Meteorological-driven rural highway risk prediction method

Abstract

The invention relates to a weather-driven rural highway risk prediction method, which comprises the following steps: constructing a road-weather double-graph structure based on input road network information and weather grid point information, and fusing road and weather information based on a cross-domain attention propagation mechanism to generate a road-weather fusion feature vector; introducing a road physical attribute vector based on the generated road-weather fusion feature vector, and performing feature interaction fusion and nonlinear modulation by fusing feature modulation of physical prior to obtain a feature vector after physical constraint modulation; the modulated feature vectors are input into an edge-center collaborative reasoning framework for risk prediction, and joint prediction of local quick response and global fine optimization is realized through collaborative calculation of edge nodes and center nodes; and based on a risk prediction result, performing model parameter updating through a prediction-inspection hierarchical feedback mechanism, and generating a visual risk map and a management strategy. According to the method, rural highway risk prediction with high precision, high real-time performance and long-term adaptive capability can be realized.

Description

Technical Field

[0001] This invention relates to the fields of intelligent transportation and meteorological information technology, specifically to a meteorological-driven method for predicting risks on rural roads. Background Technology

[0002] Rural roads are vital transportation links connecting villages with the outside world, and their safe and smooth operation is crucial for rural economic development and social stability. However, rural roads are mostly located in areas with complex terrain and variable weather. Extreme weather conditions such as blizzards, icing, heavy rain, and sandstorms can easily cause road icing, water accumulation, and low visibility, seriously threatening driving safety.

[0003] In existing technologies, several studies have attempted to predict road risks using meteorological data. For example, some schemes improve modeling accuracy by spatially matching and weighting roads with meteorological grid points (e.g., CN202310285678.9), but these methods are mostly static models and cannot reflect the spatiotemporal evolution of dynamic meteorological conditions such as real-time wind fields. Other studies have attempted to introduce physical rule embedding (e.g., CN202210987654.3), but these usually only consider basic factors such as temperature-dew point difference and slope, ignoring key physical properties such as road surface material and drainage level, resulting in insufficient physical rationality of the model. In addition, although some schemes propose edge-center hierarchical prediction to improve computational efficiency (e.g., CN202310564321.8), these mostly remain at the result interaction level and fail to effectively solve the long-term drift problem of edge models in weak network environments.

[0004] In summary, existing methods still have significant shortcomings in areas such as spatial dynamic modeling, enhanced physical constraints, real-time response, and long-term adaptability. Therefore, there is an urgent need for a systematic, highly adaptive, and physically interpretable rural road risk prediction scheme. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a weather-driven method for predicting rural road risks. This method achieves high-precision, high-real-time, and long-term adaptive rural road risk prediction by constructing a complete technical system from data perception to strategy generation.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a weather-driven method for predicting rural road risks, comprising the following steps:

[0007] Step S1: Based on the input road network information and meteorological grid information, construct a road-meteorological dual-map structure, including a road map, a meteorological map, and their cross-domain correlation matrix. Based on the cross-domain attention propagation mechanism, perform weighted perception on the dynamic impact of meteorological grid points to fuse road and meteorological information and generate a road-meteorological fusion feature vector.

[0008] Step S2: Based on the generated road-meteorological fusion feature vector, the road physical attribute vector is introduced. By fusing the physical prior feature modulation, feature interaction fusion and nonlinear modulation are performed to obtain the feature vector after physical constraint modulation.

[0009] Step S3: Input the modulated feature vector into the edge-center collaborative reasoning framework for risk prediction, and achieve joint prediction of local fast response and global fine optimization through collaborative computation between edge nodes and center nodes; wherein, edge nodes perform lightweight local reasoning, center nodes perform multi-source data fusion and global optimization, and maintain parameter consistency through a collaborative update mechanism, forming a two-layer reasoning output from local prediction to global optimization.

[0010] Step S4: Based on the risk prediction results, update the model parameters through a prediction-inspection hierarchical feedback mechanism, and generate a visualized risk map and management strategy.

