A method for evaluating urban traffic resilience based on multi-source heterogeneous data and large language model

By integrating multi-source heterogeneous data with a large language model, the urban traffic resilience assessment method solves the problems of global and local scale matching and static and dynamic responses in traffic resilience assessment, and realizes the accurate assessment and dynamic response capability of the urban traffic system under complex disturbances.

CN120562925BActive Publication Date: 2025-09-23TONGJI UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202511044830.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-07-29
Publication Date
2025-09-23
Estimated Expiration
2045-07-29

AI Technical Summary

Technical Problem

Existing traffic resilience assessment methods have shortcomings in global and local scale matching, static and dynamic response assessment, and are difficult to effectively integrate multi-source heterogeneous data. In particular, the prediction accuracy is insufficient during high-error periods, and text event information is difficult to jointly model with traffic flow data.

Method used

An urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model is adopted. Text event information and traffic flow data are encoded through the large language model. Combined with graph neural network and time series modeling technology, a spatiotemporal feature extractor is constructed to conduct a fusion evaluation of functional and structural indicators, including a peak correction mechanism and the integration of comprehensive resilience indicators.

Benefits of technology

It achieves accurate assessment of urban traffic systems under complex disturbance conditions, improves prediction accuracy during abnormal periods, provides a dynamic and adaptive multi-index evaluation system, and enhances the system's response capabilities and decision-making value in real environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120562925B_ABST
    Figure CN120562925B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of intelligent transportation, and proposes a method for evaluating the resilience of urban traffic based on multi-source heterogeneous data and a large language model. Specifically, it includes: constructing a prediction model that integrates traffic flow spatiotemporal data and event text semantics, using a pre-trained large language model to extract the intermediate semantic representation of the event text, introducing an automatic correction mechanism driven by prediction errors, focusing on identifying high error and event disturbance intervals and implementing compensation, and finally combining road network structural resilience and functional resilience indicators to form a comprehensive, dynamic, and adaptive resilience evaluation system. Experimental results on the public dataset BjTT show that the proposed method can significantly improve prediction accuracy and evaluation accuracy in multiple types of extreme event scenarios, is superior to the existing latest methods, and has good prospects for promotion and application.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation, and specifically to an urban transportation resilience assessment method based on multi-source heterogeneous data and a large language model. Background Art

[0002] The transportation network is a vital infrastructure for modern urban operations. Its function is not only to support the orderly flow of various urban traffic activities but also to ensure the efficient implementation of activities such as daily commuting, logistics transportation, and emergency rescue. An efficient transportation network system should generally possess good connectivity and accessibility, capable of providing stable and reliable transportation services under various external environmental conditions. However, with the rapid pace of urbanization and the intensification of social operations in recent years, the transportation system has become increasingly vulnerable to disturbances, including natural disasters (such as heavy rains and typhoons), man-made events (such as traffic accidents and large-scale events), and infrastructure failures (such as road construction and traffic light failures). These special events are often unpredictable, highly sudden, and multi-source, which can easily lead to significant declines in network efficiency and even traffic paralysis. In traditional research, transportation system performance is often evaluated using static indicators such as average speed, road saturation, delay time, and travel cost. While these indicators can characterize the system's operational capabilities under a given state, they fail to reflect the dynamic response of the transportation system to the impact of sudden events. Therefore, in recent years, "resilience" has been introduced into the transportation field as an emerging concept of multidisciplinary intersection. It is used to measure the pressure resistance, absorption capacity, response speed and recovery ability of the transportation system when facing disturbances. It represents a more comprehensive and dynamic system performance measurement method. [1] .

[0003] With the increasing frequency of external disturbances such as extreme climate and emergencies, traffic resilience has gradually become one of the hot topics in transportation system research and urban planning. However, most current resilience research methods still have many shortcomings. On the one hand, there is a scale mismatch between global resilience and local resilience assessment: some methods focus on the macro indicators of the entire road network, but ignore the micro performance of local key nodes and sections; on the other hand, some current methods rely on static indicator evaluation and fail to effectively capture the dynamic response of the system over time. In addition, some methods emphasize the derivation of theoretical models and lack real data verification or docking with actual application scenarios, resulting in insufficient generalizability and practicality of the models in real scenarios. These problems limit the reliability and explanatory power of resilience assessment results, and there is an urgent need for a more practical dynamic, data-driven resilience assessment framework.

[0004] In urban traffic systems, the average speed of a road section, as a core variable to characterize traffic status, is an important foundation for building a resilience assessment system. By accurately predicting traffic speed, we can grasp the operating trend of the traffic system in the future in advance, and then evaluate its ability to respond to potential events. However, high errors in speed predictions are often concentrated in "peak" periods, such as rush hour or when major disturbance events occur. These "peak" changes include both periodic fluctuations in congestion and smooth traffic, as well as event-driven sudden changes. Due to the high-dimensional, dynamic and unbalanced characteristics of traffic data, the prediction model faces challenges such as difficulty in data processing and modeling, difficulty in feature extraction, and insufficient response to "peaks"; at the same time, special event information often exists in text form, which is difficult to directly jointly model with traffic flow data. Therefore, how to effectively introduce text event information and improve the prediction accuracy of speed during abnormal periods has become the focus and difficulty of current research. [2] .

[0005] To address the above issues, the integration of multi-source heterogeneous data has become a key path to improving resilience assessment performance. Among them, traffic flow data mainly reflects the continuity and structure of traffic operation status, while text event data reflects the sparsity and semantic nature of disturbance information. The two have significant heterogeneity from data structure, expression form to temporal characteristics. In recent years, large language models (LLMs) have performed well in semantic understanding. They can encode event text into high-dimensional semantic embedding representations through deep language modeling capabilities, thereby achieving deep fusion between structured and unstructured data. In addition, by using graph neural networks and time series modeling technology, periodic pattern information can be introduced to fully explore the potential periodic laws and evolution trends in traffic flow, providing support for improving prediction and assessment accuracy.

