph neural network and comprehensive resilience calculation.The method comprises the following steps: step 1, based on original traffic speed data, a dynamic graph neural network model is used to predict a functional index, i.e., traffic speed; step 2, based on the original traffic speed data, a spatial physical topology structure of a traffic network is constructed to calculate a structural index; and step 3, a comprehensive evaluation index is constructed by fusing the functional index and the structural index, and a comprehensive resilience evaluation algorithm is designed to realize resilience evaluation of a city traffic network. According to experiments on a public data set PEMS-BAY, compared with other latest and recognized methods, the application can more accurately evaluate the resilience of a city traffic network.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation, and particularly relates to a traffic resilience evaluation method based on a graph neural network and comprehensive resilience calculation. BACKGROUND

[0002] The urban traffic network is a combination of various roads, streets, highways, intersections and connection points within a city, which forms the basic framework of the internal and external traffic flow. A reasonably planned and designed urban traffic network can effectively connect different regions, promote smooth traffic flow, reduce congestion, and provide diverse travel options. In recent years, with the expansion of city size and the frequent occurrence of global extreme weather, the planning and management of urban traffic networks face new challenges and opportunities. Cities may face floods, heavy rains, snowstorms and other extreme weather events, which can cause road damage, traffic disruption and an increase in traffic accidents. Traditional urban traffic network performance indicators mainly include dependence vulnerability, robustness, reliability and survivability. However, these indicators have different focuses and cannot comprehensively define traffic network performance. Traffic resilience is a relatively new and multidisciplinary concept that aims to evaluate the resistance, recovery and adaptation capabilities of urban traffic systems in the face of external shocks, and comprehensively assess the reliability, vulnerability and robustness of traffic systems in the face of sudden event disturbances.

[0003] At present, the research on traffic resilience has been widely concerned at home and abroad, and the traffic resilience evaluation method is the key research direction of traffic resilience. Researchers have made some progress in the research on traffic resilience evaluation methods, but most of the evaluation methods have some problems. First of all, the lack of comprehensive resilience indicators, topological, attribute-based and functional indicators often exist separately or are simply weighted and added, which can only reflect the traffic resilience performance from a certain side and lack a global perspective. Secondly, the lack of dynamic models that meet traffic abnormal event disturbances, many evaluation methods only describe the state of the traffic system after being disturbed statically, without considering the complexity and evolution of the traffic system. Then, the mismatch between theory and practical application. Many resilience evaluation methods have not been applied to actual traffic scenarios, and the effectiveness cannot be verified.

[0004] The speed of the traffic network is an important functional indicator for evaluating its resilience, and by predicting the spatiotemporal speed data, the running changes of the urban traffic network under heavy rain disturbance can be intuitively displayed, thereby reflecting the resilience level of the traffic network. Therefore, the model can maintain good prediction performance under extreme weather such as heavy rain, which is crucial for the evaluation of the resilience of the urban traffic network. The traffic prediction problem under heavy rain disturbance is more challenging than the conventional time series prediction problem. Firstly, the traffic speed data itself is high-dimensional big data, and secondly, the non-stationarity of traffic speed time series is aggravated by heavy rain, making it difficult to predict in the long term.

[0005] The spatio-temporal traffic speed prediction needs to consider both the time dependence and the space dependence. Traditional processing methods such as autoregressive model and integrated moving average model mainly focus on the periodic characteristics of historical periodic data and do not consider the space dependence of data. Some models applying deep learning such as convolutional neural network consider the space dependence of data, but the convolutional network is more suitable for Euclidean data and extracts regular space dependence information, and the application effect is poor on non-Euclidean and irregular traffic networks. Then, the graph spatio-temporal neural network model stands out, which takes the spatio-temporal graph data as input and can effectively utilize the topological graph information of traffic and deeply capture the spatio-temporal dependence of data. Compared with other prediction methods, this kind of method greatly improves the traffic prediction accuracy. However, under the heavy rain weather, with the non-stationarity of spatio-temporal speed data being intensified, the classic graph spatio-temporal neural network has poor effect in this kind of scene. SUMMARY

[0006] In order to solve the above-mentioned blank in the field of urban traffic road resilience evaluation, the present application proposes a traffic resilience evaluation method based on graph neural network and comprehensive resilience calculation. This method makes full use of the real and rich spatio-temporal traffic data under the condition of heavy rain weather, constructs a novel urban traffic network speed prediction model based on dynamic graph neural network, and further proposes a traffic network comprehensive resilience evaluation algorithm combining the functional and structural resilience of traffic network.

