Cross-domain-oriented urban traffic road network toughness pre-judgment and evaluation method

Abstract

The invention relates to the field of intelligent traffic, and provides a cross-domain-oriented urban traffic road network toughness pre-judgment and evaluation method, which comprises the following three steps: step 1, based on input source domain and target domain urban traffic data, constructing a space-time cross-domain traffic prediction model, realizing traffic state transfer learning and generalization prediction among different cities, and establishing a space-time cross-domain traffic prediction model; generating traffic prediction in a cross-domain scene and a corresponding traffic map structure; 2, designing a hysteretic toughness index in combination with a prediction result and a traffic map structure, and fusing structure-function toughness and a hysteretic toughness factor to form a toughness index capable of reflecting a road network performance evolution mechanism; and step 3, based on the hysteretic toughness index, establishing a space-time fine-grained toughness quantization algorithm for quantizing node-level toughness of the road network. According to the method, the three major problems of lack of pre-judgment and cross-domain generalization ability, single toughness index representation and rough toughness quantification of an existing evaluation method are effectively solved, and the perspectiveness, the adaptability and the accuracy of road network toughness evaluation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent transportation, in particular to a cross-domain-oriented urban traffic network resilience prediction and evaluation method. BACKGROUND

[0002] As a key infrastructure, urban traffic network is an important support for ensuring the normal operation of the city and the development of social economy. With the expansion of the city and the growth of traffic demand, the road network is more vulnerable to large-scale service capacity decline when facing extreme weather, sudden accidents or local failures. Therefore, how to accurately predict and evaluate the resilience of the road network under the disturbance situation is of great significance for resource scheduling and emergency response. However, there are differences in road network structure and operation characteristics between different cities, and the resilience evaluation method constructed in a single scene is difficult to be directly migrated to other cities. Further, it is of great significance to study a cross-domain-oriented urban traffic network resilience prediction and evaluation method to improve the forward-looking and generalization of the model, and to urban traffic management.

[0003] Existing traffic network resilience evaluation methods mainly include three types: topological model, simulation model and data-driven model. Topological model focuses on the structure of the road network, and uses structural indicators such as network connectivity, node centrality and shortest path to represent the resilience of the road network. However, in real scenarios, structural resilience indicators usually only change significantly when large-scale structural damage (such as earthquakes, floods, etc.) occurs, and it is difficult to effectively evaluate the resilience of the road network under general traffic accidents or sudden congestion and other functional disturbances.

[0004] Simulation model introduces node attack, link failure and other simulation strategies, and combines functional indicators such as traffic speed, flow or occupancy to quantify the resilience of the road network. However, such methods usually need to rely on a large number of preset assumptions and are mainly used for fixed road networks, which are difficult to realistically depict the real evolution of the road network under disturbance situations, resulting in poor flexibility and scalability.

[0005] With the popularity of deep learning techniques and large-scale traffic sensor data, data-driven methods have become mainstream. Based on spatio-temporal traffic data, these methods combine convolutional neural networks, graph neural networks, and neural differential equations to model road network states. They not only accurately depict the evolution of road networks during disturbances but also significantly improve the accuracy of resilience quantification. However, these methods still face three common problems. First, they lack predictive ability and cross-domain generalization. Existing methods are mostly based on historical observation data for post-evaluation, which cannot effectively predict future road network resilience. Moreover, they can only analyze a single city or a fixed road network scenario, lacking cross-domain migration and generalization ability. Second, the resilience index representation is single-dimensional. The road network has a hysteresis effect during disturbances, and existing methods generally ignore the evolution mechanism of road network performance with disturbance intensity, leading to distorted resilience index representation. Third, there is a lack of spatio-temporal fine-grained resilience quantification. Existing methods usually focus on the overall change of road network performance at the beginning and end of the disturbance, without conducting fine-grained analysis by stages (such as resistance and recovery). In the spatial dimension, they do not refine to the node or road segment level, making it difficult to identify local vulnerable units and limiting the guiding value of evaluation results for resource scheduling and fine-grained management. SUMMARY

[0006] To solve the problems of existing urban traffic road network resilience evaluation methods, such as lack of predictive ability and cross-domain generalization, single-dimensional resilience index representation, and lack of spatio-temporal fine-grained resilience quantification, the present invention proposes a cross-domain-oriented urban traffic road network resilience prediction and evaluation method, including a spatio-temporal cross-domain traffic prediction model, a hysteresis resilience index, and a spatio-temporal fine-grained resilience quantification algorithm. The spatio-temporal cross-domain traffic prediction model uses a pre-training-fine-tuning framework. It first pre-trains on source domain urban traffic data to learn general spatio-temporal features, then fine-tunes on target domain data while inheriting the source domain spatio-temporal pattern, achieving traffic prediction in cross-city scenarios. The hysteresis resilience index is constructed based on the prediction results, integrating structural, functional resilience indices, and hysteresis resilience factors to comprehensively represent multi-dimensional resilience characteristics during disturbances. The spatio-temporal fine-grained resilience quantification algorithm considers the performance retention rate, average resistance index, and maximum recovery rate of the road network, segmentally depicts the disturbance process in the time dimension, and refines to the node level in the spatial dimension, dynamically outputting fine-grained resilience evaluation results.

[0007] The present invention is achieved by the following technical solutions: A cross-domain-oriented urban traffic road network resilience prediction and evaluation method, comprising the following steps: Step 1: Based on source domain and target domain urban traffic data, construct a spatio-temporal cross-domain traffic prediction model to generate traffic prediction and traffic graph structure in cross-city scenarios; The construction of the spatio-temporal cross-domain traffic prediction model includes a pre-training stage and a fine-tuning stage. The model is first pre-trained on the traffic data of the source domain city to learn general spatio-temporal features, and then fine-tuned on the data of the target domain. The two stages share common-heterochronous temporal module, lightweight spatial memory module, and decoder module. In addition, a fine-grained spatio-temporal distiller module is introduced in the fine-tuning stage.

[0008] Thus, the working process of the spatio-temporal cross-domain traffic prediction model is as follows: First, the historical speed data of the source domain city is selected for pre-training. The common-heterochronous temporal module, lightweight spatial memory module, and decoder are used to capture the general pre-training spatio-temporal features of the source domain city.