[0006] In spatiotemporal traffic speed modeling, it is necessary to pay attention to both the time scale and the spatial scale. Early research focused on mining time dependencies, while in recent years, graph neural networks have [3] The introduction of [ ] significantly improved the accuracy of spatiotemporal traffic flow mining. Temporal scale: The core challenge lies in the rapid decay of long-term prediction accuracy over time. Mainstream approaches capture long-range dependencies through mechanisms such as multi-head attention to mitigate long-term prediction distortions. Spatial scale: The key lies in simultaneously capturing both the inherent road network topology and the underlying dynamic spatial relationships that change with events. Therefore, the model should take into account both the event-driven inherent graph structure changes and the real-time learnability of dynamic graphs to comprehensively characterize the spatiotemporal dynamics of urban transportation systems. Summary of the Invention

[0007] This paper proposes an urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model. It aims to comprehensively utilize traffic data and text event information, integrate functional and structural indicators, and achieve a comprehensive and accurate assessment of the resilience level of the urban traffic system under disturbance conditions.

[0008] Technical Solution

[0009] A method for evaluating urban traffic resilience based on multi-source heterogeneous data and a large language model includes the following steps:

[0010] Step 1: Functional Resilience Assessment: Construct functional indicators based on raw traffic data and text information. This step is divided into two phases: the first phase uses raw traffic data and text information to make preliminary traffic status predictions, and the second phase performs peak correction on the prediction results of the first phase to improve the prediction accuracy during high-disturbance periods;

[0011] Step 2: Structural Resilience Assessment: Based on the raw traffic data and the topological structure of the transportation network, structural resilience indicators are calculated. These structural resilience indicators include node degree centrality, node closeness centrality, and node average edge betweenness centrality.

[0012] Step 3: Integrate the aforementioned structural and functional resilience indicators to construct a comprehensive road network resilience index. Integrate the comprehensive resilience index over a period of time to obtain the final resilience assessment result, which is used to quantitatively analyze and evaluate the overall resilience level of the urban transportation system under complex disturbances.

[0013] Beneficial effects

[0014] The urban traffic resilience assessment method of the present invention, combining multi-source heterogeneous data with a large language model, can comprehensively and accurately assess the resilience of the traffic system under complex and dynamic disturbance conditions, and has high practical value and broad application prospects. Specifically:

[0015] (1) During traffic speed prediction, high errors often occur during abnormal periods and extreme disturbance scenarios. To address this, the present invention introduces an automatic correction mechanism based on training error analysis. This mechanism combines historical error patterns with event characteristics (i.e., the temporal textual characteristics of the road network) to dynamically compensate for large prediction deviations during the testing phase. This mechanism not only improves the speed prediction model's performance during the "peak" period but also provides a more accurate and stable input foundation for the subsequent resilience assessment module, effectively mitigating the impact of abnormal disturbances on assessment accuracy.

[0016] (2) Traffic resilience is a comprehensive reflection of the urban transportation system's ability to resist disturbances and recover. Its modeling relies not only on accurate state prediction, but also requires the integration of system structural attributes and functional behavior characteristics. In this paper, resilience assessment comprehensively considers multiple dimensions such as structural resilience and functional resilience, and constructs a dynamic, adaptive, and quantifiable multi-index evaluation system based on the integrated prediction results, thereby achieving accurate quantification and trend determination of the resilience level of the urban transportation system, and improving the system's response capability and decision-making value in real-world environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] FIG1 is a flow chart of the urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model according to the present invention;

[0018] FIG2 is a schematic diagram of a spatiotemporal feature extraction module in the first stage of step 1 of the present invention;

[0019] FIG3 is a visual structural diagram of an urban traffic network according to an embodiment of the present invention;

[0020] FIG4 is a comparison of MAE of the node with the largest error before correction in the training set and corrected in the test set according to an embodiment of the present invention;

[0021] FIG5 is a comparison of the RMSE of the node with the largest error before correction in the training set and corrected in the test set according to an embodiment of the present invention;

[0022] FIG6 shows the correction rate percentage of the node with the largest error in the training set before correction in the test set according to an embodiment of the present invention;

[0023] FIG7 is a curve showing the comprehensive toughness index and single speed index of node 37 (S50 East Fifth Ring Road Section 1) according to an embodiment of the present invention;

[0024] FIG8 is a curve showing the sum of the comprehensive toughness index and the sum of the single speed index for all sections of the S50 East Fifth Ring Road according to an embodiment of the present invention. DETAILED DESCRIPTION

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

[0026] Example

[0027] A method for evaluating urban traffic resilience based on multi-source heterogeneous data and large language model includes the following steps (e.g. Figure 1 ):

[0028] Step 1: Functional Resilience Assessment: This involves constructing functional indicators based on raw traffic data and textual information, i.e., conducting traffic speed forecasts. This step is divided into two phases: the first phase uses raw traffic data and textual information to conduct preliminary traffic status forecasts; the second phase applies a "peak correction" to the forecast results from the first phase to improve forecast accuracy during periods of high disturbance;

[0029] The text information includes a description of the traffic incident, which serves as an external factor to make the model prediction more accurate.

[0030] In this example, the dataset used is Beijing Text-Traffic (BjTT-January). This dataset contains 1,260 road nodes, covers a one-month period, and has a temporal resolution of one time slice every four minutes. Data features include traffic speed and congestion level. Each time slice corresponds to a text message describing an event that occurred on a particular road.