[0007] The technical scheme of the present application is as follows:

[0008] The traffic resilience evaluation method based on graph neural network and comprehensive resilience calculation comprises the following steps:

[0009] Step 1) Based on the original traffic speed data, a dynamic graph neural network model is used to predict the functional index, i.e. traffic speed.

[0010] Step 2) Based on the original traffic speed data, the spatial physical topology structure of the traffic network is constructed, and the structural index is calculated.

[0011] Step 3) The comprehensive evaluation index is constructed by fusing the functional and structural indexes, and the urban traffic network resilience evaluation is realized by combining the comprehensive resilience evaluation algorithm.

[0012] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0013] (1) Firstly, for traffic speed prediction, in view of the difficulty of existing methods in time feature extraction, such as long-time short-time time dependence, non-stationarity of storm data and lack of dynamic relationship between space and time dependence, the present application combines multi-scale periodic decomposition, global-local time sequence convolution and dynamic graph convolution network to realize effective extraction of non-stationary features under storm weather, so as to realize accurate prediction of urban road network speed and provide effective support for subsequent resilience calculation.

[0014] (2) The previous traffic resilience evaluation method only considers functional or structural resilience indicators, and fails to combine the two. The present application comprehensively considers the function and structure of the road network, combines the predicted urban road network speed with the physical topology structure of the road network, and designs a new comprehensive resilience evaluation index.

[0015] (3) On the basis of the comprehensive evaluation index, a comprehensive resilience evaluation algorithm is constructed to realize the resilience evaluation of urban traffic road network under storm weather.

[0016] The present application uses multi-scale decomposition and local-global time sequence convolution to decompose the speed data under storm weather into seasonal and trend data and focuses on extracting the non-stationary features of the data, and through the dynamic graph neural network, the dynamic traffic road network topology caused by the change of speed under storm weather is simulated, which can effectively predict the long-time speed data under storm weather. In addition, the designed comprehensive resilience evaluation index and comprehensive resilience evaluation algorithm realize the accurate evaluation of the resilience of urban traffic road network from the perspective of functional and structural resilience. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The system flow chart of the urban road network traffic resilience evaluation method of the present application.

[0018] Figure 2 The model structure diagram proposed in step 1. Figure 1

[0019] Figure 3 The visualized topology structure diagram of the urban traffic road network.

[0020] Figure 4 The fitting curve diagram of the predicted value and the true value of the instance PEMS-BAY at the 9th node under normal and storm scenarios.

[0021] Figure 5 The heat map of the predicted value and the true value of the instance PEMS-BAY under storm scenario.

[0022] Figure 6 The comprehensive resilience curve diagram, the comprehensive resilience value and the functional resilience curve diagram of the instance PEMS-BAY on a certain day under storm scenario.

[0023] ​Figure 7 Figure 1 shows the integrated resilience curve and the integrated resilience value of node 224 of PEMS-BAY for a storm scenario and a day instance, and the functional resilience curve. DETAILED DESCRIPTION

[0024] Multi-scale decomposition is a method of decomposing time series data into different periodic data at different scales to better understand the periodic and trend characteristics in the data. Decomposing historical traffic speed data into periodic and trend data under storm scenarios can facilitate the design of different feature extraction modules for these two types of data, making it easier to capture the non-stationary change patterns of traffic speed data.

[0025] Local-global temporal convolution module is a temporal feature extraction framework that combines local and global dynamic temporal data through temporal convolution. Compared with traditional temporal convolution, local-global temporal convolution can capture both local and global temporal dependencies, improving the expression and generalization ability of temporal features.

[0026] Dynamic graph neural network is a graph neural network based on dynamic spatio-temporal graph. Compared with traditional static graph neural network, the adjacency matrix input of dynamic graph neural network is adaptive and changes over time. This makes dynamic graph neural network better adapt to the dynamics and complexity of spatio-temporal data, capturing the non-linear and non-stationary characteristics in spatio-temporal data.

[0027] The present application proposes a traffic resilience evaluation method based on graph neural network and integrated resilience calculation, as shown in Figure 1 which includes three steps:

[0028] Step 1) Based on the original traffic speed data, a dynamic graph neural network model is used to predict the functional indicators, i.e. traffic speed.

[0029] Step 2) Based on the original traffic speed data, the spatial physical topology of the traffic network is constructed to calculate the structural indicators.

[0030] Step 3) The functional and structural indicators are fused to construct an integrated evaluation index, and the designed integrated resilience evaluation algorithm is used to realize the resilience evaluation of urban traffic network.

[0031] The technical solutions provided by the present application will be further described below with reference to specific embodiments and their accompanying drawings. The advantages and features of the present application will become clearer in combination with the following description.