[0011] Furthermore, step S1 specifically includes:

[0012] Construct a road map based on the input road network information. Road map Based on road topology, adjacent road nodes are connected by edges to describe the topological structure and traffic elements of road nodes; where the node set... In this context, each node represents a road node, i.e., a road segment; the edge set... An edge in the middle represents the connection relationship between adjacent road nodes;

[0013] Meteorological maps are constructed based on the input meteorological grid information. Weather map Based on meteorological nodes, adjacent meteorological nodes are connected by edges to reflect the spatial proximity and dynamic wind field evolution between meteorological grid points; among them, the node set In this context, a node represents a meteorological grid point, and the set of edges... An edge in the middle represents the spatial connection between adjacent meteorological grid points;

[0014] To establish a cross-domain mapping between road nodes in the road map and meteorological grid points in the meteorological map, a cross-domain correlation matrix is ​​constructed. ; Middle elements , Representing meteorological grid points Located at road node Within their respective spheres of influence, the two entities have a cross-domain adjacency relationship. This indicates that the two are not directly related; Only the cross-domain connection structure is depicted, without carrying propagation strength information; the propagation strength of cross-domain information is determined by the subsequent physical propagation weights. Attention coefficient Dynamically determined;

[0015] Combined with road map Weather map and cross-domain association matrix It forms a road-meteorological dual-map structure;

[0016] Design road map Middle Road Node Characteristic Weather map China Meteorological Grid Characteristic The attention coefficient is calculated according to formula (1). Used to characterize meteorological grid points For road nodes Semantic relevance:

[0017] (1)

[0018] in, It is a linear transformation matrix. For learnable attention vectors, for Defined cross-domain neighborhood, Road nodes Meteorological grids in cross-domain neighbors Features This represents the natural exponential function. represents the linear rectified activation function with leakage, and || represents the concatenation operation;

[0019] Integrating physical propagation weights into attention mechanisms The attention coefficient after physical enhancement is obtained. ;

[0020] Then, the meteorological enhancement characteristics of the road nodes are calculated according to formula (2):

[0021] (2)

[0022] in, Indicates the activation function; This is a weighted fusion result of meteorological elements and dynamic wind field at road nodes; the obtained... This is the road-meteorological fusion feature vector.

[0023] Furthermore, the physical propagation weights The calculation method is as follows:

[0024] Spatial proximity weights are defined to characterize spatial proximity constraints, reflecting the geographical attenuation characteristics of the influence of meteorological grid points on road nodes. The calculation formula is shown in equation (3):

[0025] (3)

[0026] in, Representing meteorological grid points With road nodes Euclidean distance, To prevent division by zero errors, This is the distance attenuation factor;

[0027] The directional consistency weight is defined to characterize the directional consistency constraint, and its calculation formula is shown in equation (4):

[0028] (4)

[0029] in, The direction angle of the line connecting the road node and the meteorological grid point. For local wind direction, This is an adjustment coefficient used to control the sensitivity to wind direction differences;

[0030] Then, calculate the physical propagation weights according to formula (5):

[0031] (5)

[0032] in, This is the balance coefficient.

[0033] Furthermore, in step S2, the implementation method of the feature modulation mechanism that integrates physical priors is as follows:

[0034] The generated road-meteorological fusion feature vector With the introduced road physical attribute vector Each inputs a separate feature extraction branch;

[0035] The feature interaction module dynamically fuses the features extracted from the two branches, employing an attention weighting mechanism to capture the coupling relationship between meteorological elements and road attributes. Then, a nonlinear mapping function is used to map the fused features, generating a feature vector modulated by physical constraints. Its expression is shown in equation (6):

[0036] (6)

[0037] in, Indicates feature-level interactive operations. It is a nonlinear mapping function. , , These are the weighting coefficients.

[0038] Furthermore, in step S3, the edge-center collaborative reasoning framework includes edge nodes and center nodes;

[0039] The edge nodes are deployed with lightweight temporal convolutional networks for local risk prediction; the lightweight temporal convolutional networks undergo model pruning and parameter quantization; the model pruning compresses the network structure by removing neurons and connections with low prediction contributions; the parameter quantization reduces model storage and transmission costs by converting floating-point parameters into low-ratio point-to-point representations.

[0040] The central node performs the following tasks:

[0041] Multi-source data fusion: It integrates numerical weather forecasts, radar observation data, and prediction results from multiple edge nodes, and bridges the differences in spatiotemporal resolution between different data sources through feature alignment and spatiotemporal interpolation.

[0042] Global model optimization: A deep fusion model is used to jointly model multi-dimensional meteorological elements, and error compensation is achieved through statistical regression or residual correction;

[0043] Parameter synchronization and knowledge transfer: The optimized parameter set is compressed and distributed to each edge node to realize model weight update and knowledge transfer;

[0044] The collaborative update mechanism achieves the following tasks:

[0045] Adaptive task allocation: When network latency is high, switch to edge-independent operation mode to ensure the real-time performance of local predictions; when network conditions are good, switch to central collaborative optimization mode to achieve global synchronization and joint training.