[0009] Then, the historical speed data of the target domain city is selected for fine-tuning. The same modules are used to capture the fine-tuning spatio-temporal features of the target domain city, and the fine-grained spatio-temporal distiller is used to effectively transfer the source domain features to the target domain.

[0010] Finally, the speed prediction and corresponding traffic graph structure of the target domain city under the disturbance scenario are generated.

[0011] Step 2: According to the prediction results and traffic graph structure, design the hysteresis resilience index. The hysteresis resilience index includes structure-function resilience and hysteresis resilience factor.

[0012] The establishment process is as follows: First, consider the spatial structure of the road network and the traffic speed characteristics, design the structure-function resilience index. Second, analyze the hysteresis characteristics of the road network in the disturbance phase, design the hysteresis resilience factor. Finally, integrate the hysteresis resilience factor into the structure-function resilience to construct the hysteresis resilience index. Step 3: Based on the hysteresis resilience index, establish a spatio-temporal fine-grained resilience quantification algorithm, and output the resilience results.

[0013] First, according to the target domain city traffic speed label value, target domain city traffic speed prediction value, and disturbance event, calculate the hysteresis resilience index under non-disturbance and under disturbance .

[0014] Second, based on and , quantify the performance retention rate of the road network in the disturbance phase from a macroscopic perspective .

[0015] Then, calculate the average resistance index and maximum recovery rate .

[0016] Finally, weight and integrate the three parts of the resilience quantification results to obtain , wherein, , and is a weight coefficient, R is the final spatiotemporal fine-grained resilience quantification result.

[0017] Advantages The present application aims at the three major problems faced by the existing urban traffic network resilience evaluation method: lack of predictability and cross-domain generalization ability, single resilience index representation dimension, and lack of fine-grained resilience quantification. Based on deep learning technology, the adaptability and accuracy of network resilience evaluation in cross-city market scenarios are effectively improved, providing guidance for resource scheduling, emergency scheduling and strategy optimization, as follows: (1) Based on the pre-training-fine-tuning framework, a spatiotemporal cross-domain traffic prediction model is designed, which fully utilizes the source domain spatiotemporal pattern for transfer learning to improve its traffic prediction ability for the target domain, realizing the predictability and cross-domain generalization of resilience evaluation.

[0018] (2) On the basis of traditional structure-function resilience indicators, a hysteresis resilience factor is designed, and the two are fused to build a hysteresis resilience indicator, which more comprehensively and accurately represents the dynamic evolution of the network during the disturbance period.

[0019] (3) A spatiotemporal fine-grained resilience quantification algorithm is proposed, which segments the whole disturbance process in the time dimension and refines to the node level in the space dimension, realizing the resilience evaluation of local fragile units. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is the flowchart of the cross-domain urban traffic network resilience prediction and evaluation method of the present application; Figure 2 is the structural schematic diagram of the spatiotemporal cross-domain traffic prediction model of the embodiment of the present application; Figure 3 is the visual traffic node diagram of the urban traffic network of the embodiment of the present application; Figure 4 is the hysteresis evolution mechanism diagram of the network performance of the embodiment of the present application; Figure 5 is the network resilience evaluation result diagram of the embodiment of the present application in San Jose, USA; Figure 6 is the network resilience evaluation result diagram of the embodiment of the present application in Beijing, China. DETAILED DESCRIPTION

[0021] The technical solutions provided by the present application will be further described below in combination with specific embodiments and their drawings. The advantages and features of the present application will be clearer in combination with the following description.

[0022] A cross-regional urban transportation network resilience prediction and assessment method, such as Figure 1 As shown, it includes the following steps: Step 1: Based on urban traffic data from the source and target domains, construct a spatiotemporal cross-domain traffic prediction model to generate traffic predictions and traffic map structures for cross-city scenarios; Step 2: Based on the prediction results and traffic map structure, design the hysteresis resilience index; Step 3: Based on the hysteresis toughness index, establish a spatiotemporal fine-grained toughness quantification algorithm and output the toughness results.

[0023] Example Application scenario: Resilience prediction and assessment of cross-regional urban transportation networks Step 1: Construct a spatiotemporal cross-domain traffic prediction model, such as Figure 2 As shown.

[0024] The design of the spatiotemporal cross-domain traffic prediction model includes a pre-training stage and a fine-tuning stage. First, it is pre-trained on source domain urban traffic data to learn general spatiotemporal characteristics. Then, the fine-tuning stage adapts the model to the target domain data. Both stages share a common-to-differential time-series module (…). Figure 2 (Middle ①), lightweight spatial memory module ( Figure 2 (②) and decoder module ( Figure 2 (Middle ③). Simultaneously, a refined spacetime distillation module is additionally introduced during the fine-tuning phase ( Figure 2 (4)

[0025] The working process of the spatiotemporal cross-domain traffic prediction model: First, historical velocity data of source cities are selected for pre-training. The general pre-trained spatiotemporal features of source cities are captured through a constant-dissimilar time series module, a lightweight spatial memory module, and a decoder.

[0026] Secondly, historical speed data of cities in the target domain are selected for fine-tuning. The same modules mentioned above are used to capture the fine-tuned spatiotemporal characteristics of cities in the target domain. At the same time, the source domain features are effectively transferred to the target domain through a refined spatiotemporal distillation device.

[0027] Finally, the speed prediction and corresponding traffic map structure of the target domain city under the disturbance scenario are generated.

[0028] In this embodiment, the source city data is speed data from the California road network, provided by the California Department of Transportation, and includes historical observations from hundreds of different spatial sensors. The target city data is speed data from the Beijing road network, provided by the China Transportation Authority, and also includes historical observations from hundreds of different spatial sensors.

[0029] Step 1.1: Construct the traffic map structure; Based on the theory of topological network structure, the road sensors are regarded as the nodes of the traffic network, and the connection relationship between the sensors is regarded as the edge of the traffic network, thereby constructing the spatial physical topology structure of the traffic network, i.e., the traffic graph structure. The traffic graph can be defined as: (1) In the formula, represents a set of nodes; is a set of edges connecting the nodes; is an adjacency matrix, when the node is connected with the node , then ; otherwise, . Based on , the spatiotemporal traffic data can be represented as a sequence ; wherein is the length of the time sequence; is the traffic feature (such as speed, flow, and occupancy).