[0032] EMBODIMENT

[0033] Step 1: Predict future traffic speed data based on original traffic speed data;

[0034] A dynamic graph neural network model is designed for functional index prediction based on original traffic speed data, and the structure of the model is as shown in Figure 2 The model mainly includes five sub-modules, which are a multi-scale decomposition module, a local-global time series convolution module, a linear mapping layer module, a dynamic graph convolution module, and a full connection layer module.

[0035] The multi-scale decomposition module is used to decompose the original traffic speed data into seasonal period data and trend period data.

[0036] The local-global time series convolution module is used to process the seasonal period data to obtain a seasonal period feature matrix. Specifically, the module includes two parts, one is a local gated time series convolution, and the other is a global time series convolution; the local gated time series convolution can capture local time series features and select useful information through the gating mechanism to reduce the influence of noise; the global time series convolution can capture long-term time series dependence and increase the receptive field through multi-scale convolution kernels to improve the generalization ability of the model. The combination of the two can effectively model the time series data while keeping low computational complexity and parameter quantity.

[0037] The linear mapping layer module is used to process the trend period data, and the unstable change rule of the speed data under heavy rain is extracted to obtain a trend feature matrix.

[0038] The multi-layer dynamic spatial feature extraction module increases the time series transfer matrix on the basis of the forward and backward transfer matrices of the static graph convolutional neural network. The time series transfer matrix can complement the forward and backward transfer matrices, capture the dynamic changes of the node data in the graph, and thus improve the modeling ability of the model for spatio-temporal data.

[0039] The full connection layer module connects all neurons of the previous layer with each neuron of the current layer, thereby realizing global information transmission and feature combination, and mapping the high-dimensional hidden spatio-temporal features learned by the model into spatio-temporal speed data.

[0040] The specific steps are shown in steps 1.1, 1.2, 1.3, and 1.4.

[0041] Step 1.1: Use multi-scale decomposition to decompose the original traffic speed data into seasonal period data and trend period data, and the specific operation is shown in formulas (1) and (2).

[0042]

[0043] X trend =X-X season (2)

[0044] where X is the original traffic speed data, Xseason It is seasonal cycle data, X trend This is trend-period data. AvgPool is an average pooling convolution operation, kernel i is the i-th convolution kernel, and s is the number of convolution kernels.

[0045] Step 1.2: Input the seasonal cycle data and trend cycle data obtained in Step 1.1 into the local-global temporal convolution module and the linear mapping layer respectively, process the two types of data features during the rainy period, extract the local-global seasonal cycle features and the unstable trend cycle features, fuse the spatiotemporal information features, and generate a high-dimensional spatiotemporal feature matrix.

[0046] Step 1.2.1 Extract local-global seasonal cycle features and unstable trend cycle features.

[0047] Among them, the local-global convolution module processes seasonal periodic data, and the processing is shown in formulas (3) to (7).

[0048] F 0 =X season (3)

[0049]

[0050]

[0051]

[0052] F season =F L (7)

[0053] In the formula, L is the number of layers in the local-global convolutional module, l is the l-th layer of the local-global convolutional module (belonging to layers 1 to L), the input of layer 0 is the periodic data obtained in step 1.1, and F... l It is the input of the l-th layer. This is the local seasonal cycle feature matrix of the l-th layer. Let be the global seasonal periodic feature matrix of layer l. tanh(*) is the Tanh activation function, and sigmod(*) is the Sigmoid activation function. Conv1d local and Conv1d global These represent one-dimensional local and global temporal convolutions, respectively, with padding performed by zero-padding the data. Conv1d local The kernel size is fixed at (1, 3), Conv1d global The kernel size is in yes Length in the time dimension. F Lis the input of the last layer L and the final seasonal period feature matrix F season .

[0054] wherein the linear mapping layer processes the trend period data, and the processing formula is as shown in equation (8).

[0055] F trend = W irend *X treand +b trend (8)

[0056] wherein W trend is the weight matrix of the linear mapping layer, b trend is the bias quantity of the linear mapping layer, and F trend is the trend feature matrix obtained after processing by the linear mapping layer.

[0057] Step 1.2.2 constructs adaptive day period, week period and spatial feature matrices, and the operation process is as shown in equations (9) to (13).