[0046] Results Upload and Parameter Feedback: Edge nodes periodically upload summarized prediction results, while central nodes send parameter update packages to complete bidirectional model synchronization;

[0047] Drift detection and update triggering: A local drift detection module is embedded in the edge node to monitor prediction deviations and actively trigger parameter synchronization requests.

[0048] Furthermore, in step S4, the prediction-inspection hierarchical feedback mechanism includes:

[0049] 1) Construct a multi-timescale hierarchical feedback mechanism: Construct a two-layer feedback structure with short and long cycles in the time dimension; edge nodes collect meteorological and inspection data in real time and perform incremental fine-tuning to achieve rapid self-correction; the central node performs hyperparameter optimization and model retraining based on historical and multi-source feedback to achieve continuous generalization of seasonality and regionality; a closed loop is formed through the uplink inspection feedback of edge nodes and the downlink parameter update of central nodes to achieve self-learning and dynamic evolution at multiple time scales;

[0050] 2) Construct a model self-learning and parameter evolution mechanism: Maintain a global parameter cache at the central node. Record the optimal model state under different time windows; within a short period, edge nodes update local parameters and upload summaries. Over a long period, the central node utilizes... Aggregate parameter gradients, perform global retraining and knowledge distillation, and update the main model. ;

[0051] 3) Construct a risk mapping and strategy generation module: This module generates the output risk prediction results. Transform into a geographic risk layer And automatically generate management policies. :

[0052]

[0053]

[0054] in, For space mapping functions, A rule- and case-based strategy generator;

[0055] 4) Achieving a complete closed loop of closed-loop optimization and system evolution through collaboration between upper and lower layers, including:

[0056] Edge nodes perform prediction and local correction;

[0057] Periodic optimization and parameter distribution of the central node;

[0058] The strategy generation module translates the prediction results into actionable measures;

[0059] The feedback from the execution further drives the model's iterative learning.

[0060] Furthermore, the dual-layer feedback structure includes:

[0061] Short-cycle edge feedback: Edge nodes trigger incremental updates of local parameters based on the deviation between real-time inspection data and prediction results;

[0062] Long-term central feedback: The central node performs hyperparameter optimization and global model retraining based on the bias summary and historical data uploaded by multiple edge nodes.

[0063] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method.

[0064] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0065] Compared with existing technologies, this invention has the following advantages: This invention provides a meteorology-driven rural road risk prediction method. This method achieves a detailed characterization of the spatiotemporal evolution characteristics of meteorology through a dual-graph structure and cross-domain attention, improving the dynamism and accuracy of risk perception. At the same time, by fusing physical prior feature modulation, it significantly enhances the physical consistency of the model and the interpretability of the prediction results, providing a scientific basis for decision-making. In addition, through edge-center collaborative reasoning and adaptive communication, it achieves efficient and stable distributed prediction in weak network and low computing power environments, ensuring the real-time performance and robustness of the system. Finally, through hierarchical feedback and policy generation, it realizes continuous optimization of the model and the executable implementation of prediction results, forming a self-evolving intelligent closed-loop system with long-term adaptability and practical value. Attached Figure Description

[0066] Figure 1 This is a flowchart of the meteorological-driven rural road risk prediction method in an embodiment of the present invention;

[0067] Figure 2 This is an overall structural diagram of the meteorological-driven rural road risk prediction method in this embodiment of the invention;

[0068] Figure 3 This is a schematic diagram illustrating the structural relationship between road maps, weather maps, and cross-domain correlation matrices in an embodiment of the present invention;

[0069] Figure 4 This is a schematic diagram of the feature modulation mechanism structure that integrates physical priors in an embodiment of the present invention;

[0070] Figure 5 This is a schematic diagram of a lightweight hierarchical inference framework for edge-center collaboration in an embodiment of the present invention;

[0071] Figure 6 This is a schematic diagram of the prediction-inspection hierarchical feedback and visualization strategy generation mechanism in an embodiment of the present invention. Detailed Implementation

[0072] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0073] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0074] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0075] like Figure 1 As shown, this invention provides a weather-driven method for predicting rural road risks, comprising the following steps:

[0076] Step S1: Based on the input road network information and meteorological grid information, construct a road-meteorological dual-map structure, including a road map, a meteorological map, and their cross-domain correlation matrix. Based on the cross-domain attention propagation mechanism, perform weighted perception on the dynamic impact of meteorological grid points to fuse road and meteorological information and generate a road-meteorological fusion feature vector.

