Road network function state cross-regional prediction method and system under extreme rainstorm

Abstract

The invention discloses a road network function state cross-regional prediction method and system under extreme rainstorm, and the method comprises the steps: generating a'rainstorm-traffic 'physical simulation dynamic data set based on an extreme rainstorm scene of a preset region; obtaining and preprocessing static characteristics and road network data of a preset area to generate a static data set; constructing a G-R2R model, and training the G-R2R model through a'rainstorm-traffic 'physical simulation dynamic data set and a static data set; and inputting the rainstorm time sequence, the static characteristics and the road network data of the target area into the trained G-R2R model, and outputting a road network function state prediction sequence of the whole rainstorm process by adopting an iterative prediction method. Under the condition that the target area lacks disaster response observation data, space-time prediction output of the road network function state can be realized without retraining the model for the target area, and direct popularization and application of the model in different cities or areas are supported.

Description

Technical Field

[0001] This invention relates to the field of transportation network function assessment and disaster risk prediction technology, specifically to a cross-regional prediction method and system for the functional status of a road network under extreme rainstorms. Background Technology

[0002] Extreme rainstorm disasters pose a severe challenge to traditional disaster prevention and mitigation systems due to their wide impact, destructive power, and significantly non-stationary evolution. Urban flooding induced by rainstorms leads to reduced road capacity, resulting in a chain of risks such as evacuation delays, hindered rescue efforts, difficulties in resource allocation, and delayed restoration of critical infrastructure. Therefore, when constructing standardized resilience management systems at the provincial and municipal levels, there is an urgent need for a key technology that can rapidly transform meteorological and rainfall information into spatiotemporal predictions of road network functional status to support pre-disaster early warning and rolling assessments during disasters.

[0003] For ease of description, the "Rainfall Time Series to Road Network Functionality Prediction Task" is defined as a Rainfall to Road Network Functionality (R2R prediction task) mapping prediction task. Here, "Road Network Functionality" refers to the accessibility or functional performance of the road network at a given time, which can be characterized by road segment speed, congestion level, accessibility, or functional index (Road Network Functionality, hereinafter referred to as "RNF"). The core objective of the R2R prediction task is to learn the mapping relationship from the rainfall time series to the future RNF spatiotemporal series under given road network topology and historical and forecast rainfall sequences, thereby achieving multi-step spatiotemporal prediction of the RNF. Meanwhile, the low-data environment mentioned in this application refers to the lack of sufficient data in the target area to support model construction, calibration, or validation, including at least one or more of the following situations:

[0004] (1) High-quality traffic operation response data is missing or sparsely sampled during disasters (such as road speed, traffic capacity, road closure records, etc.).

[0005] (2) Information on critical infrastructure is missing or incomplete (e.g., key parameters of drainage systems are difficult to obtain).

[0006] (3) The number of extreme rainstorm events is insufficient, making it difficult to form a representative sample set that can be used for training and validation.

[0007] (4) The data during the disaster were severely lacking, noisy, lacked spatial accuracy, or had difficulty in time alignment, making it difficult to use them directly for modeling.

[0008] Currently, the main methods for predicting the road network functional state (RNF) under extreme rainstorm conditions are as follows:

[0009] (1) Simulation prediction methods based on physical mechanisms: These methods typically simulate the evolution of urban flooding scenarios using hydrological and hydrodynamic models, and then combine this with traffic network analysis to obtain the spatiotemporal evolution results of changes in road capacity and the Recurrent Flooding (RNF). These methods have strong physical interpretability, but their computational complexity often increases rapidly with the expansion of the study area scale. Limited by the bottlenecks in model integration complexity and computational efficiency, they are difficult to meet the real-time response requirements for rapid RNF prediction at the provincial and municipal scales, and their application is particularly limited in scenarios with rapid disaster evolution and rolling update decision-making.

[0010] (2) Single-region R2R data-driven prediction method: To reduce the cost of physical simulation, existing studies have used historical observation data or physical simulation data to train surrogate models and establish a mapping relationship between rainfall and RNF within a single region, thereby achieving rapid prediction of RNF. Chinese patent CN117875190A discloses a "Spatiotemporal prediction method for the functional state of transportation networks in extreme rainstorm events", which belongs to this type of single-region prediction technology. However, this type of method has the following shortcomings: 1) The model depends on region-specific training data and region-specific parameter features, and the prediction performance often drops significantly when migrated to a new region; 2) When applied in a new region, it is usually necessary to repeatedly build physical simulation models to generate training data and retrain surrogate models, thereby weakening the efficiency advantage of data-driven methods over physical models; 3) In low-data environments, due to the lack of disaster response observation data and key basic information support, single-region data-driven models are difficult to obtain sufficient training or calibration conditions, making it difficult to quickly implement and apply them in new regions.

[0011] (3) Cross-regional modeling methods for routine traffic forecasting: In recent years, cross-regional spatiotemporal forecasting research has made progress, but most of the relevant models are based on routine traffic operation data, which makes it difficult to fully depict the strong nonlinear coupling mechanism of "rainfall-waterlogging evolution-road network function degradation and recovery" during rainstorm disasters. Therefore, it is difficult to directly meet the emergency response planning needs under rainstorm disaster scenarios, and it is also difficult to support the standardized deployment of a unified resilience management system at the provincial and municipal scale.

[0012] In summary, the existing technologies still have the following key problems: (1) The physical mechanism simulation model has high computational cost when the regional scale is expanded, making it difficult to meet the real-time requirements. (2) The single-region R2R data-driven prediction model has strong regional dependence and lacks the ability to be directly deployed to new regions, often requiring repeated modeling and retraining. (3) In low-data environments, there is a lack of training, calibration and validation data support, making it difficult for both physical models and data-driven models to effectively establish reliable prediction capabilities.

[0013] Therefore, there is an urgent need for an RNF spatiotemporal prediction method that can be deployed across regions under extreme rainstorm disaster scenarios, output standardized prediction results without repeated modeling in unobserved areas, and adapt to low-data environments, so as to reduce the implementation cost of a unified resilience management system at the provincial and municipal scale and improve emergency response efficiency. Summary of the Invention

[0014] To overcome the shortcomings of the above-mentioned technologies, this invention provides a cross-regional prediction method and system for road network functional status under extreme rainstorms. Even when disaster response observation data is lacking in the target area, the learned rainstorm-road network functional status mapping relationship can be transferred to unobserved areas without retraining the model for the target area. This enables spatiotemporal prediction output of road network functional status, supporting the direct application of the model across different cities or regions, reducing cross-regional deployment costs, and improving overall application efficiency.

[0015] Terminology Explanation:

[0016] 1) G-R2R: Generalized Rainfall-to-Road-Network Functionality, which considers cross-regional generalization of the task of mapping and predicting the functional state of road networks from rainfall time series.

[0017] 2) LSTM: Long Short-Term Memory. A type of recurrent neural network model.

[0018] 3) LISFLOOD-FP: An open-source two-dimensional hydrodynamic model for simulating flood inundation in complex terrain.