[0058]

[0059]

[0060] F day = E day [X time point , :] (11)

[0061] E week = E week [X day , :] (12)

[0062]

[0063] wherein E day and E week are the adaptive embedded day period feature matrix and the adaptive embedded week period feature matrix, respectively, N timepoint is the number of sampling speeds of the monitor within a day (288 sampling times within a day by default), N fe a ture dimension of day is the day period feature dimension, N day is the number of days within a week (7 days within a week by default), N feature dimension of week is the week period feature dimension. X time point and X day are the time point (i.e., the sampling point) of the input data and the day within a week, respectively, F day is the feature matrix corresponding to the time point of the input data, and F week is the feature matrix corresponding to the day number of the input data. N is the number of city road network monitors, is the spatial feature dimension, F space is the spatial feature matrix.

[0064] Step 1.2.3 concatenates all the feature matrices into a high-dimensional spatio-temporal feature matrix, as shown in equation (14).

[0065] F feature = Cat(F season , F trend , F day , F week , F space ) (14)

[0066] Step 1.3: input the high-dimensional spatio-temporal feature matrix obtained in step 1.2 into a multi-layer dynamic graph convolution network to obtain high-dimensional hidden spatio-temporal features.

[0067] The multi-layer dynamic graph convolution network adds a time transition matrix that changes dynamically over time based on the static forward-backward transition matrix. While capturing static graph structure dependency information, it also captures dynamic graph structure dependency under the influence of changing weather, such as the rapid decline in traffic speed data monitored by some nodes under heavy rain.

[0068]

[0069]

[0070]

[0071]

[0072] TA emd = T emd + A emd (19)

[0073]

[0074]

[0075]

[0076]

[0077] Equations (15) and (16) calculate the first-order forward and backward state transition matrices and D is the degree matrix, A is the adjacency matrix, D T is the transpose of D. Equations (17) to (20) are used to calculate the first-order dynamic time state transition matrix Temd is an adaptive time embedding matrix, N embeddding dimension is an embedding dimension, A emd is an adaptive node embedding vector, TA emd is a dynamic transition node embedding vector, is the transpose of TA emd . is the input of the 0th layer dynamic graph convolution network, F feature . is the input of the mth layer dynamic graph convolution network, is the ith order dynamic time state transition matrix, and are the ith order forward and backward state transition matrices. W m0 , W m1 and W m2 are three different weight matrices of the mth layer. is the input of the Mth layer dynamic graph convolution network; F graph is a high-dimensional hidden spatiotemporal feature, which is the output of step 1.3, and the value is

[0078] Step 1.4: The high-dimensional hidden spatiotemporal feature F graph obtained in step 1.3 is passed through a fully connected layer to obtain predicted future spatiotemporal speed data.

[0079] X future = fully_connected_layer(F graph ) (24)

[0080] where fully_connected_layer is a fully connected layer operation, X future is the predicted future spatiotemporal speed data.

[0081] Step 2: Based on the original historical traffic speed data, the spatial physical topology structure of the traffic network is constructed. Then, based on the original traffic speed data, structural index calculation is performed, i.e., three types of structural resilience indexes of node degree centrality, node betweenness centrality and node reachability. The detailed steps include:

[0082] Step 2.1: Construct the spatial physical topology structure of the traffic network, model the physical topology structure of the original traffic network, and the physical topology structure is as shown in Figure 3 The spatial physical topology structure of the traffic network is a directed graph, each speed monitor node is a point of the graph, the distance of the monitor node is the length of the edge of the point of the graph, as an example, but not limited to, the graph has 308 points and 189112 directed edges.

[0083] Step 2.2: Calculate the three structural resilience indicators of node degree centrality, node betweenness centrality and node reachability, wherein the node degree centrality is the degree of the node divided by the maximum possible degree, the node betweenness centrality is the sum of the number of times the node appears in all shortest paths divided by the number of all shortest paths, and the node reachability is the number of nodes that the node can reach within a fixed step.

[0084] The calculation methods of the three indicators are shown in formulas (25) to (27).

[0085]

[0086]

[0087]

[0088] wherein n represents the number of nodes of the graph, v and t and s represent nodes, C D (v) represents the degree centrality of node v, deg(v) represents the degree of node v, C B (v) represents the node betweenness centrality of node v, the total number of shortest paths from node s to node t is σ st , and the total number of shortest paths passing through node v is σ st (v), R e (v) represents the node reachability of node v, g is a fixed step, n k (v) is the number of nodes that the node can reach at step k.

[0089] Step 3: To improve the accuracy of the model in evaluating the resilience of the urban traffic network, the changes in the functional and structural indicators of the network need to be considered comprehensively, so the comprehensive resilience indicator of the network is constructed, and the comprehensive resilience evaluation algorithm is designed for accurate calculation of the resilience loss of the network, and finally the evaluation result of the urban traffic network is obtained.