[0077] Step S2: Based on the generated road-meteorological fusion feature vector, the road physical attribute vector is introduced. By fusing the physical prior feature modulation, feature interaction fusion and nonlinear modulation are performed to obtain the feature vector after physical constraint modulation.

[0078] Step S3: Input the modulated feature vector into the edge-center collaborative reasoning framework for risk prediction, and achieve joint prediction of local fast response and global fine optimization through collaborative computation between edge nodes and center nodes; wherein, edge nodes perform lightweight local reasoning, center nodes perform multi-source data fusion and global optimization, and maintain parameter consistency through a collaborative update mechanism, forming a two-layer reasoning output from local prediction to global optimization.

[0079] Step S4: Based on the risk prediction results, update the model parameters through a prediction-inspection hierarchical feedback mechanism, and generate a visualized risk map and management strategy.

[0080] The implementation process of the method proposed in this invention will be described in detail below.

[0081] 1. Overview of this method

[0082] The overall structure of this method is as follows: Figure 2As shown. Its inputs include road network information and meteorological grid information, which are used to form road maps. With weather map Then, through four core modules acting sequentially on the input, feature fusion, inference prediction, and policy feedback are gradually achieved.

[0083] (1) Dynamic perception model of road-meteorological dual map

[0084] based on and Construct a road-meteorological dual-map structure and use a cross-domain correlation matrix. (This can also be viewed as a road-meteorological coupling matrix or feature mapping) to achieve feature propagation between the two domains. The model combines spatial proximity and directional consistency constraints, employing a cross-domain attention mechanism to weightedly perceive the dynamic impact of meteorological grid points, outputting a road-meteorological fusion feature vector. The model enables spatiotemporal information interaction driven by meteorology, providing dynamic feature inputs for subsequent modeling.

[0085] (2) Feature modulation mechanism that integrates physical priors

[0086] In fusion features Based on this, road physical properties are introduced. (e.g., slope, road surface material, drainage grade), a feature modulation mechanism (FMM) integrating physical priors is constructed. The interaction between meteorological elements and physical characteristics is characterized by a nonlinear modulation function, generating modulation features with physical constraints. This module explicitly demonstrates the physical formation mechanisms of risks such as freezing and water accumulation, enhancing the physical plausibility and interpretability of the model.

[0087] (3) Lightweight hierarchical inference strategy with edge-center synergy

[0088] Modulation features The input is fed into the Edge-Center Collaborative Inference Framework (ECCIS) to achieve distributed prediction deployment of the model. Edge nodes deploy lightweight TCNs to perform fast local predictions, while the central node fuses NWP and radar data for global fine-grained optimization. The two levels of nodes maintain parameter consistency through a collaborative update mechanism, forming a two-layer inference output from local prediction to global optimization, based on the road node risk probability distribution. express.

[0089] (4) Prediction-inspection hierarchical feedback and visualization strategy generation

[0090] Building upon the inference output of ECCIS, a Prediction-Inspection Hierarchical Feedback Mechanism (PI-HFV) is further constructed to achieve hierarchical feedback across the time dimension. Short-cycle parameter corrections are performed at the edge, while long-cycle hyperparameter updates are performed at the center. Combining the "Technical Specifications for Rural Road Maintenance" and historical cases, risk maps and strategy recommendations are generated, achieving an intelligent closed loop of "prediction—decision-execution—relearning."

[0091] This method system achieves full-process modeling of meteorological-driven rural road risk prediction through a four-layer collaborative structure of "data modeling—feature modulation—hierarchical inference—strategy feedback." The first two modules focus on dynamic fusion and physical constraint enhancement at the feature layer, while the latter two modules are responsible for efficient inference and closed-loop feedback at the system layer. This method combines spatiotemporal accuracy, physical consistency, real-time responsiveness, and long-term adaptability, theoretically forming a systematic meteorological risk modeling framework, and possessing engineering deployability and practical application value.

[0092] 2. Road-Weather Dual-Map Dynamic Sensing Model

[0093] This method proposes a dynamic perception model based on a road-meteorological dual-map. By constructing a heterogeneous dual-map structure of road network and meteorological grid points, and introducing a cross-domain attention propagation mechanism enhanced by physical constraints, it enables road nodes to achieve weighted perception of local meteorological elements and dynamic wind fields. Under the combined effect of spatial proximity and directional consistency constraints, the model effectively characterizes the geometric structural relationship and atmospheric dynamic coupling relationship between roads and meteorology, thereby significantly improving the spatiotemporal response capability and perception accuracy of road risk prediction.