[0019] 4) InfoWorks ICM: InfoWorks Integrated Catchment Modeling, a commercial software for integrated watershed modeling, used for integrated simulation of urban drainage, river channels and hydrological and hydraulic processes.

[0020] 5) MIKE flood: A flood simulation tool, including a complete one-dimensional and two-dimensional flood simulation engine.

[0021] 6) ArcGIS: A Geographic Information System (GIS) platform software used for the creation, editing, management, analysis, mapping and sharing of geographic data to support various geospatial analysis and decision-making tasks, enabling people around the world to apply geographic knowledge to government, business, science and technology, education and media.

[0022] 7) IPW: Independent Pathway, refers to reachable paths between node pairs that do not share links (roads) with each other. It can be used to characterize the redundancy of a transportation network and is the most important topological feature of a transportation network.

[0023] 8) DEM: Digital Elevation Model, a digital model used to represent the spatial distribution of surface elevation.

[0024] The technical solution adopted by this invention to overcome its technical problems is:

[0025] A method for cross-regional prediction of road network functional status under extreme rainstorms includes the following steps:

[0026] Based on extreme rainstorm scenarios in a preset area, a dynamic dataset of "rainstorm-traffic" physical simulation is generated;

[0027] Acquire static features and road network data of a preset area and preprocess them to generate a static dataset;

[0028] A cross-regional generalized prediction model for rainstorm-road network functional state, referred to as the G-R2R model, is constructed. The G-R2R model is trained using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset.

[0029] The rainstorm time series, static features, and road network data of the target area are input into the trained G-R2R model, and the road network functional state prediction sequence of the entire rainstorm process is output using an iterative prediction method.

[0030] Furthermore, the generation of a dynamic dataset for "rainstorm-traffic" physical simulation based on extreme rainstorm scenarios in a preset area specifically includes:

[0031] Based on the design rainfall patterns and historical records of extreme rainstorm events in the preset area, several rainstorm time series are generated through sampling methods.

[0032] Based on the geographical and hydrological conditions of the preset area and the time series of the rainstorm, the urban flooding scenario is simulated, and the flooding scenario is mapped onto the road network to obtain the service level of all road sections under the extreme rainstorm scenario.

[0033] Suppose that the road network includes a node set and an edge set, where the node set represents traffic nodes and the edge set represents road segments. Based on the service level of the road segments, a traffic accessibility index based on independent paths between road nodes is constructed, and the traffic function index of each road node is calculated.

[0034] Furthermore, the step of acquiring static features and road network data of a preset area and preprocessing them to generate a static dataset specifically includes:

[0035] Obtain the static features of the preset area, including at least terrain static features, land use type features, and drainage outlet replacement features;

[0036] Align the road nodes with the rasterized static features, take each road node in the preset area as the target node, and extract the square raster window of the preset range around the target node as the node static feature slice.

[0037] For all road nodes within a preset area, generate a sampled sub-map for each road node;

[0038] The road is divided into grid cells, and the terrain feature index of each grid cell is calculated to obtain the terrain perception weight of each road segment, thereby constructing a terrain perception adjacency matrix.

[0039] Furthermore, the step of generating a sampling sub-map for each road node within the preset area specifically includes:

[0040] Each road node within the preset area is sequentially used as a prediction node. A breadth-first search algorithm is used to extract a preset number of neighboring nodes around the current prediction node from the original road network of the preset area as a neighboring node set. The current prediction node and its neighboring node set are combined to form a sampling subgraph of the road node. Then, all road nodes within the preset area are traversed until a corresponding sampling subgraph is generated for each road node.

[0041] Furthermore, the G-R2R model includes a static feature encoding module, a spatiotemporal feature encoding module, and a spatiotemporal feature decoding module; the static feature encoding module is used to encode static features to obtain static features; the spatiotemporal feature encoding module is used to encode the rainstorm time series and the historical road network functional state series of the subgraph to obtain spatiotemporal dynamic features; the spatiotemporal feature decoding module is used to fuse the static features obtained by the static feature encoding module and the spatiotemporal dynamic features obtained by the spatiotemporal feature encoding module to output a road network functional state prediction sequence for multiple future time steps.

[0042] Furthermore, the static feature encoding module includes a static feature encoder for encoding static feature slices of predicted nodes and a set of static feature encoders for encoding static feature slices of sampled subgraph nodes. Each static feature encoder includes multiple convolutional layers and pooling layers connected to each convolutional layer. The input of each static feature encoder is a multidimensional variable composed of grid static features corresponding to a road node. After multiple convolution and pooling operations, it becomes a one-dimensional vector. The one-dimensional vector is concatenated with the traffic function index of the road node under normal conditions and then subjected to one-dimensional convolution to output the static feature information of the road node.

[0043] Furthermore, the static feature variables obtained by encoding the static feature slices of the prediction nodes are broadcast and expanded along the height and width dimensions, while the static feature variables obtained by encoding the static feature slices of the sampled subgraph nodes are broadcast and expanded along the time dimension. After the broadcasting operation, two three-dimensional static feature variables are formed.

[0044] Furthermore, the spatiotemporal feature encoding module includes an LSTM module and a spatiotemporal encoder module. The LSTM module is used to process rainstorm time series data, capture the time dependency between the peak intensity of rainstorms and changes in road network functional status. The rainstorm time series processed by the LSTM module is concatenated with the historical road network functional status of the subgraph and then input into the spatiotemporal encoder module. The spatiotemporal encoder module is used to learn the spatiotemporal dynamic influence of rainfall and static features on changes in road network functional status, thereby outputting spatiotemporal feature variables. The spatiotemporal encoder module adopts a double-layer stacked spatiotemporal convolutional block structure. Each spatiotemporal convolutional block includes two time gating units and a spatial graph convolutional unit set between the two time gating units.

[0045] Furthermore, the two three-dimensional static feature variables output by the static feature encoding module are concatenated with the spatiotemporal feature variables output by the spatiotemporal feature encoding module to form a multi-dimensional feature variable, which is then input into the spatiotemporal feature decoding module. The spatiotemporal feature decoding module includes a spatiotemporal convolutional block, a temporal control unit, and a fully connected neural network. The spatiotemporal convolutional block is used to learn the joint spatiotemporal correlation between dynamic and static feature information. The temporal control unit is used to adjust the output time step. The fully connected neural network is used to perform feature decoding to complete the prediction output of the future road network functional state.

[0046] This invention also discloses a cross-regional prediction system for road network functional status under extreme rainstorms, comprising:

[0047] The dynamic dataset generation module is configured to generate a dynamic dataset of physical simulation of "rainstorm-traffic" based on extreme rainstorm scenarios in a preset area.

[0048] The static dataset generation module is configured to acquire and preprocess static features and road network data of a preset area to generate a static dataset.

[0049] The model building and training module is configured to build a cross-regional generalized prediction model for rainstorm-road network functional state, referred to as the G-R2R model. The G-R2R model is trained using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset.

[0050] The prediction application module is configured to input the rainstorm time series, static features and road network data of the target area into the trained G-R2R model, and output the road network functional state prediction sequence of the entire rainstorm process using an iterative prediction method.