[0030] For the source domain and the target domain city, the traffic graph structure is constructed.

[0031] Taking the San Jose city road network in the United States as an example, the traffic graph visualization is shown in Figure 3 . The maximum rainfall information of the regional position is used as the attribute feature of the traffic node, which can be used for the road network evaluation when the traffic node appears rainfall.

[0032] Step 1.2: Normal-Abnormal Time Sequence Module; The urban traffic network not only exists periodicity and other normal time dependence, but also often accompanies the sudden fluctuation superimposed by the perturbation effect. In order to fully capture the normal and abnormal time dependence of the traffic mode, a normal-abnormal time sequence module is designed, including: time scale decomposition, patch processing, global time attention convolution, and gated time convolution. Step 1.2.1: Time Scale Decomposition; The traffic time sequence generally presents the multi-scale characteristics, including the long-term seasonal change and the short-term trend change. Among them, the long-time seasonality embodies the periodicity and other normal traffic state of the road network, and the short-time trend reflects the abnormal traffic state of the nonlinear fluctuation caused by the disturbance. In order to facilitate the capture of the multi-scale time dependence, the scale decomposition needs to be performed first.

[0033] Specifically, the fast Fourier transform is performed on the spatiotemporal traffic data , the spatiotemporal data X is converted into a frequency domain signal, and then the significant frequency components with the top K amplitudes are selected, and the inverse fast Fourier transform is used to restore them into low-frequency sequences and high-frequency sequences, which is represented as:​​ (2) (3) (4) (5) (6) In formula (2), represents the spatio-temporal traffic data (source domain or target domain); is a fast Fourier transform function; is a frequency domain signal. In formula (3), is an amplitude function; represents the significant frequency domain components selected before the amplitude; K In formula (4), represents a frequency domain subset composed of the pre-selected K components. In formula (5), is an inverse fast Fourier transform function; represents a low-frequency sequence, reflecting the long-time characteristics of the time sequence, and is used to characterize the normal traffic state. In formula (6), is a high-frequency sequence, focusing on the short-term irregular fluctuations of the time sequence, and is used to characterize the abnormal traffic state.

[0034] Step 1.2.2: Patch processing; In the field of traffic prediction methods, 'Patch' specifically refers to a local data unit formed after structuring and blocking traffic time series data, containing traffic state characteristics (such as flow, speed, density, etc.) within a continuous time window, with clear time series continuity and feature aggregation.

[0035] Since the original input spatio-temporal traffic data often covers a long time span, directly training it will lead to feature redundancy and high computational complexity. Therefore, the present application performs Patch processing on the low-frequency sequence and the high-frequency sequence , that is, the continuous sequence ( is or ) is divided into coarse-grained time blocks. This operation helps to improve the efficiency of long sequence processing while preserving the key information of the sequence, represented as: (7) In formula (7), represents the Patched sequence; is the window length of the Patch; , represents the number of Patch blocks.

[0036] Secondly, use linear layer Map the Patch sequence to the dimension required by the model , denoted as: ,(8) Further, in order to preserve the periodicity and trend context information of the spatio-temporal data, an embedding term is introduced for each Patch block , which represents the temporal position information of the Patch.

[0037] It should be noted that, is the linear mapping result of the original Patch sequence, only containing the mapped feature information, and does not contain the order relationship of the Patch in the overall time sequence, i.e. the temporal position information, which is represented by .

[0038] The temporal information covers the periodic characteristics of the Patch corresponding data in the time sequence dimension, including daily and weekly cycles. The position information refers to the order relationship of the Patch in the overall time sequence, which is used to reflect the time sequence logic between Patches to provide cross-Patch context information. Denoted as: ,(9) ,(10) Equation (9) represents the broadcast operation; equation (10) represents feature fusion, represents the Patch sequence superimposed with position information, including the Patched low-frequency sequence and the Patched high-frequency sequence .

[0039] Step 1.2.3: Global time attention convolution; Patched low-frequency sequence has a low change frequency within a unit of time, and focuses on reflecting the normal periodic changes of the road network, and needs to capture its time dependence from a global perspective. Therefore, the present application designs a global time attention convolution, and combines a (source domain or target domain) normal time pattern library to explicitly supplement the regular periodic information of the traffic pattern, which is represented as: ,(11) Equation (11) represents the Hadamard product between the Patched low-frequency sequence and the (source domain or target domain) normal time pattern library .

[0040] wherein, is for adaptive matrix. In the model training stage, through the dynamic parameter updating mechanism, based on the mode (such as the typical flow change of the morning and evening peak of the working day, the stable speed distribution of the regular period, etc.) conforming to the normal characteristics in the input traffic data, combined with the error back propagation gradient of the model prediction and the real normal performance, the matrix parameters are iteratively updated by using the Adam optimizer, so that the matrix can adaptively capture and store different types of traffic normal mode characteristics. is for explicitly updated Patched low-frequency sequence.

[0041] ,(12) ,(13) ,(14) in formula (12), is a trainable parameter; respectively represent Query, Key and Value matrices, which are respectively used to learn the feature mapping of query, key and value between time blocks.

[0042] in formula (13), is an activation function; is the transpose of the Key matrix; is the feature dimension; represents an attention matrix.

[0043] Formula (14) is a global time attention convolution operation . represents the normal time dependence of (source domain or target domain) captured via .

[0044] Step 1.2.4: gated time convolution; Patched high-frequency sequence has a high change frequency in unit time, and focuses on reflecting the intense abnormal fluctuations of the road network under disturbance, and needs to fully capture its time dependence from a local perspective. Therefore, the invention designs a gated time convolution, and combines the (source domain or target domain) abnormal time mode library to explicitly supplement the abnormal fluctuation information of the traffic mode, and the process is represented as: ,(15) Formula (15) represents the Hadamard product of the Patched high-frequency sequence and the (source domain or target domain) abnormal time mode library . , The same is an adaptive matrix. In the model training stage, through the dynamic parameter updating mechanism, the abnormal characteristics deviating from the normal mode library in the input traffic data (such as sudden traffic surge, abnormal speed drop, etc.) are continuously optimized, combined with the back propagation gradient of the model prediction error, and the matrix parameters are iteratively updated by using the Adam optimizer, so that the matrix can adaptively capture and store different types of traffic abnormal mode characteristics. is trained Patch high-frequency sequence after explicit adaptive update.