[0031] Step 1.1: Phase 1: Preliminary prediction of traffic speed;

[0032] The first phase predicts preliminary traffic speeds based on raw traffic data and textual information. A pre-trained large language model is used to encode event text to generate high-dimensional semantic features. These high-dimensional semantic features are then fused with the raw traffic speed sequences and fed into a spatiotemporal feature extractor (ST-RetNet). The ST-RetNet module utilizes a RetNet architecture at both spatial and temporal scales, enabling it to simultaneously capture the topological relationships of the road network and the spatiotemporal dependencies of the speed sequences. Subsequently, two convolutional layers are used to output preliminary traffic speed predictions.

[0033] The specific steps are as follows:

[0034] Step 1.1.1: Construct a two-dimensional time encoding for each time step in the dataset and concatenate it with the feature dimension of the original velocity vector as the exogenous time context feature of the network ;

[0035] First, construct a two-dimensional time encoding for each time step in the dataset:

[0036]

[0037] in, It represents the standardized coding method of time of day (TOD), where is the timestamp, is the total number of time periods in a day ( ). It represents the encoding method of the day of the week (DOW). It first passes Timestamp Map to the corresponding day. Then, use The two encoding methods represent the time characteristics and the characteristics of the day of the week respectively, and convert the periodic characteristics into a numerical range through standardization. . At the corresponding road node and time slice and Concatenate with the feature dimension of the original velocity vector as the exogenous temporal context feature of the network , where N is the number of road nodes in the road network and T is the time step.

[0038] Step 1.1.2: Encode text information and generate temporal text features ;

[0039] Encode text information, which includes a description of a traffic event. Suppose the event text sequence is:

[0040]

[0041] The text corresponding to the t-th prediction time step, For words in the text, is the length of the text. Input the text into a pre-trained large language model (such as "Longformer-base-4096"), and the model will calculate and generate the corresponding hidden state based on the text content:

[0042]

[0043] Next, a convolution operation is performed ( ) Process the hidden state sequence, reduce the token dimension to 1, and generate the final time text feature The temporal text features provide real-time, supplementary data for traffic speed prediction, enabling the model to make more accurate predictions in complex traffic environments.

[0044] For example, if the text source repeatedly displays "prohibit left turn on Xiangheyuan Road" over a period of time, relying solely on a time series model for prediction would often result in worse predictions for this road section than under normal traffic conditions, as the model would be unable to detect the temporary traffic control information.

[0045] After the text information is introduced, the event text "prohibit left turn on Xiangheyuan Road" is first encoded using the large language model, and then the obtained representation is The model is then integrated with the original traffic speed time series features. This allows the model to explicitly understand the impact of the "No Left Turn" policy on local traffic flow, thereby providing a more accurate prediction of vehicle speed changes on Xiangheyuan Road during the duration of the policy.

[0046] Step 1.1.3: Combine the temporal text features with the road nodes where the events occurred;

[0047] Assume the event node index set is:

[0048]

[0049] in, , is the number of road nodes in the road network, Indicates that an event occurred on road k.

[0050] The time text feature only performs feature splicing on the road nodes where the event occurred, that is, according to The temporal text features for constructing the road network are:

[0051]

[0052] For event node index collection Each road node in , the time text feature (and The values ​​in each time dimension are the same, that is, splicing on consecutive time slices) is assigned to , ensuring that all event-related nodes share the same time-context feature vector, which is convenient for subsequent interaction with exogenous time context features. Prepare for splicing. and They are all three-dimensional features, and their three dimensions represent time and position in turn , Road Index and feature dimension (size is ).symbol" " means selecting all elements on this dimension, and formula (6) means At the moment The complete features under are assigned to the tensor At the same time And the road is characteristics.

[0053] Step 1.1.4: Concatenate exogenous temporal context features and temporal text features to obtain feature vectors ;

[0054] Exogenous temporal context features obtained in step 1.1.1 After the convolution operation, the feature dimension is expanded to 64 dimensions; similarly, the time text feature obtained in step 1.1.3 The feature dimension is also expanded to 64 dimensions through convolution operation. The two features are concatenated according to the feature dimension to obtain the feature vector .

[0055] Eigenvector Provided to the spatiotemporal feature extractor.

[0056] Step 1.1.5: Build a spatiotemporal feature extractor and process feature vectors , and obtain preliminary traffic speed results.

[0057] The number of layers of the spatiotemporal feature extractor (spatiotemporal network) ST-RetNet is one or more layers in series. In the embodiment, only the case where the number of network layers is 1 is introduced, such as Figure 2 .

[0058] The spatiotemporal network consists of a spatial module, S-RetNet, and a temporal module, T-RetNet. Both modules use a spatial-temporal serial approach. The output of the spatial module is fed into the temporal module, which then feeds two convolutional layers to produce the final prediction result.

[0059] (1) Space module

[0060] The input of the spatial module is and the adjacency matrix set ,in, is the inherent adjacency matrix of the road network, is the adaptive adjacency matrix.

[0061] Intrinsic adjacency matrix It is dynamically generated based on the road congestion level in the original data. The inherent adjacency matrix of the road network is known, which contains only two values: 0 and 1: 0 means that the road shown in the corresponding row and column is not connected, and 1 means it is connected. When a road is detected as "closed" (corresponding to a congestion level of 5 and a traffic speed of 0 on this road section), the row and column of the corresponding node in the adjacency matrix are set to 0, forming a dynamic mask for the inherent adjacency matrix, and obtaining a dynamic inherent adjacency matrix that reflects the real-time accessibility of the road network. .

[0062] Adaptive adjacency matrix The model automatically learns the weights of this matrix through the training process, thereby describing the dynamic spatiotemporal relationship of the road network. and , the adaptive adjacency matrix is ​​generated as follows:

[0063]

[0064] in, and is the activation function.