[0051] The beneficial effects of this invention are:

[0052] 1) It possesses cross-regional generalization capabilities, solving region-specific modeling challenges. Since most existing technologies rely on data from specific regions for model training, retraining or data generation is required when applying to new regions, limiting the cross-regional application of the model. This invention constructs a unified cross-regional generalized prediction model for rainstorm-road network functional state, enabling direct transfer between different regions without retraining for each new region. This significantly improves the model's cross-regional adaptability and makes large-scale deployment of disaster management systems possible.

[0053] 2) Strong adaptability to low-data environments. Traditional methods typically rely on high-quality real-time disaster response data, which is difficult to obtain in many regions. In contrast, this invention effectively reduces reliance on real-time disaster data by utilizing readily available data sources such as rainfall forecasts, static features, and road network data. This allows the model to maintain high prediction accuracy in data-scarce environments, ensuring its stability and practicality under different data conditions.

[0054] 3) Reduce computational costs and improve decision-making efficiency. Traditional physical simulation methods typically require complex computational processes, and the computational complexity increases dramatically as the study area expands. In contrast, this invention employs a data-driven neural network model, which not only avoids the cumbersome hydrodynamic simulation process but also significantly improves prediction efficiency, providing real-time support for rapid post-disaster decision-making and emergency response.

[0055] 4) Improve the model's generalization, interpretability, and accuracy. This invention enhances the model's adaptability to regional differences and changes in the geographical environment through innovative designs such as terrain-aware adjacency matrix and static feature dual-path encoding. It can more accurately capture the impact of terrain changes on urban flooding diffusion, while also improving the model's interpretability. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the static feature extraction of road network prediction nodes and neighboring nodes according to an embodiment of the present invention.

[0057] Figure 2 This is a schematic diagram illustrating the calculation of road segment terrain perception weights according to an embodiment of the present invention.

[0058] Figure 3 This is a schematic diagram of the architecture of the G-R2R model described in an embodiment of the present invention.

[0059] Figure 4 This is a schematic diagram of the structure of the static feature encoder described in an embodiment of the present invention.

[0060] Figure 5 This is a schematic diagram of the spatiotemporal encoder module according to an embodiment of the present invention.

[0061] Figure 6 This is a schematic diagram of the iterative prediction method described in an embodiment of the present invention.

[0062] Figure 7 This is an error histogram of the ablation experiment results described in this embodiment of the invention.

[0063] Figure 8 This is an analysis chart of the mean absolute percentage error and coefficient of determination for the generalization test described in the embodiments of the present invention.

[0064] Figure 9 This is a diagram illustrating the short-time single-prediction time error analysis of the G-R2R model described in an embodiment of the present invention.

[0065] Figure 10 This is a graph showing the error analysis of the 12-hour iterative prediction time of the G-R2R model described in this embodiment of the invention.

[0066] Figure 11 This is the road network function prediction curve of the G-R2R model described in this embodiment of the invention in an extreme rainstorm scenario.

[0067] Figure 12 This is a graph showing the relative error analysis results of the G9 region as described in an embodiment of the present invention.

[0068] Figure 13 This is a spatial comparison diagram of the prediction results of the physical simulation model and the G-R2R model described in the embodiments of the present invention. Detailed Implementation

[0069] To facilitate a better understanding of the present invention by those skilled in the art, exemplary embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These are merely exemplary embodiments of the present invention; however, it should be understood that the present invention can be implemented in various forms and is not limited to the embodiments described herein. These embodiments are provided to enable those skilled in the art to gain a clearer and more thorough understanding of the present invention.

[0070] Example 1:

[0071] The method for cross-regional prediction of road network functional status under extreme rainstorms described in this embodiment includes the following steps:

[0072] S1. Generate a dynamic dataset of physical simulation of "rainstorm-traffic" based on extreme rainstorm scenarios in a preset area.

[0073] This embodiment uses a physical simulation link of "extreme rainstorm event - urban flooding evolution simulation - road network functional status assessment" to construct supervised learning samples for training the G-R2R model. Specifically, step S1 includes the following steps:

[0074] S11. Constructing extreme rainstorm events: Based on the design rainfall pattern and historical records of extreme rainstorm events in the preset area, several rainstorm time series are generated through sampling methods.

[0075] Specifically, based on the design rainfall patterns and historical records of extreme rainstorm events in the preset area, the Monte Carlo sampling method is used to generate several rainstorm time series with a time resolution of 1 hour and a duration of 24 hours.

[0076] S12. Urban Flood Simulation: Based on the geographical and hydrological conditions of the preset area and the time series of the rainstorm, simulate the urban flooding scenario and map the flooding scenario onto the road network to obtain the service level of all road sections under the extreme rainstorm scenario.

[0077] Specifically, based on the geographical and hydrological conditions of the preset area and the aforementioned rainfall time series, an urban flooding scenario is simulated. Flood simulation models can be selected, including classic 2D flood models (such as LISFLOOD-FP or cellular automata-based flood simulation) or mature commercial models (such as InfoWorks ICM or MIKE flood). After acquiring the flooding scenario, ArcGIS software is used to map the flooding scenario onto the road network to extract the road flooding status and assess the impact of urban flooding on the road network capacity, including throughput and maximum safe driving speed, thereby obtaining the service level of all road segments in the road network under extreme rainfall scenarios.

[0078] S13. Traffic Network Analysis: Suppose that the road network includes a node set and an edge set, where the node set represents traffic nodes and the edge set represents road segments. Based on the service level of the road segments, construct a traffic accessibility index based on independent paths between road nodes, and calculate the traffic function index of each road node.

[0079] Specifically, this embodiment proposes a traffic accessibility index based on Independent Paths (IPWs) between road nodes. Assuming the road network comprises a set of nodes and a set of edges, the road network can be described as a directed graph. ,in, Represents a set of nodes (representing traffic nodes, such as residential areas, economic hubs, and major road intersections). Let edge set (representing road segments) be defined. This represents the total number of IPWs, where each IPW is a concatenated, ordered set of edges. Therefore, road nodes... At any moment The traffic function index NIPW (Node Independent Pathway) is defined as the average reliability of all its IPWs, i.e.:

[0080] (1)

[0081] In formula (1), Road nodes With road nodes Between IPW at time Weighting factors; Road nodes With road nodes Between IPW at time The reliability (assuming all edges in the road network are independent). They are expressed as follows:

[0082] (2)

[0083] (3)

[0084] In formulas (2) and (3), Represents road nodes and Between the first One IPW; Indicates the road segment number; For road section Length (unit: kilometers); Indicates road segment At any moment The level of service.

[0085] S2. Obtain static features and road network data of the preset area and preprocess them to generate a static dataset.

[0086] The data generated by the physical simulation contains both structured raster information and graph network information. In order to meet the unified input requirements of cross-regional and cross-scale road network prediction models (i.e., G-R2R models), this embodiment proposes a data preprocessing mechanism for G-R2R models.

[0087] Step S2 in this embodiment specifically includes the following steps:

[0088] S21. Obtain the static features of the preset area, including at least the static features of terrain, land use type, and drainage outlet replacement features.