[0090] Step 3.1: Construct the comprehensive resilience indicator of the network

[0091] The specific process of calculating the comprehensive resilience indicator is as follows. The calculation method of the structural resilience indicator is shown in formula (28), the calculation method of the functional resilience indicator is shown in formula (29), the calculation method of the comprehensive resilience indicator is shown in formulas (30) and (31), and the calculation method of the comprehensive resilience capacity is shown in formula (32).

[0092] RI structure (v) = Norm(C D (v)) + Norm(C B (v)) + Norm(R e (v)) (28)

[0093] RIfunction (v) = Norm(Speed(v)) (29)

[0094] RI com (v) = RI structure (v) * RI function (v) (30)

[0095]

[0096]

[0097] C D (v), C B (v), R e (v) are the node degree centrality, node closeness centrality and node reachability of node v in step 2 respectively, Speed(v) is the output value X in step 1 future the speed value of node v at a certain time. Norm(C D (v)), Norm(C B (v)), Norm(R e (v)) and Norm(Speed(v)) are the normalized node degree centrality, normalized node betweenness centrality and normalized node reachability and normalized speed of node v calculated by the min-max normalization operation respectively. RI structure (v) is the structural resilience index, RI function (v) is the functional resilience index. RI com (v) is the comprehensive resilience index of node v, which is the product of the structural resilience index and the functional resilience index. RI com is the comprehensive resilience index of urban traffic road network, RI rain，com and RI norm，com are the comprehensive resilience indexes of urban traffic road network under rainstorm and normal weather respectively, RV rain，com is the comprehensive resilience capacity of urban traffic road network under rainstorm weather.

[0098] Step 3.2: Based on the comprehensive resilience index of road network, design a comprehensive resilience evaluation algorithm to realize the evaluation of road network resilience.

[0099] The specific details of the comprehensive resilience evaluation algorithm are shown in the following pseudo code.

[0100] The inputs of the comprehensive resilience evaluation algorithm include: historical speed label values of urban traffic road network; historical precipitation of urban traffic road network; the speed prediction value of urban traffic road network in the evaluation period T fut , t fut0 , t fut1 are the starting point and ending point of T fut respectively, tfut0 , t fut1 ∈{0, 1, 2,..., 287}; non-rainy day set D norm ; non-rainy day set on the jth day of the week

[0101] The output of the algorithm is the time-varying comprehensive resilience capability under the influence of rainstorms and the comprehensive resilience evaluation value

[0102] The algorithm is divided into three steps:

[0103] Step 3.2.1 calculates the time-varying comprehensive resilience index under the influence of non-rain, wherein is the non-rainy day set on the jth day of the week, is the jth day of the week, is the comprehensive resilience index calculated by the ith detector node at the kth sampling time point on the dth day in the set.

[0104] Step 3.2.2 calculates the time-varying comprehensive resilience index under the influence of rainstorms.

[0105] Step 3.2.3 calculates the time-varying comprehensive resilience capability under the influence of rainstorms, and integrates the time-varying comprehensive resilience capability to obtain the comprehensive resilience evaluation value.

[0106]

[0107]

[0108] To prove the speed prediction performance of the present example under regular weather and rainstorm weather, the present application performs experiments on the public data set PEMS-BAY. The data set contains a total of 308 monitor nodes, covering a continuous 1 year, with 5-minute interval statistical observation of speed data. The input of the past one hour speed (12 discrete points) is used to predict the future one hour speed (12 discrete points). When the urban traffic network is affected by rainstorm weather, the prediction accuracy of the speed prediction model at different spatio-temporal levels will be biased. Therefore, in order to prove the overall speed prediction robustness and generalization ability of the present application at different spatio-temporal levels, the road network nodes are tested at different times. Selecting nodes 9 and 13 as test nodes, the fitting curves of the prediction results and the true values of the speed prediction model proposed by the present application under non-rain and rainstorm weather are as shown in Figure 4 . Figure 4 In the figure, the horizontal coordinate is time, from 00:00 to 24:00, and the vertical coordinate is the speed value measured by the 9th monitor. The left graph is the true value and the predicted value of the speed under non-rain (regular weather), and the right graph is the true value and the predicted value of the speed under rainstorm weather (rainy day).

[0109] To further visualize the prediction effect of the application, a heat map of the predicted value and the true value under non-rainstorm and rainstorm weather is drawn, as shown in Figure 5 Figure 5 The abscissa is time, from 00:00 to 24:00, and the ordinate is the number of monitor nodes, from 0 to 307. The greater the speed measured by the monitor at a certain time point, the darker the color of the point on the heat map.