[0045] ,(16) In formula (16), is a dilated convolution operation; represents an inflation factor; is the size of the convolution kernel; ,(17) In formula (17), is a gated time convolution operation; and are activation functions; represents a time convolution operation; represents a high-frequency sequence captured (source domain or target domain) abnormal time dependence.

[0046] Then, the normal and abnormal time dependence is fused to obtain comprehensive (pre-training or fine-tuning) time features: , (18) In formula (18), represents feature fusion; represents pre-training time features (pre-training stage) or fine-tuning time features (fine-tuning stage).

[0047] Step 1.3: lightweight space memory module; The existing method often faces two shortcomings in the space feature modeling method: first, the memory ability of the space dependence between traffic nodes is weak, and it is difficult to continuously pay attention to the key space mode information before and after the disturbance occurs. Second, in the pre-training-fine-tuning cross-domain scene, when facing large-scale traffic road network nodes, the traditional model has large parameter quantity, high training cost and limited migration generalization ability.

[0048] Therefore, the lightweight space memory module is designed to enhance the dynamic memory ability of the space relationship and reduce the parameter overhead.

[0049] The lightweight space memory module comprises a space memory module and a LoRA enhanced training strategy.

[0050] Step 1.3.1: Spatial Memory Module; For the (pre-trained or fine-tuned) temporal features captured by constant-discrete time series modules The spatial memory module is used to further extract spatial features between traffic nodes and, in conjunction with a spatial pattern library (source or target domain), strengthens the model's memory of spatial relationships, as shown below: (19) (20) ,(twenty one) In equation (19), These are trainable parameters; Let represent the Query, Key, and Value matrices, respectively, used to learn the feature mappings of the query, key, and value in the spatial pattern. Equation (20) represents the acquisition of intermediate layer spatial dependencies with attention mechanism. In equation (21), For activation functions; , ,and These are trainable parameters; This indicates the spatial dependencies captured by the spatial memory module; For spatial pattern library, .

[0051] Specifically, the spatial pattern library is obtained by combining the traffic speed characteristics of different spatial locations (such as different road segments, intersections, and areas) in the input traffic data with the backpropagation gradient of the error between the model's spatial traffic speed prediction results and the actual situation, and iteratively updating the matrix parameters using the Adam optimizer. The spatial pattern library stores diverse spatial relationship templates of the road network and can be adaptively updated during training, helping the model share key spatial dependency patterns when facing different disturbance events and different urban scenarios.

[0052] Step 1.3.2: LoRA-enhanced training strategy; Due to the trainable matrices of Query and Value Large-scale spatial graphs involve a large number of parameters, leading to high model training costs. Therefore, this invention introduces LoRA low-rank decomposition to reduce the number of trainable parameters and lower the transfer learning cost across cities. The low-rank decomposition is expressed as: ,(twenty two) (twenty three) In equation (22), denotes a low-rank matrix; in formula (23), and is an incremental term.

[0053] Based on the low-rank matrix and the incremental term, the original weight is updated , denoted as: (24) Further, the enhanced Query and Value are denoted as: ,(25) Finally, the spatial dependency of the LoRA enhanced training is changed from formula (20) and formula (21) to formula (26) and formula (27): ,(26) ,(27) In formula (26), is the intermediate layer spatial dependency optimized by LoRA; in formula (27), denotes the pre-training spatial feature captured through the lightweight spatial memory module or the fine-tuning spatial feature .

[0054] Step 1.4: Fine-tuning spatio-temporal distiller; In the pre-training-fine-tuning cross-domain scenario, the time and spatial features learned by the source domain often mix noise that is not available for the target domain city, and direct migration can reduce the adaptability of the model. Therefore, the present application designs a fine-tuning spatio-temporal distiller to improve the consistency and effectiveness of the spatio-temporal features in the pre-training and fine-tuning stages, including: pre-training-fine-tuning spatio-temporal distillation and pre-training spatio-temporal feature refinement.

[0055] Step 1.4.1: Pre-training-fine-tuning spatio-temporal distillation; Distillation using time feature similarity: the similarity of the pre-training time feature and the fine-tuning time feature obtained by formula (18) is evaluated using the KL divergence loss, denoted as: ,(28) In formula (28), promote and feature alignment; M is the number of samples.

[0056] Distillation with spatial feature similarity: the pre-trained spatial features are enhanced using the contrastive learning loss evaluation formula (27) and fine-tune the consistency of spatial features , denoted as: ,(29) In formula (29), is a similarity measure function; denotes the exponential function; can effectively align and .

[0057] By minimizing the distillation loss with time feature similarity and the distillation loss with spatial feature similarity, the updated distillation time feature and distillation spatial feature are obtained, which provide input for subsequent pre-training spatio-temporal feature refinement.

[0058] Step 1.4.2: pre-training spatio-temporal feature refinement; The time features and spatial features captured during pre-training are refined using adaptive filters: ,(30) In formula (30), and denote adaptive time and spatial filters, respectively, , are element-wise multiplied with time features and spatial features , respectively, and the parameters are optimized in a task-oriented manner during pre-training, automatically identifying and discarding noise and retaining relevant features. Specifically, during the pre-training process, the filter parameters and are jointly optimized with the task-related loss function (KL divergence loss, contrastive learning loss). The parameter update direction is determined by the task objective function, thereby automatically suppressing noise and retaining relevant features.

[0059] Step 1.5: decoder The decoder is used to adjust the number of feature channels and output the traffic prediction result. The decoder of the present application is composed of multiple layers of perceptrons . Taking the decoder of the fine-tuning stage as an example, it is denoted as: ,(31) In formula (31), denotes the Concat operation; This is the cross-city traffic prediction result output by the spatiotemporal cross-domain traffic prediction model (in this embodiment, it is the traffic speed prediction result). The subsequent step 2 can construct a hysteresis resilience index based on this speed prediction result.

[0060] Step 2: Design hysteresis toughness parameters, such as Figure 1 As shown.

[0061] The construction of the hysteresis toughness index includes structural-functional toughness and hysteresis toughness factor.