[0089] Specifically, the selection of static features in this embodiment follows the following principles: (a) closely related to the water accumulation mechanism; (b) easily obtainable across regions; (c) able to unify spatial scale; and (d) able to improve the generalization ability of unknown regions.

[0090] Accordingly, this embodiment selects at least the following static feature categories as inputs to the G-R2R model: (1) Topographic static features. These include digital elevation models (DEMs) and their derived topographic features, such as slope, aspect, curvature, and roughness, to characterize surface runoff paths, low-lying and flood-prone features, and topographic control. (2) Land use type features. Land use type indicators that reflect the degree of impermeability of the ground are selected to characterize differences in surface runoff generation and runoff and water accumulation sensitivity. (3) Drainage outlet substitution features. When the actual location of drainage outlets is difficult to obtain, since rainwater drainage pipes are usually laid along roads and eventually flow into natural water bodies, this embodiment uses the point set formed by the spatial intersection of the road network and natural water bodies (rivers, lakes, etc.) as the substitution feature of the spatial distribution of drainage outlets (hereinafter referred to as the rw point set), and calculates accordingly. (For any grid cell, take its Euclidean distance to the nearest rw point as...) The value is one of the static features input to the G-R2R model.

[0091] S22. Align the road nodes with the rasterized static features, take each road node in the preset area as the target node, and extract the square raster window of the preset range around the target node as the node static feature slice.

[0092] Specifically, such as Figure 1 As shown, road nodes are aligned with rasterized static features. Each road node in a preset area is taken as a target node. Information within a square raster window centered on the target node is extracted and used as a static feature slice of that node, i.e., the prior input of the node's static features. This slice is used to characterize the impact of static features such as topography and hydrology on the node's waterlogging sensitivity. Generally, the side length of the square raster window can be selected between 100 meters and 1000 meters, with the value that minimizes the training error of the G-R2R model being the optimal value.

[0093] S23. For all road nodes within the preset area, generate a sampled sub-map for each road node.

[0094] Because the size and topological complexity of road network nodes vary significantly across different regions, directly modeling the entire graph would result in inconsistent model input sizes and make cross-regional deployment difficult. Therefore, this embodiment constructs a fixed-size sampling subgraph for each road node. Specifically, each road node within a preset area is sequentially used as a prediction node. A breadth-first search algorithm is employed to extract a preset number of neighboring nodes from the original road network of the current prediction node, forming a set of neighboring nodes. The current prediction node and its set of neighboring nodes together form the sampling subgraph for that road node. Then, all road nodes within the preset area are traversed until a corresponding sampling subgraph is generated for each road node.

[0095] S24. Divide the road along the line into grid cells, calculate the terrain feature index of each grid cell to obtain the terrain perception weight of each road segment, and then construct the terrain perception adjacency matrix.

[0096] To enable the G-R2R model to better capture the differences in flood risk along roads, this embodiment introduces terrain-aware weights into the subgraph adjacency matrix, such as... Figure 2 As shown, the roadside is divided into grid cells, and the terrain feature index of each grid cell is calculated. The median of the terrain feature indices of all road segment cells is used as the terrain-aware weight for that edge, thus constructing a "terrain-aware adjacency matrix" to replace the traditional binary adjacency matrix. This improves the G-R2R model's ability to express the impact of terrain and hydrodynamics and its cross-regional portability. The formula for calculating the terrain feature index is as follows:

[0097] (4)

[0098] In formula (4), Represents grid cells Elevation value at the location (unit: m); Represents grid cells Elevation value at the location (unit: m); This represents the side length of the square neighborhood window used when calculating the terrain feature index. This side length can be selected between 100 meters and 1000 meters, with the value that minimizes the training error of the G-R2R model being the optimal value. In this embodiment, it is preferred. It is 300 meters; This represents the average elevation value (in meters) of all grid cells within the square neighborhood window.

[0099] Note: The G-R2R model in this embodiment can be trained using only the "rainstorm-traffic" physical simulation dynamic dataset described in step S1 and the static data described in step S2. After training, it can be directly applied to new areas, avoiding the need to repeatedly construct the "rainstorm-traffic" physical simulation dynamic dataset for new areas.

[0100] S3. Construct a cross-regional generalized prediction model for rainstorm-road network functional state, referred to as the G-R2R model, and train the G-R2R model using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset.

[0101] (I) Construction of the G-R2R model

[0102] like Figure 3 As shown, the G-R2R model in this embodiment includes three core modules: a static feature encoding module, a spatiotemporal feature encoding module, and a spatiotemporal feature decoding module. The static feature encoding module encodes static features to obtain static features. The spatiotemporal feature encoding module encodes the rainstorm time series and the historical road network functional state sequence of the subgraph to obtain spatiotemporal dynamic features. The spatiotemporal feature decoding module fuses the static features obtained by the static feature encoding module and the spatiotemporal dynamic features obtained by the spatiotemporal feature encoding module to output a predicted sequence of road network functional states for the entire rainstorm process. These three modules work together, enabling the G-R2R model to simultaneously utilize exogenous rainfall drivers, prior differences in static features, and road network topology propagation, thereby improving its generalization prediction capability in unobserved areas.

[0103] (1) Static feature coding module

[0104] The static feature encoding module includes a static feature encoder for encoding static feature slices of the prediction node and a set of static feature encoders for encoding static feature slices of the sampled subgraph node. The static feature encoder for encoding the static feature slices of the prediction node has the same structure as the static feature encoder for encoding the static feature slices of the sampled subgraph node. The structure of each static feature encoder is as follows: Figure 4 As shown, its core is a spatial feature learner based on a convolutional neural network. Each static feature encoder includes multiple convolutional layers and pooling layers connected to each convolutional layer. The input of each static feature encoder is a multidimensional variable composed of the grid static features corresponding to a road node. ,in, This embodiment uses a number of terrain features that reflect the risk of urban flooding; This represents the size of the static feature slice of the node; spatial features are extracted step by step through multiple convolution operations, and pooling operations are used to reduce the spatial dimension of the feature variables into a one-dimensional vector, that is, when it becomes a single pixel, the hidden vector is obtained. The one-dimensional vector (hidden vector) The feature vector is concatenated with the traffic function index (NIPW) of the road node under normal conditions and then subjected to one-dimensional convolution to further extract features, thereby outputting a one-dimensional feature vector to represent the static feature information of the road node.

[0105] In this embodiment, not only is the aforementioned static encoding performed on the predicted node, but static encoding is also performed on all nodes in its sampled subgraph, i.e., a dual-path static encoding mechanism. Specifically, in addition to using a single static feature encoder to extract the static features of the predicted node, a set of static feature encoders is constructed to capture the static features of the sampled subgraph nodes centered on the predicted node. The number of static feature encoders corresponds to the number of subgraph nodes, and they extract the static features of each subgraph node respectively. The output variables of each static feature encoder are concatenated along the height order (i.e., the road network node dimension, such as...). Figure 3 As shown, this forms a two-dimensional variable representing the static features of the sampled subgraph. This design, which uses multiple static feature encoders to extract the static features of the prediction nodes and the sampled subgraph nodes respectively, significantly enhances the prediction accuracy and generalization performance of the G-R2R model.