[0110] In addition, to prove the superiority of the application compared with other prediction methods. Select the latest and well-known prediction model on the public data set PEMS-BAYS for comparison experiment, and the experimental results show that the speed prediction model of the application is superior to similar models in the average absolute error (MAE), root mean square error (RMSE), and average absolute percentage error (MAPE) indicators, as shown in Table 1.

[0111] Table 1 Comparison results with well-known models

[0112]

[0113] As can be seen from Table 1, compared with the current mainstream and advanced method, the prediction performance of the application reaches a comprehensive leading position in various evaluation indicators under both normal weather and rainstorm weather. As can be seen from Table 1, compared with the current mainstream and advanced method, the prediction performance of the application almost reaches a comprehensive leading position in various evaluation indicators under both normal weather and rainstorm weather. Among them, the Graph-WaveNet[1] model combines graph convolutional neural network and causal convolution for spatio-temporal prediction of traffic data. ASTGCN[2] combines an adaptive attention mechanism to mine traffic correlations between different locations and times based on Graph-WaveNet. STNorm[3] proposes two types of normalization modules, time and space normalization, which are respectively used to significantly enhance the local and high-frequency processes of the original data. STID[4] reconsiders the traffic speed prediction problem, finds the inseparability problem of spatio-temporal data prediction, and attempts to increase the adaptive spatio-temporal matrix in the input data to improve the prediction accuracy. The prediction model proposed by the application designs and introduces multi-scale periodic decomposition, global-local time series convolution, dynamic graph neural network, etc., which greatly improves the model's ability to capture spatio-temporal dependence features under rainstorm weather, and the prediction model proposed by the application also reaches the optimal speed prediction accuracy under normal weather.

[0114] According to the comprehensive resilience evaluation algorithm, the resilience under rainstorm weather is calculated, and the overall comprehensive and functional resilience curves of the urban traffic network and the evaluation results are as shown in Figure 6 Figure 6 ​​The horizontal coordinate of the first column of the graph is time, from 00:00 to 24:00. The vertical coordinate of the first graph in the first column is the comprehensive resilience index; “-.-” is the comprehensive resilience index of a rainy day within a day, “---” is the comprehensive resilience index of the corresponding day of the week of a rainy day according to historical regular days, and “—” in the shaded interval is the comprehensive resilience index calculated according to the prediction speed of a certain rainy period. The second graph in the first column changes the corresponding comprehensive resilience to functional resilience (speed value). The first graph in the second column is a further explanation of the shaded part of the first graph in the first column. The horizontal coordinate is time, the range is the shaded interval, and the vertical coordinate is the comprehensive resilience capacity, which is the ratio of the comprehensive resilience index of the rainy period to the corresponding period of the corresponding normal day. The comprehensive resilience evaluation value is the integral of the comprehensive resilience capacity with respect to time, and the closer to 1, the stronger the resilience capacity. The second graph in the second column is a further explanation of the shaded part of the second graph in the first column, which changes the corresponding comprehensive resilience to functional resilience (speed value).

[0115] The comprehensive resilience and functional resilience curves of the node with the maximum structural resilience index and the evaluation results are shown in Figure 7

[0116] As shown in Figure 6 and Figure 7 , the comprehensive resilience index is more sensitive to heavy rain events than the functional resilience index, and the decline is greater. At the same time, the overall comprehensive resilience evaluation result under heavy rain weather is higher than that of the node with the maximum structural resilience index, which shows that the influence of heavy rain weather on the resilience capacity of a single node is higher than that of the overall.

[0117] The above is a further detailed description of the present application, which cannot be considered as limiting the specific embodiments of the present application. For ordinary skilled persons in the technical field to which the present application belongs, without departing from the concept of the present application, a number of simple deductions or substitutions can be made, which should be considered as belonging to the protection scope of the present application as defined in the claims.

[0118] The references are as follows:

[0119] [1] Z. Wu, S. Pan, G. Long, J. Jiang, and C. Zhang, “Graph wavenet for deep spatial-temporal graph modeling,” in Proc. of IJCAI, 2019.

[0120] ​[2] Guo S, Lin Y, Feng N, “Attention based spatial-temporal graph convolutional networks for traffic flow forecasting”, Proceedings of the AAAI conference on artificial intelligence, 33(01):922-929, 2019.

[0121] [3] Jinliang Deng, Xiusi Chen, Renhe Jiang, Xuan Song, and Ivor W. Tsang. 2021. ST-Norm: Spatial and Temporal Normalization for Multi-variate Time Series Forecasting. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (KDD’21).