[0062] The process of establishing the hysteresis resilience index is as follows: First, considering both the road network spatial structure and traffic speed characteristics, a structural-functional resilience index is designed. Second, the hysteresis characteristics of the road network during disturbance phases are analyzed, and a hysteresis resilience factor is designed. Finally, the hysteresis resilience factor is integrated into the structural-functional resilience to construct the hysteresis resilience index.

[0063] Step 2.1: Structural-functional resilience index; Traditional resilience indices focus on a single structural or functional attribute of the road network, failing to comprehensively reflect the coupling effect between the two. Therefore, this invention constructs a structure-function resilience index, which includes speed-coupled node accessibility, speed-coupled node importance, local diffusion congestion index, speed-coupled connectivity integrity, and speed fluctuation index, defined as follows: (32) (33) (34) (35) (36) (37) In equation (32), For nodes n exist t Moment-time velocity coupling node accessibility; Represents a node o and d The number of accessible road segments between them; Represents a node o and d Between, and through nodes n The number of accessible road sections; This is a function for calculating node degree; Represents a node m and n Whether the connection is connected is indicated by a value of 1 if connected and 0 otherwise. This is the normalization function; Represents a node n existt the speed at time t.

[0064] Equation (33) represents the node n In t the speed coupling node importance at time t. .

[0065] In Equation (34), represents the node n In t the local diffusion congestion index at time t. is the node n the number of reachable nodes at diffusion step k . represents the total number of nodes in the road network.

[0066] In Equation (35), represents the speed coupling connectivity integrity of the road network at time t. t is the number of nodes in the largest connected subgraph; represents the average speed of all nodes in the road network at time t. t

[0067] In Equation (36), is the speed fluctuation index of the road network at time t; t represents the set of speeds of all nodes in the road network at time t; and t represent the standard deviation and mean of the set of speeds, respectively.

[0068] In Equation (37), represents the structure-function resilience index of the road network at time t; t is the weight coefficient, which is obtained by the entropy method, and . Specifically, the indices are first standardized, then their entropy values are calculated (the smaller the entropy value, the more significant the difference in index data and the more important the information carried), and finally the weights are allocated according to the proportion of information importance.

[0069] In Equations (32)~(37), , , , , , and reflect the spatial structure characteristics of the road network, while and and their derivatives correspond to the traffic speed characteristics. The combined effects of the two types of characteristics ensure that the structure-function resilience index can comprehensively depict the spatial topological properties and functional operating characteristics of the road network.​​​​​

[0070] Step 2.2: Hysteresis resilience factor; During the perturbation, existing methods usually only focus on the performance-time variation process of the road network, ignoring the hysteresis effect, such as (a) in Figure 4 To depict the non-ideal fluctuation of road network performance with perturbation intensity during the perturbation phase, this embodiment constructs a road network hysteresis performance variation graph, as shown in (b) in Figure 4 , and further establishes a hysteresis resilience factor.

[0071] Specifically, first, the hysteresis time point of road network performance is defined, denoted as: ,(38) In formula (38), is the hysteresis time point, indicating the inflection point at which the road network performance changes from decline to recovery; represents taking the lower bound value of the set; is the time search window step; is the performance value of the road network at t’ time; represents the minimum value of the road network performance; is the tolerance threshold.

[0072] Secondly, the original road network performance variation curve is divided into resistance-recovery processes: ,(39) In formula (39), and are the performance curves in the resistance and recovery stages, respectively; and represent the time point at which the road network starts to be disturbed and the time point at which the road network recovers to the steady state, respectively.

[0073] In addition, since the resistance and recovery times are usually asymmetric, an interpolation method is used to realize curve continuity and time dimension alignment, denoted as: , ,(40) wherein, is a cubic spline interpolation function; and are the resistance and recovery stage performance curves after cubic spline interpolation processing. The road network hysteresis performance variation graph constructed based on the above formula is shown in (b) in Figure 4 .

[0074] Then, the hysteresis factor is established according to the road network hysteresis performance variation graph, as follows: ,(41) ,(42) (43) In formula (41), represents the hysteresis index of the road network, which is measured from the area of the hysteresis loop.

[0075] In formula (42), is the ideal hysteresis effect of the road network; and are maximum and minimum value calculation functions, respectively.

[0076] In formula (43), is the hysteresis factor, which comprehensively considers the balance decay and hysteresis recovery evolution mechanism of the road network under the influence of disturbance intensity.

[0077] Finally, the hysteresis factor is fused with the structure-function resilience index to generate the hysteresis resilience index : (44) Step 3: Establishing a spatiotemporal fine-grained resilience evaluation algorithm; Therefore, the design process of the spatiotemporal fine-grained resilience evaluation algorithm is as follows: first, according to the target domain city traffic speed label value, the target domain city traffic speed prediction value and the disturbance event, the hysteresis resilience index under non-disturbance and the hysteresis resilience index under disturbance are calculated, respectively.

[0078] Secondly, based on and , the performance retention rate of the road network in the disturbance stage is quantified from a macroscopic perspective, which is represented as: (45) Formula (45) calculates the performance retention rate of the road network from the beginning to the end of the disturbance.

[0079] Then, the average resistance index and the maximum recovery rate are calculated, respectively, to segmentally quantify the resilience in the resistance and recovery stages from a fine-grained perspective, which is represented as: (46) (47) In formula (46), is the hysteresis point, which is calculated by formula (38); represents the average resistance index of the road network from period to ; and formula (47) measures the maximum recovery rate of the road network from period to ​ .

[0080] Finally, the three-part toughness quantification results are weighted and fused: ,(48) In formula (48), , and are weight coefficients. From the super parameter analysis experiment, , and the optimal values of a, b and c are 0.4, 0.3 and 0.3 respectively. R is the final spatiotemporal fine-grained toughness quantification result.

[0081] Verification example In this embodiment, traffic speed data in California, USA is selected as the source domain data in the pre-training stage. Traffic speed data in San Jose, USA and Beijing, China is selected as the target domain data in the fine-tuning stage. Specifically, the source domain data set includes: PEMS-D4, PEMS-D7 and PEMS-D8. The target domain data set includes: PEMS-Rainy and Beijing.