[0106] Furthermore, the static feature variables obtained by encoding the static feature slices of the prediction nodes are broadcast and expanded along the height and width dimensions. At the same time, the static feature variables obtained by encoding the static feature slices of the sampled subgraph nodes are broadcast and expanded along the time dimension. After the broadcasting operation, two three-dimensional static feature variables are formed, which are then concatenated with the output features of the spatiotemporal feature encoding module and used together as the input variables of the spatiotemporal feature decoding module.

[0107] (2) Spatiotemporal feature coding module

[0108] like Figure 3 As shown, the spatiotemporal feature encoding module includes an LSTM module and a spatiotemporal encoder module.

[0109] Given that the evolution of urban flooding is primarily driven by rainstorm time series data, with peak rainfall intensity and cumulative rainfall having a particularly significant impact on the Recurrent Functional Array (RNF), and considering the significant non-stationary characteristics of extreme rainstorm events, this embodiment employs an LSTM network with a gating mechanism to learn the influence of peak rainfall and cumulative rainfall on urban flooding evolution. The LSTM module processes the rainstorm time series data, capturing the time dependency between peak rainfall intensity and changes in road network functional status. The rainstorm time series processed by the LSTM module is then concatenated with historical road network functional status data and input into the spatiotemporal encoder module. Specifically, the rainstorm time series includes past... Historical rainfall records for hours and future Hourly rainfall forecast, where H and All are positive integers greater than or equal to 1. After processing by the LSTM module, the output feature variables are... Include One time step; subsequently, feature variables Expand the height dimension to [a higher dimension] via broadcast operation. (i.e., the total number of nodes in the sampled subgraph) to obtain the feature variables. ; characteristic variables With subgraph history RNF sequence By concatenating along the time dimension, feature variables are formed. As input to the spatiotemporal encoder module.

[0110] The spatiotemporal encoder module is used to learn the spatiotemporal dynamic influence of rainfall and static features on changes in the functional state of the road network, thereby outputting spatiotemporal feature variables. The spatiotemporal encoder module employs a double-layer stacked spatiotemporal convolutional block structure, such as... Figure 5 As shown, each spatiotemporal convolutional block includes two temporal gating units and a spatial graph convolutional unit positioned between them. The temporal gating units learn temporal dependencies using temporal gating convolution, which essentially applies one-dimensional (1-D) convolution along the time dimension to avoid the parameter redundancy and time-consuming iterations of RNNs (Recurrent Neural Networks). Subsequently, nonlinear data is captured by gating linear units. Due to the temporal convolution operation, the time step of the input variables is reduced by [amount missing]. ( (This refers to the temporal convolution kernel size). The spatial graph convolution unit captures spatial features through graph convolution operations. This embodiment uses the spectral graph method for learning spatial dimensionality information through convolution, employing a subgraph adjacency matrix. ( Indicates road segment The terrain-aware weights are the terrain-aware adjacency matrix. In this embodiment, the Chebyshev approximation method is used to reduce the computational complexity of graph convolution operations.

[0111] The feature variables The input variable is processed by the spatiotemporal encoder module, and the output is a feature variable. The feature variables The two three-dimensional static feature variables generated by the static feature encoding module are processed according to... Figure 3 The components are concatenated as shown to form multidimensional feature variables, which are then input into the spatiotemporal feature decoding module for subsequent calculations.

[0112] (3) Spatiotemporal feature decoding module

[0113] The spatiotemporal feature decoding module adopts a three-level architecture design, such as... Figure 3As shown, it includes a spatiotemporal convolutional block, a temporal control unit, and a fully connected neural network. The spatiotemporal convolutional block is used to learn the joint spatiotemporal correlation between dynamic and static feature information, particularly the spatiotemporal influence mechanism of rainfall, geographic, hydrological, and traffic features on the evolution of the RNF during a disaster. Its structure is the same as that of the spatiotemporal encoder module (e.g., ...). Figure 5 (As shown). The timing control unit is used to regulate the output time step. Specifically, the timing control unit performs a one-dimensional convolution operation in the time dimension, using a specifically designed convolution kernel size. Adjusting the output time step The fully connected neural network is used to perform feature decoding, thereby predicting the future functional state of the road network.

[0114] (II) Training of the G-R2R model:

[0115] In this embodiment, the G-R2R model is trained under supervised supervision using a dynamic dataset of "rainstorm-traffic" physical simulation. The loss function can be a measure of the error between the predicted RNF and the true RNF (such as mean squared error), and the parameters are updated through gradient descent. After training, the G-R2R model only needs to input the rainstorm time series, static features, and road network data of the new region to directly output the RNF prediction results, without needing to retrain for the new region, thus achieving cross-regional generalization deployment.

[0116] S4. Input the rainstorm time series, static features and road network data of the target area into the trained G-R2R model, and use the iterative prediction method to output the road network functional state prediction sequence of the entire rainstorm process.

[0117] Step S4 involves the application of the G-R2R model. Specifically, this embodiment uses an iterative prediction method to predict the road network functional status throughout the entire rainstorm process. For example... Figure 6 As shown, the specific process of this iterative prediction method is as follows: the entire prediction range is divided into multiple time steps with a duration of... Hourly forecast range ( The G-R2R model outputs the number of prediction time steps. For each prediction interval, the network function state (RNF) of the previous prediction can be used as the input variable for the current prediction. This is part of the process to predict the road network functional status throughout the entire rainstorm process. For the input variables of the first prediction interval... If the rainstorm has not yet started, the road network can be considered to be in good working order.

[0118] The following is a detailed explanation of the cross-regional prediction method for road network functional status under extreme rainstorms described in this embodiment, using a specific case.

[0119] This case study focuses on predicting the functional status of the road network during extreme rainstorm events in a coastal region (City B). Extreme rainstorm events occur frequently in City B, particularly during typhoon season, significantly impacting the transportation system. The specific study area includes 10 flood-prone areas selected from City B: G1, G2, G3, G4, G5, G6, G7, G8, G9, and G10. These areas exhibit significant heterogeneity in geography, socio-economic conditions, and transportation infrastructure, reflecting a geographical gradient from coastal plains to inland mountains, spanning multiple hydrological basins, and are highly representative, making them suitable for the regional generalization research of this invention.

[0120] Specifically, the cross-regional prediction method for road network functional status under extreme rainstorms described in this case includes the following steps:

[0121] S1. Generate a dynamic dataset for the physical simulation of "rainstorm-traffic".

[0122] S11. Constructing extreme rainstorm events: Based on the design rainfall pattern and historical records of extreme rainstorm events in City B, 10,000 rainstorm time series samples are generated using the Monte Carlo sampling method. Each rainstorm time series sample has a time resolution of 1 hour, a duration of 24 hours, a peak precipitation intensity range of 20 mm to 80 mm, and a 24-hour cumulative precipitation range of 45.5 mm to 652.5 mm.