[0122] [4] Zezhi Shao, Zhao Zhang, Fei Wang, Wei Wei, and Yongjun Xu, “Spatial-Temporal Identity: A Simple yet Effective Baseline for Multivariate Time Series Forecasting”, Proceedings of the 31st ACM International Conference on Information & Knowledge Management (CIKM’22).

Claims

1. A traffic resilience evaluation method based on a graph neural network and comprehensive resilience calculation, characterized in that, The method comprises the steps of: Step 1: based on the original traffic speed data, a dynamic graph neural network model is used to predict the functional index, i.e. traffic speed; Step 2: based on the original traffic speed data, the spatial physical topology structure of the traffic network is constructed to calculate the structural index; Step 3: the functional index and the structural index are fused to construct a comprehensive evaluation index, and a comprehensive resilience evaluation algorithm is used to realize the resilience evaluation of the urban traffic network; The step 1 comprises the following steps: The dynamic graph neural network model comprises five sub-modules, namely a multi-scale decomposition module, a local-global time series convolution module, a linear mapping layer module, a dynamic graph convolution module and a full connection layer module; wherein: The multi-scale decomposition module is used to decompose the original traffic speed data into seasonal period data and trend period data; The local-global time series convolution module is used to process the seasonal period data to obtain a seasonal period feature matrix; The linear mapping layer module is used to process the trend period data, and the non-stationary change rule of the speed data under heavy rain is extracted to obtain a trend feature matrix; The multi-layer dynamic spatial feature extraction module increases a time series transfer matrix on the basis of the forward and backward transfer matrices of the static graph convolutional neural network, the time series transfer matrix and the forward and backward transfer matrices complement each other, and the dynamic change of the node data in the graph is captured, thereby improving the modeling capability of the model for the space-time data; The full connection layer module connects all the neurons of the previous layer with each neuron of the current layer, thereby realizing global transmission of information and combination of features, mapping high-dimensional hidden space-time features learned by the model into space-time speed data; The specific steps of traffic speed prediction comprise: Step 1.1: the original traffic speed data is decomposed into seasonal period data and trend period data by using multi-scale decomposition; Step 1.2: the seasonal period data and the trend period data obtained in step 1.1 are respectively input into the local-global time series convolution module and the linear mapping layer, the two types of data features in the rain stage are processed respectively, the local-global seasonal period feature and the non-stationary trend period feature are extracted, the space-time information features are fused, the adaptive day period, week period and space feature matrices are constructed, and finally a high-dimensional space-time feature matrix is generated; Step 1.3: the high-dimensional space-time feature matrix obtained in step 1.2 is input into the multi-layer dynamic graph convolution network to obtain a high-dimensional hidden space-time feature; Step 1.4: the high-dimensional hidden space-time feature obtained in step 1.3 is input into the full connection layer to obtain the predicted future space-time speed data; The step 3 comprises: Step 3.1: the functional index and the structural index are fused to construct a road network comprehensive resilience index The structural resilience index calculation method is shown as formula (1) in the drawing. The functional resilience index calculation method is shown as formula (2) in the drawing. The comprehensive resilience index calculation method is shown as formula (3) in the drawing. The comprehensive resilience capability calculation method is shown as formula (4) in the drawing. The comprehensive resilience capability calculation method is shown as formula (4) in the drawing. In the formula: are node in-degree centrality, node out-degree centrality, node closeness centrality, node betweenness centrality, respectively are node in-degree centrality, node out-degree centrality, node closeness centrality, node betweenness centrality, is a velocity value of a node at a certain time in the predicted future spatio-temporal velocity data in step 1 is a velocity value of a node at a certain time in the predicted future spatio-temporal velocity data in step 1 , , and are the standardized node degree centrality, the standardized node betweenness centrality and the standardized node closeness centrality and the standardized speed calculated by the minimax standardization operation of the node , is the structural resilience indicator, is the functional resilience indicator; is the integrated resilience index of the node is the product of the structural resilience index and the structural resilience index of the node is a comprehensive resilience index of the urban traffic road network, and are a comprehensive resilience index of the urban traffic road network under rainstorm and regular weather, respectively, is a comprehensive resilience capacity of the urban traffic road network under rainstorm weather; Step 3.2: based on the road network comprehensive resilience index, a comprehensive resilience evaluation algorithm is designed to realize the resilience evaluation of the road network; The comprehensive resilience evaluation algorithm, which includes the following inputs: historical speed label values of the urban traffic road network, historical precipitation of the urban traffic road network, speed prediction values of the urban traffic road network in the evaluation period, a set of non-rainy days, and a set of non-rainy days in a week The outputs are time-varying comprehensive resilience capabilities and comprehensive resilience evaluation values under the influence of rainstorms. The comprehensive resilience evaluation algorithm comprises the following steps: Step 3.2.1: calculate the one-week time-varying comprehensive resilience index under the influence of non-rain; Step 3.2.2: calculate the time-varying comprehensive resilience index under the influence of rain; Step 3.2.3: calculate the time-varying comprehensive resilience capacity under the influence of rain, and integrate the time-varying comprehensive resilience capacity to obtain the comprehensive resilience evaluation value.