[0065] The spatial module S-RetNet includes: GCN, space preservation mechanism, spatial feature aggregation, gated multi-scale space preservation mechanism and feedforward network.

[0066] Wherein, GCN: Based on the spatial relationship of sensors in a given road network, GCN can effectively capture the static spatial characteristics of the road network. In this embodiment, GCN is represented by the following formula:

[0067]

[0068] is the activation function, and is the parameter matrix, is the input of S-RetNet, and same.

[0069] Space reservation mechanism: e.g. Figure 2 , the space preservation mechanism is defined as follows, where the three linear layers complete 、 operate:

[0070]

[0071] in and Mutually conjugate complex numbers, a relative position embedding method [4] . , and is the parameter matrix. is the Hadamard product (element-wise product).

[0072] Spatial feature aggregation: In order to obtain both the static and dynamic spatial dependencies of the road network, information aggregation is performed according to the following formula:

[0073]

[0074] is a one-dimensional convolution, For splicing operation.

[0075] Gated multi-scale retention mechanism: In order to better capture the complex relationship between input sequences and extract useful information, the present invention adopts a gated multi-scale retention mechanism method. scale Analyze the input sequence , through multiple scales and different parameter matrices, to capture the dynamic characteristics of the sequence at multiple scales in parallel. Specifically, according to formulas (7) to (10), the corresponding multi-scale calculation results are given Then, the multi-scale retention mechanism is calculated according to formula (11). , and finally use the gated structure to increase the nonlinearity of the model.

[0076]

[0077] in is the normalization operation, and is the parameter matrix, is the activation function, This is the output result of spatial multi-scale retention (MSR).

[0078] Feedforward network: Process as follows, output The final output of S-RetNet is:

[0079]

[0080] in and is the parameter matrix, SiLU is the activation function.

[0081] Output of the spatial module and input Provided to the time module.

[0082] (2) Time module

[0083] The temporal module T-RetNet includes a time-preserving mechanism, a gated multi-scale time-preserving mechanism, and a feedforward network.

[0084] Among them, the time retention mechanism introduces , which combines the causal mask and the exponential decay of relative distance into a matrix. The gated multi-scale time preservation aggregates the features from each head and captures the temporal dependencies more comprehensively. Finally, the output of T-RetNet is obtained through the feedforward network. .

[0085] Time retention mechanism: As shown in Figure 2, the time retention mechanism is defined as follows: 、 For linear layer operation:

[0086]

[0087] in ; and Mutually conjugate complex numbers; , and is the parameter matrix; Is a relative position coding matrix that gradually decays as the time interval increases [4] .

[0088] Gated multi-scale time retention mechanism: In order to better capture the complex relationship between input sequences and extract useful information, this paper adopts a gated multi-scale time retention mechanism. Analyze the input sequence , and the corresponding multi-scale calculation results are given according to formula (13) According to formula (14), the multi-scale retention mechanism is calculated , and then use the gated structure to increase the nonlinearity of the model.

[0089]

[0090] in is the output result of Multi-Scale Retention (MSR). Used to generate h constants with evenly spaced values, Assign different constants to each scale .

[0091] Feedforward network: Process as follows, output The final output of T-RetNet is:

[0092]

[0093] in and is the parameter matrix.

[0094] Finally, through two convolution operations, the final prediction result is obtained .

[0095] Step 1.2: Second stage: Use the peak deviation calculation algorithm to extract the cumulative average error for each road; input this deviation value, along with the periodic factor and the high-dimensional semantic features of the text, into the gated fusion module to adjust the initial prediction results; and output the final future traffic speed prediction value.

[0096] To further enhance the model's ability to predict periodic anomalies such as morning and evening peaks, a periodic error correction mechanism and a text event information fusion mechanism are introduced based on the first-stage prediction. The specific process is as follows:

[0097] Step 1.2.1: Peak deviation calculation algorithm;

[0098] First, the basic model predicts data in batches, with each batch being a sample. The sample size (i.e., the amount of data processed in each batch) is set based on the device and model scale. In this example, it is set to 32.

[0099] For the training set, the predicted values ​​and true values ​​obtained by the first-stage basic model are expressed as and , representing the samples at the prediction time step and the road node number is The traffic speed prediction and true value on Corresponding to the first stage output The absolute error at each time period and node is calculated using the following formula:

[0100]

[0101] Then, the error is periodically mapped in the time dimension (for example, a day is divided into time slices), and obtain the periodic average error matrix , which is calculated as:

[0102]

[0103] in, For the corresponding time and nodes The cumulative number matrix of the upper error, is the number of time slices in a day, Indicates that in all and In the combination, only those that meet the "sum of the two models" are selected. equal "In this embodiment, , The value range is 0 to 359.

[0104] Finally, the predicted value of the basic model and the average error are corrected for the peak value at the corresponding time and node:

[0105]

[0106] in, Indicates time and nodes superior, In all samples The value of .

[0107] The pseudo code of the above process (peak deviation calculation algorithm) is as follows:

[0108]

[0109] Step 1.2.2: Time period feature encoding;

[0110] Based on the revised prediction results, two periodic time dimensions (equations (1) and (2)) are introduced as auxiliary features:

[0111] Time of day:

[0112] Day of the week:

[0113] Compare it with the corrected speed Perform splicing to construct a three-dimensional periodic time feature tensor:

[0114]

[0115] Step 1.2.3: Text feature alignment and gated fusion;

[0116] In order to integrate external event information, we use formula (6) to get , the time feature vector With text feature vector Fusion via gating mechanism:

[0117]

[0118] in, is the parameter matrix, represents the activation function, Represents a splicing operation, represents element-wise multiplication, is the final fused representation vector.