[0094] 2.1 Construction and Dynamic Perception of Dual-Graph Structure

[0095] To characterize the spatial, physical, and dynamic coupling characteristics between the road system and the meteorological field, this method constructs a system based on road maps. Weather map Cross-domain association matrix The resulting heterogeneous bigraph structure. Among them, Describe the topology and traffic elements of road nodes; Reflecting the spatial proximity and dynamic wind field evolution between meteorological grid points; and Define the meteorological neighborhood structure within the sensing range of road nodes. This is a binary (0 / 1) matrix that only characterizes the cross-domain connection structure and does not carry propagation strength information. The cross-domain information propagation strength is determined by the subsequent physical constraint weights. Attention coefficient It is determined dynamically. Figure 3 It illustrates the structural relationships between road maps, weather maps, and cross-domain correlation matrices.

[0096] (1) Road map construction

[0097] Construct a road map based on the input road network information. Road map Based on road topology, adjacent road segment nodes are connected by edges to describe the topological structure and traffic elements of road nodes. The node set... In this context, each node represents a road node, i.e., a road segment; the edge set... This represents the connection relationship between adjacent roads. (Edge set) Each edge represents a connection between adjacent road nodes. weight Indicates traffic flow or connectivity strength. Node feature vector. It includes factors such as traffic flow, vehicle speed, vehicle density, road conditions, and geographical attributes.

[0098] (2) Construction of meteorological maps

[0099] Meteorological maps are constructed based on the input meteorological grid information. Weather map Based on meteorological nodes, adjacent meteorological nodes are connected by edges to reflect the spatial proximity and dynamic wind field evolution between meteorological grid points. The node set... In this context, a node represents a meteorological grid point, and the set of edges... Represents the spatial relationship between adjacent grid points. Edge set. Each edge represents a spatial connection between adjacent meteorological grid points. weight Calculated based on grid distance or wind field trajectory, it reflects spatial proximity or atmospheric dynamic correlation. Node characteristics. Meteorological elements include temperature, humidity, precipitation, visibility, wind speed, and wind direction. Among these, wind speed and wind direction constitute the dynamic wind field input and are key factors affecting cross-regional propagation.

[0100] (3) Construction of cross-domain association matrix

[0101] To establish a cross-domain mapping between road nodes in the road map and meteorological grid points in the meteorological map, a cross-domain correlation matrix is ​​constructed. This matrix takes into account both geographical proximity and wind propagation characteristics, enabling road nodes to receive dynamic information from neighboring meteorological grid points. Middle elements , Representing meteorological grid points Located at road node Within their respective spheres of influence, the two entities have a cross-domain adjacency relationship. This indicates that the two are not directly related. This matrix is ​​only used to define the neighborhood structure in the cross-domain attention propagation mechanism; its role is to specify which meteorological nodes can provide information to road nodes, but it does not carry propagation strength or physical weights. Subsequent cross-domain attention mechanisms... Calculate the attention coefficient within the defined neighborhood range And further combined with physical propagation weights Constructing physical enhancement of attention This enables road nodes to achieve weighted perception of meteorological elements and dynamic wind fields.

[0102] Combined with road map Weather map and cross-domain association matrix It forms a dual-map structure of road and weather maps. Among them, Provide road traffic characteristics, Provide meteorological grid features, Establish information mapping relationships between the two domains.

[0103] 2.2 Cross-domain attention propagation mechanism

[0104] To enable road nodes to dynamically perceive local weather changes and wind field transmission, this method constructs a cross-domain attention propagation mechanism based on Graph Attention Network (GAT) and combines it with a physical augmentation strategy to form a hybrid weighted model of "data-driven + physical guidance".

[0105] Design road map Middle Road Node Characteristic Weather map China Meteorological Grid Characteristic .

[0106] (1) Calculation of cross-domain attention coefficient

[0107] Calculate attention coefficient Used to characterize meteorological grid points For road nodes The semantic relevance is defined as shown in equation (1).

[0108] (1)

[0109] in, It is a linear transformation matrix. For learnable attention vectors, for Defined cross-domain neighborhood, Road nodes Meteorological grids in cross-domain neighbors Features This represents the natural exponential function. represents the linear rectified activation function with leakage, and || represents the concatenation operation.