[0123] S12. Urban Flood Simulation: The LISFLOOD-FP flood simulation model is used to simulate the urban flooding process under extreme rainstorms. The flooding scenario is then mapped onto the road network of City B using ArcGIS software to assess the impact of the flood on the traffic capacity of each road segment and obtain the service level of all road segments under extreme rainstorm scenarios.

[0124] S13. Traffic Network Analysis: This case study proposes a traffic accessibility index based on Independent Paths (IPW) between road nodes to quantify the functional degradation of the road network under extreme rainstorm events. The Traffic Function Index (NIPW) of each road node is evaluated by calculating the average reliability of all its independent paths.

[0125] S2. Obtain static features and road network data of City B and preprocess them to generate a static dataset.

[0126] S21. Obtaining Static Features: This case study selects static features closely related to the formation mechanism of urban flooding, including: topographic static features (including DEM, slope, aspect, curvature, and roughness), land use type features, and drainage outlet replacement features (i.e., the point set (rw point set) formed by the intersection of the road network and the natural water body space, and the Euclidean distance to the nearest drainage outlet is calculated). value)).

[0127] S22. Extract Static Feature Slices for Nodes: Extract static features from a 600-meter square grid area surrounding each road node. Align static features such as terrain-related static features, land use similar features, and drainage outlet replacement features with the rasterized geographic data to form static feature slices for the nodes. In this case, a side length of 600 meters for the square grid window is preferred, as this value can preserve key terrain change information while avoiding excessive redundancy.

[0128] S23. Generate a sampling subgraph for each road node: Each road node is used as a prediction node in turn. A breadth-first search algorithm is used. In this case, it is preferred to extract 15 neighboring nodes around the current prediction node and form a sampling subgraph of 16 nodes together with the current prediction node to ensure that a model input of the same scale is generated in different regions. Extracting 15 neighboring nodes can balance prediction accuracy and computational efficiency.

[0129] S24. Constructing Terrain-Aware Weights: To enable the G-R2R model to better capture the differences in flood risk along roads, this case introduces a terrain feature index as a terrain-aware weight in the graph adjacency matrix, replacing the traditional binary adjacency matrix, which improves the performance of the graph convolutional network in the G-R2R model during spatial propagation.

[0130] Construction and training of S3 and G-R2R models.

[0131] Construction of S31 and G-R2R models.

[0132] The G-R2R model comprises three modules: a static feature encoding module, a spatiotemporal feature encoding module, and a spatiotemporal feature decoding module.

[0133] (1) Static feature coding module

[0134] In this case, the static feature encoding module includes 17 static feature encoders. Each static feature encoder processes the static raster features of the nodes through a convolutional neural network (CNN) and reduces the feature dimension through pooling operations, ultimately outputting a one-dimensional static feature vector of length 128. These static encoders are divided into two parts: one part includes one static feature encoder for extracting the static features of the predicted node; the other part includes a set of 16 static feature encoders for extracting the static features of the sampled subgraph nodes (the sampled subgraph includes one predicted node and 15 neighboring nodes).

[0135] (2) Spatiotemporal feature coding module

[0136] The spatiotemporal feature encoding module includes an LSTM module and a spatiotemporal encoder module. The LSTM module processes rainstorm time series data to capture the temporal dependence between peak rainstorm intensity and changes in road network functional status. The spatiotemporal encoder module learns the spatiotemporal dynamic impact of rainfall and static features on changes in road network functional status. After rainfall data is input, it is processed by the LSTM module, and the output feature sequence is then concatenated with the historical RNF sequence of the subgraph before being input into the spatiotemporal encoder module.

[0137] (3) Spatiotemporal feature decoding module

[0138] The spatiotemporal feature decoding module includes a spatiotemporal convolutional block, a temporal control unit, and a fully connected neural network. The spatiotemporal convolutional block is used to learn the joint spatiotemporal correlation between dynamic and static feature information. The temporal control unit is used to regulate the output time step. Finally, feature decoding is achieved through the fully connected neural network to complete the prediction output of the future road network functional state.

[0139] Training of S32 and G-R2R models.

[0140] (1) In this case, it is preferable to use the historical time step. Single-step prediction duration The time windows were set to 12 hours and 3 hours respectively. A sliding time window method was used to segment the spatiotemporal sequences of 10,000 24-hour rainstorm time series and their corresponding road network functional status. The total length of the time window was 15 hours (including a 12-hour historical period and a 3-hour predicted period), and the sliding step size was consistent with the prediction duration. The RNF data from the first 12 hours were used as the input variable for the G-R2R model. The data from the following three hours were used as supervision labels for the output of the G-R2R model. For the region partitioning strategy, eight regions were selected to construct the training set, while the remaining two regions served as the validation and test sets, respectively, to evaluate the generalization performance of the G-R2R model in unobserved regions.

[0141] This case study uses the PyTorch deep learning framework to build a G-R2R agent model. The experimental environment is a Linux (Ubuntu 20.04 LTS) operating system, and the hardware configuration is an Intel Xeon Silver 4314 CPU (2.40GHz) with an NVIDIA GeForce RTX 4090 graphics processor.

[0142] The parameters of the G-R2R model are configured as follows: the static feature encoder adopts an alternating structure of convolutional and pooling layers, with a convolutional kernel size of 3×3 and a pooling kernel size of 2×2, and the feature channels of each layer vary. Figure 4As shown; the LSTM module has two hidden layers of size 32, and outputs the time length of a one-dimensional variable. = 4; The temporal and graph convolutional kernels in the spatiotemporal convolutional block are both 3 in size, with channel dimensions set to [1, 32, 64] and [64, 32, 128] respectively; The one-dimensional convolutional kernel size of the temporal control unit is... .

[0143] The training process of the G-R2R model uses the RMSprop optimizer, with a batch size of 1000 and an initial learning rate of 1×10⁻⁶. -3 The coefficient decays by 0.7 every 5 training epochs. To prevent overfitting, a dropout probability of 0.2 is introduced during the training of the G-R2R model, and an early stopping mechanism is employed.

[0144] (2) Model performance analysis

[0145] 1) Ablation Experiment Analysis

[0146] To verify the effectiveness of the static feature encoding module, three variant models were constructed for comparative analysis: (a) the G-R2R-nN model, which disables the predictive node feature encoding module; (b) the G-R2R-nS model, which disables the sampled subgraph feature encoding module; and (c) the G-R2R-nNS model, which completely disables the static feature encoding module, i.e., both the predictive node feature encoding module and the sampled subgraph feature encoding module are disabled. The ablation experiments used region G10 as the test set, region G9 as the validation set, and regions G1 to G8 as the training set for the model.

[0147] like Figure 7As shown, quantitative analysis based on mean absolute error (MAE) and mean squared error (MSE) indicates that the G-R2R model (the model of the method in this invention) integrating both prediction node and sampled subgraph static feature encoding modules exhibits the best prediction performance, with its MAE (0.0167) and MSE (0.0013) significantly lower than the other three ablation variant models. Notably, the G-R2R-nNS model, which completely disables the static feature encoding module, actually demonstrates superior error characteristics compared to the partially ablation models (G-R2R-nN and G-R2R-nS). This phenomenon suggests that enabling either prediction node feature encoding or sampled subgraph feature encoding modules alone not only fails to improve model performance but also weakens the model's learning ability. This case suggests that this may stem from two key factors: firstly, the high similarity of static features of adjacent nodes increases the difficulty of feature discrimination in the model; secondly, the overlap of sampled subgraph nodes makes it difficult for a single feature encoder to effectively extract discriminative subgraph features. Experimental results show that only through the co-coding mechanism of predicted node features and sampled subgraph features can the G-R2R model effectively identify neighboring road network nodes.