[0119] Step 2: Structural resilience assessment phase: Based on the original traffic data and combined with the topological structure of the traffic network, calculate the structural resilience indicators.

[0120] Based on the original historical traffic speed and congestion data, the real-time spatial physical topology of the traffic network is constructed (i.e., the inherent adjacency matrix of the road network). Structural indices are then calculated, namely, three types of structural resilience indices: node degree centrality, node closeness centrality, and node average edge betweenness centrality.

[0121] The specific steps are as follows:

[0122] Step 2.1: Construct the spatial physical topology of the traffic network and model the physical topology of the original traffic network. Figure 3 shown.

[0123] Step 2.2: Calculate three types of structural resilience indicators of the road network: node degree centrality, node closeness centrality, and node average edge betweenness centrality.

[0124] Node degree centrality : From a local perspective, it measures the number of available connections (connection redundancy) of a node, and reflects the node's fault tolerance and traffic diversion capabilities in resilience.

[0125]

[0126] Node proximity centrality : From a global perspective, it measures the efficiency of information or traffic dissemination and reflects the node's ability to respond quickly and reconstruct in resilience.

[0127]

[0128] Node average edge betweenness centrality : Measures the ability of a node to control critical path traffic, represents the impact of node failure on the shortest path, and reflects the impact on the vulnerability of the road network structure in resilience.

[0129]

[0130] in, Representation node The degree, Represents the total number of road network nodes in the graph, that is, the scale of the graph. Representation node and nodes The shortest path length between . Representative Node The set of neighbor nodes of node All directly connected nodes. Represents an edge The betweenness centrality refers to the frequency of occurrence of this edge on the shortest path between nodes.

[0131] Step 3: Integrate the aforementioned structural and functional resilience indicators to construct a comprehensive road network resilience index. Integrate the comprehensive resilience index over a period of time to obtain the final resilience assessment result, which is used to quantitatively analyze and evaluate the overall resilience level of the urban transportation system under complex disturbances.

[0132] In order to improve the accuracy of the model's assessment of urban traffic road network resilience, it is necessary to comprehensively consider the changes in the functional and structural indicators of the road network. Therefore, a comprehensive road network resilience index was constructed, and a comprehensive resilience assessment algorithm was designed to accurately calculate the road network resilience loss.

[0133] Step 3.1: Calculate the comprehensive resilience index;

[0134] For the dataset , its inner elements The minimum and maximum standardization method is as follows: As shown,

[0135]

[0136] The calculation method of structural toughness index is as follows: As shown,

[0137]

[0138] The calculation method of functional resilience index is as follows: As shown, Represents the node obtained in step 1 speed.

[0139]

[0140] The calculation method of comprehensive toughness index is as follows: As shown,

[0141]

[0142] The calculation method of comprehensive resilience capacity is as follows: shown.

[0143]

[0144] 、 、 and Node The standardized node degree centrality, node closeness centrality, node average edge betweenness centrality and standardized speed, is the structural toughness index, It is an indicator of functional resilience. For nodes The comprehensive toughness index is the product of the structural toughness index and the structural toughness index. It is a comprehensive resilience indicator of the urban transportation network.

[0145] Step 3.2: Based on the comprehensive resilience index of the urban traffic network, integrate the comprehensive resilience index over a period of time to obtain the final resilience assessment result:

[0146]

[0147] That is arrive The resilience performance of section i.

[0148]

[0149] That is arrive The resilience of the entire road network.

[0150] Comparison of overall performance of prediction models:

[0151] To verify the effectiveness of the proposed prediction model, this example conducts experiments on the public dataset BjTT-1. This dataset contains 1,260 road nodes, covering a period of one month, with a temporal resolution of 4 minutes per slice. The experiment uses 15 historical time steps (one hour) as input and performs multi-step predictions for the next 15 time steps (one hour), evaluating the prediction error at each time step.

[0152] The evaluation indicators used are as follows: Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE).

[0153] Table 1 shows the comparison results with various baseline models, which show the prediction errors for time steps of 20 minutes, 40 minutes, and 60 minutes, as well as the average prediction errors for all time steps. As can be seen, the proposed model achieves superior performance across all evaluation metrics, fully demonstrating the effectiveness of the proposed overall prediction framework.

[0154] The comparison models are described as follows:

[0155] Baseline model: iTransformer [5]It is a time series prediction model based on the Transformer architecture, mainly used for tasks such as traffic flow prediction. Unlike the standard Transformer model, iTransformer is specially optimized for the temporal dependencies of time series data and adopts an enhanced time encoding method. [6] It is a model for time series prediction, especially in complex time series tasks. CATS uses a conditional self-attention mechanism to capture the complex dependencies of time series data. SOFTS [7] It is a model based on self-organizing feature transformation, specially designed for processing multi-dimensional time series data. The SOFTS model can adaptively extract meaningful time series features in different feature dimensions. [8] It is a hybrid model that combines graph neural networks and time series analysis, specifically for processing spatiotemporal data with graph structures such as traffic networks. GTS captures spatial relationships through graph neural networks and uses time series models to model temporal dependencies. [9] It is a graph neural network designed for spatiotemporal data modeling, especially suitable for tasks such as traffic flow prediction. D2STGNN handles spatiotemporal dependencies through a dual-stage model (spatial and temporal modeling).

[10] It is a spatiotemporal interaction decomposition model specifically used for spatiotemporal data modeling, especially in the fields of traffic flow prediction and environmental monitoring.

[11] It is a hybrid model that combines graph convolutional networks and recurrent neural networks for modeling spatiotemporal data. DGCRN can effectively handle spatiotemporal dependencies in tasks such as traffic flow prediction.