[0110] Equation (1) can be used to capture the semantic impact of meteorological variables on road conditions.

[0111] (2) Physical mechanisms to enhance attention

[0112] Data-driven attention methods may overlook physical propagation patterns, such as wind direction and propagation paths, and the extent of local spatial influence. Therefore, this method constructs physical constraint weights. It is then incorporated into the attention mechanism to obtain a physically enhanced attention coefficient. This physical enhancement ensures that: spatially, more attention is paid to grid points that are closer together; directionally, propagation paths aligned with the wind direction are prioritized; and the dynamic evolution of the meteorological field more closely matches the actual wind field propagation patterns.

[0113] (3) Update of meteorological enhancement features at road nodes

[0114] The meteorological enhancement characteristics of the final road node are defined by equation (2).

[0115] in: For the activation function ReLU, It is a linear transformation matrix, and is the same as W in equation (1), while This represents the attention coefficient after physical enhancement.

[0116] (2)

[0117] in, Represents the activation function ReLU; It is a linear transformation matrix; The weighted fusion result of meteorological elements and dynamic wind field at road nodes is a key input for subsequent feature modulation and risk prediction modules. This is the road-meteorological fusion feature vector.

[0118] 2.3 Spatial Proximity and Orientation Consistency Constraints

[0119] To ensure that cross-domain information propagation conforms to spatial geometric relationships and atmospheric dynamics, this method introduces spatial proximity constraints and directional consistency constraints to jointly construct physical propagation weights. .

[0120] (1) Spatial proximity constraint

[0121] Spatial proximity weights are defined to characterize spatial proximity constraints, reflecting the geographical attenuation characteristics of the influence of meteorological grid points on road nodes. The calculation formula is shown in equation (3):

[0122] (3)

[0123] in, Representing meteorological grid points With road nodes Euclidean distance, To prevent division by zero errors, This is the distance attenuation factor.

[0124] (2) Directional consistency constraint

[0125] Based on the direction angle of the line connecting road nodes and meteorological grid points With local wind direction The relationship between them is defined by the directional consistency weight, which characterizes the directional consistency constraint. Its calculation formula is shown in equation (4):

[0126] (4)

[0127] in, The direction angle of the line connecting the road node and the meteorological grid point. For local wind direction, This is an adjustment coefficient used to control the sensitivity to wind direction differences.

[0128] The directional consistency weight defined in equation (4) can make information propagation follow the dominant wind direction and reflect the transmission characteristics of dynamic wind fields.

[0129] (3) Physical propagation weight fusion

[0130] Taking into account both spatial and wind field factors, the physical propagation weight is finally calculated according to equation (5):

[0131] (5)

[0132] in, This is a balance coefficient used to adjust the contribution ratio of the two types of physical constraints.

[0133] The weights defined in Equation (5) can ensure that the cross-domain propagation path and the intensity of influence simultaneously satisfy the spatial structure and atmospheric dynamic laws, thereby improving the spatial accuracy and time response capability of road risk prediction.

[0134] 3. Feature modulation mechanism integrating physical priors

[0135] The input to the Feature Modulation Mechanism (FMM), which incorporates physical priors, includes road-meteorological fusion features from Part 2. With road physical properties The output is a feature vector modulated by physical constraints. FMM significantly improves both prediction accuracy and physical consistency by explicitly modeling the interaction between meteorological elements and physical characteristics.

[0136] 3.1 The Necessity and Challenges of Physical Prior Features

[0137] Existing weather-driven risk prediction models often rely solely on meteorological data or oversimplify the treatment of physical attributes, failing to reflect the moderating effect of road structure on weather risk. Under extreme weather conditions, these physical characteristics can amplify or suppress the impact of meteorological factors; ignoring these elements will lead to systematic prediction biases in models under extreme conditions. Key physical factors of concern include:

[0138] (1) Road surface material

[0139] Different materials (asphalt, cement, gravel, etc.) respond significantly differently to freezing and water accumulation. Asphalt pavements are more prone to water accumulation and freezing, while gravel pavements have higher permeability.

[0140] (2) Drainage grade

[0141] Drainage systems affect the rate at which water is discharged after rainfall. High-quality drainage systems can quickly remove accumulated water and reduce the risk of skidding; poor drainage, on the other hand, can lead to localized flooding.

[0142] (3) Gradient and traffic flow

[0143] Road sections with steeper slopes are more prone to slipping or landslides under icy conditions, and traffic flow characteristics under different slopes will also change the risk level.