[0082] To prove the superiority of the spatiotemporal cross-domain traffic prediction model of the present application. The following prediction models are selected for comparison, including: DGCRN, ST-Wave, BigST, FlashST and OpenCity model. Among them, DGCRN is based on dynamic graph convolution recurrent network, combining graph convolution and recurrent unit to model the dynamic spatial dependence and time sequence characteristics of the traffic network. ST-WAVE uses spatiotemporal wavelet transform to extract multi-scale spatiotemporal features, improving the model's ability to express complex traffic patterns. BigST is a distributed spatiotemporal prediction framework designed for large-scale urban road networks, focusing on computational efficiency and multi-region data collaboration. FlashST realizes rapid response and prediction of sudden traffic events through lightweight structure and fast training mechanism. OpenCity is based on open-domain multi-source traffic data, supporting multi-task cross-city traffic state prediction and generalization ability verification.

[0083] The evaluation indexes include mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). MAE reflects the average deviation degree of the predicted value and the true value, and the smaller the value, the smaller the overall deviation; RMSE represents the square root of the average value of the square of the difference between the predicted value and the true value, and is more sensitive to extreme deviation, highlighting the influence of large errors, and the smaller the value, the better the prediction stability; MAPE reflects the relative error in percentage form, and is not affected by the dimension, making it easy to understand the error proportion. The three evaluate the accuracy of the model from different dimensions, and the smaller the value, the better the performance of the model.

[0084] This embodiment conducts a comparative experiment on the PEMS-Rainy dataset, and the results are shown in Table 1.

[0085] Table 1 shows the experimental results compared with the latest prediction model. As can be seen from Table 1, regardless of whether it is a cross-domain scenario in the United States or China, the MAE, RMSE and MAPE in this invention are all superior to similar models.

[0086] To visually verify the effectiveness of the method of this invention in assessing urban road network resilience, heavy rainfall was used as a sudden disturbance event, and the assessment results of this invention were visualized on the PEMS-Rainy dataset, as shown below. Figure 5 As shown, the darker the color, the stronger the toughness. A heavy rainfall event that occurred in the San Jose area on February 14, 2019 was selected. Figure 5 The data showcases fine-grained road network resilience before, during, and after rainfall. From a macro perspective, during periods of sustained heavy rainfall, road network resilience in various regions exhibits a dynamic evolution trend of first decreasing and then recovering. From a local perspective, urban centers demonstrate strong resilience, while coastal areas show relatively weaker resilience.

[0087] Using a snowfall event as a sudden disturbance, the effectiveness of the method described in this invention for assessing the resilience of urban road networks is verified. Specifically, on January 21, 2022, Beijing experienced snowfall accompanied by fog, such as... Figure 6 The results show the fine-grained resilience of Beijing's road network before, during, and after snowfall. The results reveal significant spatial differences and dynamic evolution in Beijing's road network resilience. The core urban area has a high road network density and strong traffic capacity, reflecting a good level of resilience. However, while the resilience of the main roads in the core area decreases during snowfall, it recovers quickly afterward due to high road network redundancy and priority given to snow removal resources. The western suburban areas along mountains and rivers are significantly constrained by topography, resulting in a relatively simple road network structure. Furthermore, some roads are close to mountains and rivers, making them susceptible to snow accumulation and icing, causing their resilience recovery to lag behind that of the urban area. The southern suburbs have a relatively sparse road network, mainly consisting of radial corridors, and their resilience was at a moderate level before snowfall. During snowfall, resilience decreased significantly, and the recovery speed after snowfall also lagged behind that of the core area. Zhongguancun, as a typical high-density work area, had high resilience before snowfall, but the snowfall occurred during rush hour, and the combined disturbance effect of commuter traffic caused a significant decrease in the resilience of local ring roads and connecting roads. After the snowfall, the resilience recovered relatively quickly as traffic demand decreased and road conditions improved, but there were still some points where recovery was delayed, indicating that densely populated office areas are more sensitive to extreme weather disturbances.

[0088] Therefore, the application can accurately depict the space-time resilience changes of the road network during the disturbance, and timely identify the local fragile unit, thereby providing support for the traffic management department to carry out infrastructure construction and external intervention.

[0089] The above description is only a description of the preferred embodiments of the application, and is not any limitation on the scope of the application. Any modification or modification made by any person skilled in the art according to the above disclosed technical content shall be regarded as an equivalent effective embodiment, and belongs to the protection scope of the technical scheme of the application.

[0090] The references are as follows: [1] F. Li, J. Feng, H. Yan, et al, “Dynamic graph convolutional recurrent network for traffic prediction: Benchmark and solution,” ACM Transactions on Knowledge Discovery from Data, 2023, 17(1): 1-21. [2] Y. Fang, Y. Qin, H. Luo, F. Zhao, B. Xu, L. Zeng, and C. Wang, “When spatio-temporal meet wavelets: Disentangled traffic forecasting via efficient spectral graph attention networks,” in 2023 IEEE 39th International Conference on Data Engineering (ICDE). IEEE, 2023, pp. 517-529. [3] J. Han, W. Zhang, H. Liu, et al. “Bigst: Linear complexity spatio-temporal graph neural network for traffic forecasting on large-scale road networks,” in Proceedings of the VLDB Endowment, 2024, 17(5): 1081-1090. [4] Z. Li, L. Xia, Y. Xu, and C. Huang, “Flashst: A simple and universal prompt-tuning framework for traffic prediction,” in International Conference on Machine Learning (ICML), ser. ICML’24. JMLR.org, 2024. [5] Z. Li, L. Xia, L. Shi, Y. Xu, D. Yin, and C. Huang, “Opencity: Open spatio-temporal foundation models for traffic prediction,” 2024.