[12] It is a time series prediction model based on the Transformer architecture, specifically designed to capture spatiotemporal dependencies over long time periods. STAEformer improves upon the original Transformer model, making it more effective in spatiotemporal series prediction.

[13] STWave is a spatiotemporal prediction model based on the WaveNet architecture, designed for modeling spatiotemporal data (such as traffic flow). STWave leverages WaveNet's temporal modeling capabilities while also incorporating spatial feature modeling.

[0156] Table 1 Comparison results with well-known models

[0157]

[0158] Impact of the Two-Stage Strategy: As shown in Figures 4 and 5, the prediction errors of all road nodes in the training set were ranked from highest to lowest. The top 1%, 2%, 5%, 8%, and 10% of road nodes with the highest prediction errors were analyzed in the test set before and after correction. At all quantiles, the MAE and RMSE after correction were significantly lower than those before correction, indicating that the peak deviation correction module effectively reduced prediction errors. The more extreme the peak, the more significant the benefit. This indicates that the correction module is most sensitive to rare, large traffic peaks and can significantly improve prediction accuracy for extreme scenarios. As shown in Figure 6, the peak deviation correction strategy not only reduced the overall MAE and RMSE but also achieved a relative improvement of 40–50% on nodes with extremely high errors, significantly improving the model's adaptability and resilience to abnormal traffic fluctuations.

[0159] Resilience Assessment: Figure 7 shows the comprehensive resilience index and single speed index curves for node 37 (S50 East Fifth Ring Road Section 1). During most periods, the comprehensive resilience index remained around 0.8 to 0.9, significantly higher than the single speed index, indicating that the section exhibits robust performance across both functional and structural dimensions. However, when traffic accidents occurred—specifically, from 08:32 to 08:44, 09:40 to 09:56 on January 26, 2022, and from 08:04 to 10:40 on January 27, 2022—the comprehensive resilience plummeted to 0.05–0.25 within two minutes of the accident, representing only 20–30% of the single speed index. It remained sluggish for approximately 10–20 minutes after the accident was resolved before gradually recovering. This "deep V"-shaped decline is due to the fact that comprehensive resilience is not only affected by the sudden drop in speed but also by structural shocks such as weakened topological accessibility and changes in node centrality, making it more sensitive to network chain reactions. In contrast, a single speed metric only reflects localized traffic changes and fails to capture the impact of accidents on the recovery of surrounding road sections. This suggests that a comprehensive resilience metric provides earlier warnings and more realistically depicts the impact and recovery trajectory of accidents. As shown in Figure 8, starting at 5:36 PM on January 26, 2022, the S50 East Fifth Ring Road entered a period of sustained disruption. Road closures, temporary traffic controls, and multiple accidents combined led to a significant decline in both the comprehensive resilience score (solid line) and the single speed score (dashed line), with the difference peaking at over 2 points. In particular, during the concentrated accident window of 6:00 PM–7:12 PM on January 27, the comprehensive resilience curve rapidly fell below 14 points, while the speed metric remained around 15 points, highlighting the greater sensitivity of the comprehensive metric to network-level topology disruptions. Subsequently, as obstacles were cleared and controls were gradually lifted, the speed metric rebounded first. However, due to node accessibility and remaining queues, the comprehensive resilience lagged behind by approximately 10–15 minutes before returning to its normal fluctuation range. This phenomenon shows that in resilience assessments of large road sections, relying solely on speed information can easily underestimate the chain reaction caused by accidents. The comprehensive resilience index can more accurately capture the combined impact of road closures and accidents on the overall network structure and function, providing a more reliable basis for scheduling resources and evaluating recovery effects.

[0160] The above content is a further detailed description of the present invention, and it cannot be considered that the specific implementation methods of the present invention are limited to this. For ordinary technicians in the technical field to which the present invention belongs, they can make several simple deductions or substitutions without departing from the concept of the present invention, which should be regarded as belonging to the scope of protection of the invention determined by the claims of the present invention.

[0161] References are as follows:

[0162] [1]Hosseinian, Seyed Mohsen, and Hamid Mirzahossein. "Efficiency andsafety of traffic networks under the effect of autonomous vehicles." Iranian Journal of Science and Technology, Transactions of Civil Engineering 48.4(2024): 1861-1885.

[0163] [2]He, Yuxin, et al. "In-depth insights into the application ofrecurrent neural networks (rnns) in traffic prediction: A comprehensivereview." Algorithms 17.9 (2024): 398.

[0164] [3]Liu, Jinbo, et al. "GNN-based long and short term preferencemodeling for next-location prediction." Information Sciences 629 (2023): 1-14.

[0165] [4]Sun, Yutao, et al. "Retentive network: A successor to transformerfor large language models." arXiv preprint arXiv:2307.08621 (2023).

[0166] [5]Liu, Yong, et al. "itransformer: Inverted transformers areeffective for time series forecasting." arXiv preprint arXiv:2310.06625 (2023).

[0167] [6]Kim, Dongbin, et al. "Are Self-Attentions Effective for TimeSeries Forecasting?." arXiv preprint arXiv:2405.16877 (2024).

[0168] [7]Han, Lu, et al. "Softs: Efficient multivariate time seriesforecasting with series-core fusion." arXiv preprint arXiv:2404.14197 (2024).

[0169] [8]Shang, Chao, Jie Chen, and Jinbo Bi. "Discrete graph structurelearning for forecasting multiple time series." arXiv preprint arXiv: 2101.06861 (2021).

[0170] [9]Shao, Zezhi, et al. "Decoupled dynamic spatial-temporal graphneural network for traffic forecasting." arXiv preprint arXiv:2206.09112 (2022).

[0171]

[10] Shao, Zezhi, et al. "Spatial-temporal identity: A simple yeteffective baseline for multivariate time series forecasting." Proceedings of the 31st ACM international conference on information&knowledge management .2022.