[0144] Therefore, how to integrate these physically constrained static features into the model and coordinate them with dynamic meteorological elements in modeling is a key challenge in achieving high-reliability predictions.

[0145] 3.2 Design Principles of Physical Prior Feature Modulation

[0146] The physical prior feature modulation mechanism aims to adaptively adjust the influence of physical features on meteorological elements in response to environmental changes through dynamic interactive modeling. Its design follows three principles:

[0147] (1) Dynamic interaction between physical and meteorological characteristics

[0148] FMM explicitly models the non-steady-state correlation between road surface material, drainage grade and meteorological conditions, enabling the model to adaptively adjust feature weights under different meteorological scenarios and reflect real physical laws.

[0149] (2) Multi-level feature fusion

[0150] The FMM employs a dual-branch structure to extract meteorological and physical features separately, and then fuses them at a higher level through an interaction layer, thereby capturing physical-meteorological dependencies at different levels.

[0151] (3) Nonlinear mapping modeling

[0152] FMM uses nonlinear activation functions (such as sigmoid and tanh) to map interaction features in order to characterize complex nonlinear response features and avoid information loss caused by linear assumptions.

[0153] 3.3 Implementation process of feature modulation mechanism

[0154] To achieve the Feature Modulation Mechanism (FMM) that integrates physical priors, this study designed the following... Figure 4 The multi-level feature interaction structure is shown. Module input includes fused features from the output of the road-meteorological bi-map model in Part 2. With the physical property vector of the road After the two types of features are extracted through independent branches, they are dynamically fused in the feature interaction module, outputting a feature vector modulated by physical constraints. This provides high-dimensional, interpretable input for the subsequent edge-center collaborative reasoning module.

[0155] (1) Feature interaction modeling

[0156] After extracting meteorological and physical features, an interactive network is used to achieve feature fusion. This network employs an attention weighting mechanism to capture typical coupling relationships such as "temperature-material" and "precipitation-drainage," enabling explicit modeling and reinforcement of the nonlinear dependence between meteorological elements and road physical factors.

[0157] (2) Design of characteristic modulation function

[0158] Based on the interactive features, a nonlinear modulation function is introduced to achieve dynamic weight adjustment of the meteorological-physical correlation. For example, in freezing conditions, the interactive weight of temperature and road surface material is automatically amplified; in heavy rainfall conditions, the weight of drainage level increases significantly. The modulation function is defined as shown in equation (6).

[0159] (6)

[0160] in, Indicates feature-level interactive operations. It is a non-linear mapping function.

[0161] Through the above modulation process, the model can adaptively adjust the importance of road physical features according to different meteorological scenarios, and realize the dynamic fusion of meteorological and physical dual-domain features.

[0162] 3.4 Enhanced Physical Rationality and Model Interpretability

[0163] The feature modulation mechanism that integrates physical priors significantly improves the physical consistency and interpretability of the model. Specifically, in icing risk prediction, the model can quantify the interaction weights of temperature, precipitation, and road surface material to reveal the physical reasons for the differences in icing risk under different materials; at the same time, in water accumulation risk prediction, the model explains the differences in water accumulation on various roads based on the interaction term between drainage level and precipitation intensity.

[0164] This feature modeling with physical constraints provides a clear physical basis for the prediction results, enhancing the model's credibility and offering scientific support for road maintenance and emergency dispatch. The modulation features output by this module... This will serve as input for subsequent edge-center collaborative reasoning strategies, providing physically consistent feature support for multi-level reasoning.

[0165] 4. Lightweight hierarchical inference strategy with edge-center collaboration

[0166] The Edge-Center Collaborative Inference Strategy (ECCIS) uses the modulation features of the output in Part 3. Using the edge nodes and the central nodes as input, joint prediction of local rapid response and global fine optimization is achieved through collaborative computing.

[0167] This strategy achieves "spatiotemporal decoupling and two-layer collaboration" in its architecture and introduces bandwidth-aware parameter synchronization and drift suppression mechanisms in its operational mechanism, thereby enabling efficient inference of meteorological risks in weak network and low computing power environments. The final output is the road node risk prediction result. .

[0168] 4.1 Overview of the Periphery-Center Hierarchical Reasoning Framework

[0169] To adapt to the limited network and computing power environment in rural areas, ECCIS designed a hierarchical inference system of "edge fast inference - central fine-grained optimization - bidirectional collaborative update". For example... Figure 5 As shown, the system consists of three parts: an edge node layer, a central node layer, and a collaborative update mechanism.