2. The method of claim 1, wherein, Step 1.1: The operation of decomposing seasonal period data and trend period data is as shown in the formulas and ​ in It is raw traffic speed data. It is seasonal cycle data. It is trend cycle data. It is an average pooling convolution operation. It is the first One convolutional kernel, It represents the number of convolution kernels.

3. The method of claim 1, wherein, Step 1.2 includes: Step 1.2.1: Extracting local-global seasonal periodic features and non-stationary trend periodic features Wherein, through the local-global convolution module to extract seasonal cycle characteristics, the processing process is as shown in formula to shown: In the formula, It refers to the number of layers in the local and global convolutional modules. It is the first local-global convolution module Layer, belonging to 1 to The input for layer 0 is the periodic data obtained in step 1.

1. It is the first Layer input, For the first The local seasonal periodic feature matrix of the layer, For the first The global seasonal cycle feature matrix of the layer; is a Tanh activation function, is a Sigmoid activation function; and are one-dimensional local and global temporal convolutions, respectively, and Padding is a data padding operation with zeros; The kernel size is fixed at 1. , The kernel size is ,in yes Length in the time dimension; is the last layer input of the layer, also the final seasonal period feature matrix obtained ; In the formula, the non-stationary trend cycle characteristics are extracted by a linear mapping layer, and a processing formula is as follows : In the formula, is a weight matrix of the linear mapping layer, is a bias of the linear mapping layer, is a trend feature matrix obtained after processing by the linear mapping layer; Step 1.2.2: Constructing the adaptive day cycle, week cycle, and spatial feature matrices, the operation process is as formula to formula : wherein and are an adaptive embedded day-period feature matrix and an adaptive embedded week-period feature matrix, respectively, is the number of times the monitor samples within a day, is the day-period feature dimension, is the number of days within a week, is the week-period feature dimension; and is the time point of the input data, i.e., the number of the sampling point, and is the feature matrix corresponding to the time point of the input data, is the feature matrix corresponding to the day of the input data; is the number of urban road network monitors, is the spatial feature dimension, is the spatial feature matrix; Step 1.2.3: All the feature matrices are spliced into a high-dimensional space-time feature matrix, as shown in the formula ​ 。 4. The method of claim 1, wherein, Step 1.3: The high-dimensional space-time feature matrix is input into a multi-layer dynamic graph convolution network to obtain a high-dimensional hidden space-time feature; the specific operation is as shown in formula to : Formula and Computing the first order forward and backward state transition matrices and , is the degree matrix, is the adjacency matrix, is the transpose of Formula to is a first-order dynamic time state transition matrix , is an adaptive time embedding matrix, is an embedding dimension, is an adaptive node embedding vector, is a dynamic transition node embedding vector, is transpose of is the input of the 0th layer dynamic graph convolutional network, and the value is ; is the first layer dynamic graph convolution network input, is the order dynamic time state transition matrix, and is the order forward and backward state transition matrix; and are three different first layer weight matrices; is the input of the M-th layer dynamic graph convolutional network; is the high-dimensional hidden spatio-temporal feature, whose value is .

5. The method of claim 1, wherein, Step 2 includes: Step 2.1: Constructing the spatial physical topology structure of the traffic road network, modeling the original traffic road network in the physical topology structure; Step 2.2: Calculate the node degree centrality , node betweenness centrality and node reachability The three types of structural resilience indicators are node degree centrality, node betweenness centrality and node reachability. The node degree centrality is the degree of a node divided by the maximum possible degree, the node betweenness centrality is the sum of the number of shortest paths in which the node appears divided by the total number of shortest paths, and the node reachability is the number of nodes that the node can reach within a fixed step.

6. The method of claim 5, wherein, In step 2: The calculation methods of the network node degree centrality, node betweenness centrality and node reachability are respectively shown as formulas to ​ wherein, denotes the number of nodes of a graph, and and denotes a node, denotes a node the degree centrality of a node, denotes the degree of a node , denotes the node betweenness centrality of a node , the total number of shortest paths from node s to node t is , the total number of shortest paths passing through node v is , denotes the node reachability of a node , is a fixed step size, is the number of nodes reachable by a node at step k.

Images