[0172]

[11] Li, Fuxian, et al. "Dynamic graph convolutional recurrent networkfor traffic prediction: Benchmark and solution." ACM Transactions on Knowledge Discovery from Data 17.1 (2023): 1-21.

[0173]

[12] Liu, Hangchen, et al. "Spatio-temporal adaptive embedding makesvanilla transformer sota for traffic forecasting." Proceedings of the 32nd ACM international conference on information and knowledge management . 2023.

[0174]

[13] Fang, Yuchen, et al. "When spatio-temporal meet wavelets:Disentangled traffic forecasting via efficient spectral graph attentionnetworks." 2023 IEEE 39th international conference on data engineering (ICDE) .IEEE, 2023.

Claims

1. A method for evaluating urban traffic resilience based on multi-source heterogeneous data and a large language model, characterized by: The following steps are involved: Step 1: Functional resilience assessment phase: construct functional indicators based on raw traffic data and text information; The method consists of two phases: the first phase uses raw traffic data and text information to make preliminary traffic status predictions; the second phase performs peak correction on the prediction results of the first phase to improve the prediction accuracy during high-disturbance periods; Step 2: Structural resilience assessment: Based on the original traffic data and the topological structure of the transportation network, structural resilience indicators are calculated. These structural resilience indicators include node degree centrality, node closeness centrality, and node average edge betweenness centrality. Step 3: Integrate the structural and functional resilience indicators to construct a comprehensive road network resilience index. Integrate the comprehensive resilience index over a period of time to obtain the final resilience assessment result, which is used to quantitatively analyze and evaluate the overall resilience level of the urban transportation system under complex disturbances. Step 2 is as follows: Based on the original historical traffic speed and congestion data, the real-time spatial physical topology of the traffic network is constructed ; Then the structural indicators are calculated, namely the three types of structural resilience indicators: node degree centrality, node closeness centrality and node average edge betweenness centrality; The specific steps are as follows: Step 2.1: Construct the spatial physical topology of the traffic network and model the physical topology of the original traffic network; Step 2.2: Calculate three types of structural resilience indicators of the road network: node degree centrality, node closeness centrality, and node average edge betweenness centrality; Node degree centrality : From a local perspective, it measures the number of available connections of a node, and reflects the node's fault tolerance and traffic diversion capabilities in resilience; Node proximity centrality : From a global perspective, it measures the efficiency of information or traffic dissemination and reflects the node's ability to respond quickly and reconfigure in resilience; Node average edge betweenness centrality : Measures the ability of a node to control critical path traffic, represents the impact of node failure on the shortest path, and reflects the vulnerability of the road network structure in resilience; in, Representation node The degree, represents the total number of road network nodes in the graph, Representation node and nodes The shortest path length between Representative Node The set of neighbor nodes of Represents an edge Betweenness centrality; Step 3 is as follows: Taking into account the changes in functional and structural indicators of the road network, a comprehensive road network resilience index was constructed, and a comprehensive resilience assessment algorithm was designed to accurately calculate the road network resilience loss; Step 3.1: Calculate the comprehensive resilience index; For the dataset , its inner elements The minimum and maximum standardization method is as follows: As shown, The calculation method of structural toughness index is as follows: As shown, The calculation method of functional resilience index is as follows: As shown, Represents the node obtained in step 1 speed; The calculation method of comprehensive toughness index is as follows: As shown, The calculation method of comprehensive resilience capacity is as follows: As shown, 、 、 and Node The standardized node degree centrality, node closeness centrality, node average edge betweenness centrality and standardized speed, is the structural toughness index, It is an indicator of functional resilience; For nodes The comprehensive toughness index is the product of the structural toughness index and the structural toughness index. It is a comprehensive resilience indicator of the urban transportation network; Step 3.2: Based on the comprehensive resilience index of the urban traffic network, integrate the comprehensive resilience index over a period of time to obtain the final resilience assessment result: That is arrive The resilience performance of section i, That is arrive The resilience of the entire road network.

2. According to claim 1, the urban traffic resilience assessment method based on multi-source heterogeneous data and large language model is characterized by: Step 1 specifically includes: Step 1.1: Phase 1: Preliminary prediction of traffic speed; The first stage predicts preliminary traffic speeds based on raw traffic data and text information. The event text is encoded using a pre-trained large language model to obtain high-dimensional semantic features. These high-dimensional semantic features are then fused with the raw traffic speed sequence and fed into the spatiotemporal feature extractor ST-RetNet. The ST-RetNet module uses the RetNet architecture for both spatial and temporal scales. Subsequently, two convolutional layers are used to output preliminary traffic speed predictions. ; Step 1.2: Second stage: Use the peak deviation calculation algorithm to extract the cumulative average error for each road; input this deviation value, along with the periodic factor and the high-dimensional semantic features of the text, into the gated fusion module to adjust the initial prediction results; and output the final future traffic speed prediction value.