Claims

1. A cross-domain-oriented urban traffic road network resilience pre-judgment evaluation method, characterized in that, The method comprises the following steps: Step 1: constructing a spatio-temporal cross-domain traffic prediction model based on source domain and target domain city traffic data to generate traffic prediction and traffic graph structure under a cross-city scenario; The construction of the spatio-temporal cross-domain traffic prediction model comprises a pre-training stage and a fine-tuning stage: the model is first pre-trained on source domain city traffic data to learn general spatio-temporal features, and then fine-tuned on target domain data; the two stages share a common-irregular time sequence module, a lightweight space memory module and a decoder module; additionally, the fine-tuning stage introduces a fine spatio-temporal distiller module; The working process of the spatio-temporal cross-domain traffic prediction model comprises the following steps: First, the historical speed data of the source domain city is selected for pre-training, and the common-irregular time sequence module, the lightweight space memory module and the decoder are used to capture the general pre-training spatio-temporal features of the source domain city; Then, the historical speed data of the target domain city is selected for fine-tuning, and the common-irregular time sequence module, the lightweight space memory module and the decoder are used to capture the fine-tuning spatio-temporal features of the target domain city, and the fine spatio-temporal distiller is used to realize the effective migration of the source domain features to the target domain; Finally, the speed prediction and the corresponding traffic graph structure of the target domain city under the disturbance scenario are generated; Step 2: designing a hysteresis resilience index according to the prediction result and the traffic graph structure; The hysteresis resilience index comprises a structure-function resilience and a hysteresis resilience factor; The establishment process comprises the following steps: first, the structure-function resilience index is designed by simultaneously considering the spatial structure of the road network and the traffic speed characteristics; second, the hysteresis resilience factor is designed by analyzing the hysteresis characteristics of the road network in the disturbance stage; and finally, the hysteresis resilience factor is integrated into the structure-function resilience to construct the hysteresis resilience index; Step 3: establishing a spatio-temporal fine-grained resilience quantification algorithm based on the hysteresis resilience index to output the resilience result; First, according to the target domain city traffic speed label value, the target domain city traffic speed prediction value and the disturbance event, respectively calculate the hysteresis resilience index under non-disturbance And the hysteresis resilience index under disturbance ; Secondly, based on and , the performance retention rate of the road network in the disturbance stage is quantified from a macroscopic perspective ; Then, the average resistance index and the maximum recovery rate are calculated, respectively. Finally, the three parts of the resilience quantification result are weighted and fused to obtain: , wherein, , and is a weight coefficient, R is the final spatiotemporal fine-grained resilience quantification result.

2. The method of claim 1, wherein, The common-irregular time sequence module comprises time scale decomposition, patch processing, global time attention convolution and gated time convolution; specifically as follows: Step 1.2.1: Time scale decomposition Spatiotemporal traffic data is transformed into frequency domain signal by fast Fourier transform X , and then the significant frequency components with the largest K amplitudes are selected, and inverse fast Fourier transform is used to restore them into low frequency sequence and high frequency sequence, which is represented as ,(2) ,(3) ,(4) ,(5) ,(6) In formula (2), spatiotemporal traffic data representing source domain or target domain; is a fast Fourier transform function; is a frequency domain signal; in formula (3), is an amplitude function; represents the selected amplitude before K significant frequency domain components; in formula (4), represents a frequency domain subset composed of the former K significant frequency domain components; in formula (5), is an inverse fast Fourier transform function; represents a low frequency sequence, reflecting long-time characteristics of the time sequence, and is used to characterize normal traffic state; and is a high frequency sequence, focusing on short-term irregular fluctuations of the time sequence, and is used to depict abnormal traffic state; Step 1.2.2: Patch processing For low-frequency sequences and high frequency sequences Patch processing is performed on continuous sequences. Divided into coarse-grained time blocks, for or , is represented as; ,(7) In formula (7), represents the sequence after patching; is the window length of the patch; represents the number of patch blocks; Using linear layers Mapping the Patch sequence to the dimensions required by the model is represented as: ,(8) Further, to preserve the periodic and trend context information of the spatio-temporal data, an embedding term is introduced for each Patch block which represents the temporal position information of the Patch. The time information covers the periodic characteristics of the patch corresponding data in the time sequence dimension, including daily and weekly cycles; the position information refers to the sequential relationship of the patch in the overall time sequence, which is used to reflect the time sequence logic between the patches to provide cross-patch context information; and is expressed as: ,(9) ,(10) Equation (9) represents a broadcast operation; Equation (10) represents feature fusion, Patch sequence representing superimposed position information, including patching low frequency sequence and patching high frequency sequence ; Step 1.2.3: Global time attention convolution A global time attention convolution is designed, and the normal periodicity information of the traffic mode is supplemented explicitly by combining the source domain or target domain normal time mode library, and the process is expressed as: ,(11) Equation (11) represents patching the low frequency sequence with the source domain or target domain normal temporal pattern library performing a Hadamard product ; wherein, is an adaptive matrix; is a patching low frequency sequence; patched low frequency sequence after explicit adaptive update ,(12) In formula (12), are trainable parameters; Query, Key and Value matrices respectively for learning the feature mapping of queries, keys and values across time blocks; ,(13) In formula (13), is an activation function; is a transpose of the Key matrix; is a feature dimension; denotes an attention matrix; ,(14) Equation (14) represents the global-time attention convolution operation. ; Indicates via Captured source or target domain normal time dependency ; Step 1.2.4: Gated time convolution The gated time convolution combines the abnormal fluctuation information of the traffic mode by combining the source domain or target domain abnormal time mode library, and the process is expressed as: ,(15) Equation (15) represents patching the high frequency sequence with the source domain or target domain abnormal time pattern library performing a Hadamard product; for an adaptive matrix, for a patched explicitly adaptive updated patched high frequency sequence; ,(16) In Equation (16), for a dilated convolution operation; denotes a dilation factor; is the size of the convolution kernel; ,(17) In formula (17), is a gated time convolution operation; and is an activation function; denotes a time convolution operation; denotes a time-dependent captured source domain or target domain anomaly Then, the normal and abnormal time dependencies are fused to obtain comprehensive time features: ，(18) wherein, represents feature fusion; represents pre-trained temporal feature or fine-tuned temporal feature .