[0148] 2) Regional generalization ability assessment

[0149] To comprehensively evaluate the regional generalization ability of the G-R2R model described in this invention, a systematic cross-validation experiment was conducted in 10 flood-prone areas of City B. The experiment employed a leave-one-out strategy, selecting data from nine areas sequentially as the training set and using the remaining area as the test set to simulate the model's application in unobserved areas. The evaluation results of the 10 independent experiments are as follows: Figure 8 As shown, Figure 8 This demonstrates the mean absolute percentage error (MAPE) of the G-R2R model in each unobserved region and the coefficient of determination in the regression analysis. This was used to compare the model's predictive performance in different regions. The MAPE values ​​for all regions remained consistently between 2.3% and 3.54%. The model fit was lowest in region G8. The G7 region exhibited the best prediction accuracy. Despite the coefficient of determination While variations exist across different regions, the overall fluctuation range is relatively small, between 0.75 and 0.9, indicating that the G-R2R model exhibits stable predictive performance and excellent generalization ability. (Summary) Figure 8 The mean absolute percentage error (MAPE) of the forecast and the coefficient of determination for smoothing are relatively stable. As shown by the broken line, the G-R2R model possesses reliable migration potential in unobserved areas. This characteristic is of great value for constructing a standardized emergency response system at the provincial and municipal levels, and can provide technical support for unified and scientific disaster prevention and mitigation decision-making.

[0150] 3) Spatiotemporal error analysis

[0151] This case study uses region G9 as the unobserved area and analyzes the predictive performance and error distribution characteristics of the G-R2R model in rainstorm events from both temporal and spatial dimensions. In the temporal dimension, the focus is on examining the error distribution patterns of the model under different prediction step sizes in a single prediction and in iterative strategies. Figure 9 and Figure 10 As shown, the G-R2R model exhibits good stability in both short-term single-shot predictions and long-term iterative predictions. In this case, the prediction step size of the G-R2R model is... The time frame is set to 3 hours, meaning the G-R2R model can simultaneously predict the RNF changes over the next 3 hours in a single instance. Figure 9 Subplot (a) shows the distribution of the comprehensive relative error of the prediction results for the next 3 hours in a short-term single prediction. The results show that the relative errors of most road network nodes are concentrated in the range of ±1%, with 95% of the road network nodes having relative errors distributed in the range of -6.7% to 6.1%. This indicates that the G-R2R model also has excellent prediction performance in unobserved areas. It is worth noting that the maximum relative error reaches -49.1%, indicating that some road network nodes may experience prediction failure. This phenomenon will be explored in depth in subsequent studies with spatial analysis. Figure 9 Subplot (b) shows a box plot that quantitatively analyzes the evolution of the relative error at different time steps in a short-term single prediction with the prediction step length. The relative error at the first time step is 0.117%, while the relative errors at the second and third time steps only increase to 0.1322% and 0.1324%, respectively. Simultaneously, the interquartile range (IQR) gradually increases from an initial 0.53% (first time step) to 0.66% (third time step). These results indicate that although the dispersion of the relative error distribution gradually increases with the increase of the prediction time step, the error accumulation rate gradually decreases and converges. This characteristic fully demonstrates the excellent stability of the G-R2R model in the time dimension.

[0152] Considering that extreme rainfall events (especially persistent rainfall triggered by tropical cyclones) typically last for more than 3 hours, this case study adopts... Figure 6 The iterative prediction strategy shown is used to predict the road network functional status throughout the entire rainstorm process. For example... Figure 10As shown, the error range of the 12-hour iterative prediction of the G-R2R model remains within the range of -2% to 2%, and the interquartile range of the relative error exhibits a clear convergence characteristic. This excellent characteristic stems from the effective modeling capability of the LSTM module of the G-R2R model for the temporal dependence of rainfall, enabling it to suppress error divergence during long-term prediction and provide sufficient time windows for emergency planning. This stability of iterative prediction is the core advantage of the G-R2R model and is of great value for improving the reliability of emergency response planning.

[0153] This case study selects a single-peak rainstorm event as the test scenario to further analyze the predictive performance of the G-R2R model. The peak rainfall intensity of this rainstorm event reached 79 mm / hour, with a 24-hour cumulative rainfall of 313 mm. The test areas are G5, G7, G9, and G10. Each of these four test areas independently trains a G-R2R model, using data from its corresponding test area. Therefore, these four test areas are unobserved areas for their respective G-R2R models. Figure 11 As shown, where, Figure 11 Subplot (a) fully presents the hourly cumulative rainfall sequence of this rainstorm event. Figure 11 Subplot (b) shows the average functional state time histories of each road network node in the four test areas G5, G7, G9, and G10. It can be observed that the G-R2R model (solid line) and the physical simulation model (dashed line) exhibit a high degree of consistency in predicting the time-varying curves of the road network functional state (RNF). The results indicate that even in unobserved areas, the G-R2R model can accurately capture the dynamic characteristics of RNF changes, including key parameters such as the start time of RNF loss, the process of change, and the degree of maximum loss.

[0154] like Figure 12 As shown, where, Figure 12 Subplot (a) shows the spatial distribution of the prediction relative error in region G9 at hour 18 (the time of greatest road network functional loss). Spatially, the relative errors of the vast majority (over 90%) of the road network nodes are below 40%, and are clustered within 10%. To further analyze the characteristics of the error distribution, this case study correlates the relative error data with the RNF distribution, generating a scatter plot, as shown below. Figure 12 Subplot (b) shows the distribution of relative error with functional index. The analysis found that the relative error only exceeds 100% when the RNF value is small (less than 0.1). This is mainly due to the numerical amplification caused by the small denominator. The relative error in other cases is within an acceptable range.

[0155] Finally, focusing on the 18th hour of the aforementioned rainstorm event, which was the most destructive period, this case study further demonstrates the prediction results of different G-R2R models in the four regions of G5, G7, G9, and G10. Figure 13 As shown in the diagram, comparison reveals that the G-R2R model not only accurately predicts the temporal trend of RNF changes across the entire road network but also demonstrates strong spatial feature learning capabilities, particularly in identifying hotspots for emergency response planning. Therefore, the G-R2R model described in this invention can accurately predict the distribution characteristics of high-risk areas, thus providing a scientific basis for the formulation of refined emergency response plans.

[0156] II. Example 2:

[0157] This embodiment discloses a cross-regional prediction system for road network functional status under extreme rainstorms, including:

[0158] The dynamic dataset generation module is configured to generate a dynamic dataset of physical simulation of "rainstorm-traffic" based on extreme rainstorm scenarios in a preset area.