3. The urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model according to claim 2 is characterized by: Step 1.1 specifically includes: Step 1.1.1: Construct a two-dimensional time encoding for each time step in the dataset and concatenate it with the feature dimension of the original velocity vector as the exogenous temporal context feature of the network ; First, construct a two-dimensional time encoding for each time step in the dataset: (1) (2) in, represents a standardized encoding method for the time of day, where is the timestamp, is the total number of time periods in a day; Indicates the encoding method of the day of the week; At the corresponding road node and time slice and Concatenate with the feature dimension of the original velocity vector as the exogenous temporal context feature of the network , where N is the number of road nodes in the road network and T is the time step; Step 1.1.2: Encode text information and generate temporal text features ; Assume the event text sequence: The text corresponding to the t-th prediction time step, For words in the text, is the length of the text; the text is input into the pre-trained large language model, and the model calculates and generates the corresponding hidden state based on the text content: Next, the convolution operation is used to process the hidden state sequence, reducing the token dimension to 1 and generating the final time text feature. ; Step 1.1.3: Combine the temporal text features with the road nodes where the events occurred; Assume the event node index set is: in, , is the number of road nodes in the road network, Indicates that an event occurred on road k; according to The temporal text features for constructing the road network are: For event node index collection Each road node in , the time text feature Assign to ; and They are all three-dimensional features, and their three dimensions represent time and position in turn , Road Index and characteristic dimensions; Step 1.1.4: Concatenate exogenous temporal context features and temporal text features to obtain feature vectors ; Exogenous temporal context features After the convolution operation, the feature dimension is expanded to 64 dimensions; the time text feature obtained in step 1.1.3 The feature dimension is also expanded to 64 dimensions through convolution operation; the two features are spliced ​​according to the feature dimension to obtain the feature vector ; Eigenvector Provided to the spatiotemporal feature extractor; Step 1.1.5: Build a spatiotemporal feature extractor and process feature vectors , get the preliminary traffic speed results; The number of layers of the spatiotemporal feature extractor ST-RetNet is one or more layers in series; The spatiotemporal feature extractor includes a spatial module S-RetNet and a temporal module T-RetNet. The two modules adopt a spatial-temporal serial approach. The output of the spatial module Input time module, output of time module Pass in two convolution operation layers to get the prediction results .

4. The urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model according to claim 3 is characterized by: The spatial modules are as follows: The input of the spatial module is and the adjacency matrix set ,in, is the inherent adjacency matrix of the road network, is the adaptive adjacency matrix; Adaptive adjacency matrix Composed of learnable parameters; randomly initialized parameters to learnable column vectors and , the adaptive adjacency matrix is ​​generated as follows: in, and is the activation function; The spatial module S-RetNet includes: GCN, spatial preservation mechanism, spatial feature aggregation, gated multi-scale spatial preservation mechanism and feedforward network; Among them, GCN is expressed as follows: is the activation function, and is the parameter matrix, is the input of S-RetNet, and same; Space preservation mechanism: It is defined as follows, where three linear layers are completed respectively 、 operate: in and are mutually conjugate complex numbers, , and is the parameter matrix, is the Hadamard product; Spatial feature aggregation: Information aggregation is performed according to the following formula: is a one-dimensional convolution, For splicing operation; Gated multi-scale preservation mechanism: scale Analyze the input sequence Specifically, the corresponding multi-scale calculation results are given according to formulas (7) to (10): Then, the multi-scale retention mechanism is calculated according to formula (11): , and finally use the gated structure To increase the nonlinearity of the model; in is the normalization operation, and is the parameter matrix, is the activation function, Output results preserved for spatial multi-scale; Feedforward network: Process as follows, output The final output of S-RetNet is: in and is the parameter matrix, SiLU is the activation function; Output of the spatial module and input Provided to the time module.

5. The urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model according to claim 3 is characterized by: The time module is as follows: The temporal module T-RetNet includes a temporal retention mechanism, a gated multi-scale temporal retention mechanism, and a feedforward network; Among them, the time retention mechanism introduces , which combines the causal mask and the exponential decay of relative distance into a matrix; the gated multi-scale time preservation aggregates the features from each head and fully captures the temporal dependencies; finally, the output of T-RetNet is obtained through the feedforward network ; The time retention mechanism is defined as follows, where 、 For linear layer operation: in ; and Mutually conjugate complex numbers; , and is the parameter matrix; It is a relative position coding matrix that gradually decays as the time interval increases; Gated multi-scale time preservation mechanism: at multiple scales Analyze the input sequence , and the corresponding multi-scale calculation results are given according to formula (13) ; Calculate the multi-scale retention mechanism according to formula (14) , and then use the gated structure To increase the nonlinearity of the model; as follows: in Output results preserved for multiple time scales; Feedforward network: Process as follows, output The final output of T-RetNet is: in and is the parameter matrix.

6. The urban traffic resilience assessment method based on multi-source heterogeneous data and a large language model according to claim 2 is characterized by: Step 1.2 specifically includes: Step 1.2.1: Peak deviation calculation algorithm; First, the basic model predicts data in batches, with each batch as a sample; the sample size is set according to the device and model scale; For the training set, the predicted values ​​and true values ​​obtained by the first-stage basic model are expressed as and , representing the samples at the prediction time step and the road node number is The traffic speed prediction and true value on Corresponding to the first stage output The absolute error at each time period and node is calculated using the following formula: Then, the error is periodically mapped in the time dimension to obtain the periodic average error matrix , which is calculated as: in, For the corresponding time and nodes The cumulative number matrix of the upper error, is the number of time slices in a day; Finally, the predicted value of the basic model and the average error are corrected for the peak value at the corresponding time and node: in, Indicates time and nodes superior, In all samples The value of Step 1.2.2: Time period feature encoding; Based on the corrected prediction results, the periodic time dimension is introduced as an auxiliary feature: Time of day: ; Day of the week: ; Compare it with the corrected speed Perform splicing to construct a three-dimensional periodic time feature tensor: Step 1.2.3: Text feature alignment and gated fusion; The time feature vector With text feature vector Fusion via gating mechanism: in, is the parameter matrix, represents the activation function, Represents a splicing operation, represents element-wise multiplication, is the final fused representation vector.

Citation Information

Patent Citations

  • Traffic toughness evaluation method based on diffusion probability model data reconstruction

    CN117765718A

  • Traffic demand estimation method based on multi-source traffic big data and computational graph

    CN118822185A