3. The method of claim 1, wherein, The lightweight space memory module comprises a space memory module and a LoRA enhanced training strategy; specifically as follows: Step 1.3.1: Space memory module Pre-training or fine-tuning temporal features captured by constant-discrete time-series modules The spatial memory module further extracts spatial features between traffic nodes and strengthens the model's memory of spatial relationships by combining source or target domain spatial pattern libraries, as shown below: ,(19) ,(20) ,(21) In formula (19), are trainable parameters; respectively represent Query, Key and Value matrices, which are respectively used for learning feature mapping of queries, keys and values of the spatial pattern; formula (20) represents obtaining an intermediate layer spatial dependency with an attention mechanism ; in formula (21), is an activation function; , , and are trainable parameters; represents spatial dependency captured by the spatial memory module; is a spatial pattern library, . Step 1.3.2: LoRA enhanced training strategy LoRA low-rank decomposition is introduced to reduce trainable parameters and reduce the cost of model transfer learning across cities, and the low-rank decomposition is represented as: ,(22) (23) In formula (22), denotes a low-rank matrix; in formula (23), and is an increment term; updating the original weights based on a low-rank matrix and an increment term is expressed as: (24) Further, the enhanced Query and Value are represented as: ,(25) Finally, the spatial dependence of LoRA enhanced training is changed from equation (20) and equation (21) to equation (26) and equation (27): ,(26) ,(27) In formula (26), is the LoRA optimized intermediate layer spatial dependency; in formula (27), represents the pre-trained spatial features captured via the lightweight spatial memory module or the fine-tuned spatial features .

4. The method of claim 1, wherein, The fine spatio-temporal distiller includes pre-training-fine-tuning spatio-temporal distillation and pre-training spatio-temporal feature refinement. Step 1.4.1: pre-training-fine-tuning spatio-temporal distillation; Distillation using similarity of temporal features: the pre-trained temporal features are evaluated using a KL divergence loss and fine-tuned temporal features The similarity of the pre-trained and fine-tuned temporal features is expressed as: ,(28) In formula (28), facilitate with character alignment; M is the number of samples; Distillation with spatial feature similarity: Evaluate pre-trained spatial features using a contrastive learning loss and fine-tuned spatial features consistency, denoted as: ,(29) In formula (29), is a similarity measure function; denotes an exponential function; can be effectively aligned and ; by minimizing a distillation loss using temporal feature similarity and a distillation loss using spatial feature similarity, to obtain updated distilled temporal features with the distilled spatial features to provide input for subsequent pre-training of spatio-temporal features refinement; Step 1.4.2: pre-training spatio-temporal feature refinement; Refining temporal features captured during pre-training using adaptive filters and spatial features are represented as: ,(30) In formula (30), and respectively represent adaptive temporal and spatial filters, , respectively element-wise multiply with temporal feature , spatial feature , and optimize parameters in pre-training task-oriented.

5. The method of claim 1, wherein, The decoder is composed of a multi-layer perceptron The decoder of the fine-tuning phase is represented as: ,(31) wherein, represents a Concat operation; is a cross-city traffic prediction result output by the spatio-temporal cross-domain traffic prediction model.

6. The method of claim 1, wherein, The traffic graph is defined as: ,(1) wherein represents a set of nodes; is a set of edges connecting nodes; is an adjacency matrix, when a node is connected to a node ; and ; Otherwise, ; based on , the spatio-temporal traffic data is represented as a sequence ; wherein, is the length of the time series; is the traffic feature.

7. The method of claim 1, wherein, Step 2 is specifically: Step 2.1: structure-function resilience index The structure-function resilience index includes speed coupling node accessibility, speed coupling node importance, local diffusion congestion index, speed coupling connectivity integrity and speed fluctuation index, which are defined as: ,(32) ,(33) ,(34) ,(35) ,(36) ,(37) In formula (32), is a node n In t the speed coupling node reachability at time t; is a node o between nodes d ; is a node o between nodes d , and passing through node n ; is a function of the node degree; is a node m between nodes n ; is a normalization function; is a node n speed at time t t ; Equation (33) represents the node n At t Velocity coupling node importance at time ; In formula (34), representing a node n At t the local diffusion congestion index at time t; is a node n the number of reachable nodes at diffusion step k t; representing the total number of nodes of the road network; In formula (35), represents the velocity coupling connectivity integrity of the road network at t time; is the number of nodes of the largest connected subgraph; represents the average velocity of all nodes of the road network at t time; In formula (36), is the speed fluctuation index of the road network at time t ; is the speed fluctuation index of the road network at time t ; is the speed fluctuation index of the road network at time ; In formula (37), The structure-function resilience index of the road network at the time t represents the structure-function resilience index of the road network at the time t. t The structure-function resilience index of the road network at the time t represents the structure-function resilience index of the road network at the time t. is a weight coefficient, which is obtained by an entropy value method, and is a weight coefficient, which is obtained by an entropy value method, and In formulas (32) to (37), , , , , , and reflect the spatial structure characteristics of the road network, while and and their derivatives correspond to the traffic speed characteristics; Step 2.2: hysteresis resilience factor The hysteresis time point of road network performance is defined as: ,(38) In formula (38), is the inflection point of the road network performance, indicating the turning point from decline to recovery of the road network performance; represents taking the lower bound value of the set; is the time search window step; is the performance value of the road network at the time point; t’ is the performance value of the road network at the time point; represents the minimum value of the road network performance; is the tolerance threshold; Secondly, the original road network performance change curve is divided into resistance-recovery process: ,(39) In formula (39), and are the resistance and recovery phase performance curves, respectively; and denote the time point at which the road network begins to be disturbed and the time point at which the road network recovers to the steady state, respectively. The curve is continuous and the time dimension is aligned using the interpolation method, which is represented as: , ,(40) In formula (40), is a cubic spline interpolation function; and is a cubic spline interpolation processed resistance and recovery phase performance curve; Then, the hysteresis factor is established according to the road network hysteresis performance change graph: ,(41) ,(42) ,(43) In formula (41), denotes the hysteresis index of the road network, which is measured from the area of the hysteresis loop; in formula (42), is the ideal hysteresis effect of the road network; and are maximum and minimum value calculation functions, respectively; in formula (43), is the hysteresis factor, which comprehensively considers the balance decay and hysteresis recovery evolution mechanisms of the road network under the influence of disturbance intensity; Eventually, the hysteresis factor is merged with the structure-function resilience indicator to generate a hysteresis resilience indicator : ,(44)。 8. The method of claim 1, wherein, In step 3, The performance retention rate in the disturbance stage is represented as: ,(45) Equation (45) computes the performance retention of the road network from the beginning to the end of the perturbation ; The resistance and recovery stage resilience is represented as: ,(46) ,(47) In formula (46), is the hysteresis point, denotes the average resistance index of the road network in the time period to ; formula (47) measures the maximum recovery rate of the road network in the time period to .

Images