[0159] The static dataset generation module is configured to acquire and preprocess static features and road network data of a preset area to generate a static dataset.

[0160] The model building and training module is configured to build a cross-regional generalized prediction model for rainstorm-road network functional state, referred to as the G-R2R model. The G-R2R model is trained using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset.

[0161] The prediction application module is configured to input the rainstorm time series, static features and road network data of the target area into the trained G-R2R model, and output the road network functional state prediction sequence of the entire rainstorm process using an iterative prediction method.

[0162] For the system implementation, since it basically corresponds to the method implementation, the specific implementation process of the functions and roles of each module in the system can be found in the implementation process of the corresponding steps in the method described in Embodiment 1, and will not be repeated here.

[0163] The above description only outlines the basic principles and preferred embodiments of the present invention. Those skilled in the art can make many changes and modifications based on the above description, and these changes and modifications should fall within the protection scope of the present invention.

Claims

1. A cross-regional prediction method for the functional status of a road network under extreme rainstorms, characterized in that, Includes the following steps: Based on extreme rainstorm scenarios in a preset area, a dynamic dataset of "rainstorm-traffic" physical simulation is generated; Acquire static features and road network data of a preset area and preprocess them to generate a static dataset; A cross-regional generalized prediction model for rainstorm-road network functional state, referred to as the G-R2R model, is constructed. The G-R2R model is trained using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset. The rainstorm time series, static features, and road network data of the target area are input into the trained G-R2R model, and the road network functional state prediction sequence of the entire rainstorm process is output using an iterative prediction method.

2. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 1, characterized in that, The extreme rainstorm scenario based on the preset area generates a dynamic dataset of "rainstorm-traffic" physical simulation, specifically including: Based on the design rainfall patterns and historical records of extreme rainstorm events in the preset area, several rainstorm time series are generated through sampling methods. Based on the geographical and hydrological conditions of the preset area and the time series of the rainstorm, the urban flooding scenario is simulated, and the flooding scenario is mapped onto the road network to obtain the service level of all road sections under the extreme rainstorm scenario. Suppose that the road network includes a node set and an edge set, where the node set represents traffic nodes and the edge set represents road segments. Based on the service level of the road segments, a traffic accessibility index based on independent paths between road nodes is constructed, and the traffic function index of each road node is calculated.

3. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 1, characterized in that, The step of acquiring static features and road network data of a preset area and preprocessing them to generate a static dataset specifically includes: Obtain the static features of the preset area, including at least terrain static features, land use type features, and drainage outlet replacement features; Align the road nodes with the rasterized static features, take each road node in the preset area as the target node, and extract the square raster window of the preset range around the target node as the node static feature slice. For all road nodes within a preset area, generate a sampled sub-map for each road node; The road is divided into grid cells, and the terrain feature index of each grid cell is calculated to obtain the terrain perception weight of each road segment, thereby constructing a terrain perception adjacency matrix.

4. The method for cross-regional prediction of road network functional status under extreme rainstorms according to claim 3, characterized in that, The step of generating a sampled sub-map for each road node within a preset area specifically includes: Each road node within the preset area is sequentially used as a prediction node. A breadth-first search algorithm is used to extract a preset number of neighboring nodes around the current prediction node from the original road network of the preset area as a neighboring node set. The current prediction node and its neighboring node set are combined to form a sampling subgraph of the road node. Then, all road nodes within the preset area are traversed until a corresponding sampling subgraph is generated for each road node.

5. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 4, characterized in that, The G-R2R model includes a static feature encoding module, a spatiotemporal feature encoding module, and a spatiotemporal feature decoding module. The static feature encoding module encodes static features to obtain static features. The spatiotemporal feature encoding module encodes rainstorm time series and subgraph historical road network functional state series to obtain spatiotemporal dynamic features. The spatiotemporal feature decoding module fuses the static features obtained by the static feature encoding module and the spatiotemporal dynamic features obtained by the spatiotemporal feature encoding module to output a road network functional state prediction sequence for multiple future time steps.

6. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 5, characterized in that, The static feature encoding module includes a static feature encoder for encoding static feature slices of predicted nodes and a set of static feature encoders for encoding static feature slices of sampled subgraph nodes. Each static feature encoder includes multiple convolutional layers and pooling layers connected to each convolutional layer. The input of each static feature encoder is a multidimensional variable composed of grid static features corresponding to a road node. After multiple convolution and pooling operations, it becomes a one-dimensional vector. The one-dimensional vector is concatenated with the traffic function index of the road node under normal conditions and then subjected to one-dimensional convolution to output the static feature information of the road node.

7. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 6, characterized in that, The static feature variables obtained by encoding the static feature slices of the prediction nodes are broadcast and expanded along the height and width dimensions. At the same time, the static feature variables obtained by encoding the static feature slices of the sampled subgraph nodes are broadcast and expanded along the time dimension. After the broadcasting operation, two three-dimensional static feature variables are formed.

8. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 7, characterized in that, The spatiotemporal feature encoding module includes an LSTM module and a spatiotemporal encoder module. The LSTM module is used to process rainstorm time series data, capture the time dependency between the peak intensity of rainstorms and changes in road network functional status. The rainstorm time series processed by the LSTM module is concatenated with the historical road network functional status of the subgraph and then input into the spatiotemporal encoder module. The spatiotemporal encoder module is used to learn the spatiotemporal dynamic influence of rainfall and static features on changes in road network functional status, thereby outputting spatiotemporal feature variables. The spatiotemporal encoder module adopts a double-layer stacked spatiotemporal convolutional block structure. Each spatiotemporal convolutional block includes two time-gated units and a spatial graph convolutional unit set between the two time-gated units.

9. The cross-regional prediction method for road network functional status under extreme rainstorms according to claim 8, characterized in that, The two three-dimensional static feature variables output by the static feature encoding module are concatenated with the spatiotemporal feature variables output by the spatiotemporal feature encoding module to form a multi-dimensional feature variable, which is then input into the spatiotemporal feature decoding module. The spatiotemporal feature decoding module includes a spatiotemporal convolutional block, a temporal control unit, and a fully connected neural network. The spatiotemporal convolutional block is used to learn the joint spatiotemporal correlation between dynamic and static feature information. The temporal control unit is used to adjust the output time step. The fully connected neural network is used to perform feature decoding to complete the prediction output of the future road network functional state.

10. A cross-regional prediction system for the functional status of a road network under extreme rainstorms, characterized in that, include: The dynamic dataset generation module is configured to generate a dynamic dataset for physical simulation of "rainstorm-traffic" based on extreme rainstorm scenarios in a preset area. The static dataset generation module is configured to acquire and preprocess static features and road network data of a preset area to generate a static dataset. The model building and training module is configured to build a rainstorm-road network functional state prediction model for cross-regional generalization prediction, referred to as the G-R2R model. The G-R2R model is trained using the dynamic dataset of the "rainstorm-traffic" physical simulation and the static dataset. The prediction application module is configured to input the rainstorm time series, static features and road network data of the target area into the trained G-R2R model, and output the road network functional state prediction sequence of the entire rainstorm process using an iterative prediction method.

Images