[0170] (1) Edge nodes

[0171] Deploy a lightweight temporal convolutional network (TCN) to perform fast inference on time-series data collected by meteorological sensors and achieve low-latency response.

[0172] (2) Central node

[0173] It integrates numerical weather prediction (NWP), radar observations, and prediction results from multiple edge nodes, and performs global fine-grained optimization and parameter consistency updates.

[0174] (3) Same update mechanism:

[0175] The prediction results are uploaded and parameters are returned through bidirectional information flow, and the synchronization frequency is adaptively adjusted according to network bandwidth and computing load.

[0176] The system achieves spatiotemporal decoupling in its structure: edge nodes are responsible for short-term high-frequency forecasts, while central nodes perform long-term global optimization, thereby constructing a meteorological risk classification inference system that combines real-time performance, accuracy, and scalability.

[0177] 4.2 Lightweight Inference Design for Edge Nodes

[0178] Edge nodes are deployed along rural roads and need to perform local weather forecasting under conditions of low power consumption, limited storage, and limited bandwidth. This study adopts a lightweight TCN model at the edge, captures long-term dependent features through convolutional dilation structures, and significantly improves operating efficiency by combining model compression techniques.

[0179] (1) Model pruning

[0180] By removing neurons and connections with low prediction contributions, computational redundancy is reduced and the network structure is compressed.

[0181] (2) Parameter quantization

[0182] Converting floating-point parameters to low-ratio fixed-point representations significantly reduces model storage and transmission costs.

[0183] (3) Energy consumption optimization

[0184] While maintaining a performance rate of over 95%, computational load is reduced by approximately 60%, enabling millisecond-level real-time prediction on low-power devices.

[0185] The optimized TCN model can run stably in weak network and low computing power environments, providing continuous and reliable local prediction inputs for the central node.

[0186] 4.3 Multi-source fusion and global optimization of the central node

[0187] The central node has high computing power and large storage capacity, and mainly undertakes three tasks: multi-source meteorological data fusion, global model optimization, and parameter synchronization.

[0188] (1) Multi-source data fusion

[0189] By integrating NWP, radar observation, and multi-edge node prediction results, and using feature alignment and spatiotemporal interpolation methods, the differences in spatiotemporal resolution between different data sources are bridged, thereby improving the integrity and consistency of input data.

[0190] (2) Global model optimization

[0191] A deep fusion model is used to jointly model multidimensional meteorological elements, and error compensation is achieved through statistical regression or residual correction to enhance the response capability to extreme weather events.

[0192] (3) Parameter synchronization and knowledge transfer

[0193] The optimized parameter set After compression, the data is distributed to each edge node to update model weights and transfer knowledge, thereby maintaining the consistency and long-term stability of the global prediction results.

[0194] 4.4 Collaborative Update and Adaptive Communication Mechanism

[0195] The collaborative update mechanism is the core of ECCIS, used to maintain information consistency and performance balance among multiple nodes. This mechanism achieves adaptive task allocation and continuous optimization through a dynamic scheduling strategy that is both bandwidth-aware and computationally load-aware.

[0196] (1) Adaptive task allocation

[0197] When network latency is high, the system automatically switches to "edge independent operation" mode to ensure the real-time performance of local predictions; when network conditions are good, it switches to "central collaborative optimization" mode to achieve global synchronization and joint training.

[0198] (2) Result Upload and Parameter Feedback

[0199] Edge nodes periodically upload summarized prediction results (bias statistics, confidence scores, etc.), and the central node generates parameter update packages after comprehensive analysis and distributes them to complete the two-way synchronization of the model.

[0200] (3) Drift detection and update triggering

[0201] A local drift detection module is embedded in the edge nodes to monitor prediction bias and proactively trigger parameter synchronization requests, effectively suppressing long-term model drift problems. Through the "low-frequency global synchronization + high-frequency local adaptation" operating mode, ECCIS achieves a dynamic balance between computational efficiency and prediction accuracy under low bandwidth conditions.

[0202] The edge-center collaborative lightweight hierarchical inference strategy (ECCIS) achieves a hierarchical allocation of computational tasks in its spatial structure: edge nodes deploy pruned and quantized lightweight temporal convolutional networks (TCNs) to quickly respond to local meteorological changes; the central node integrates numerical weather prediction (NWP), radar observations, and prediction results from multiple edge nodes to achieve global refined prediction and parameter synchronization. Simultaneously, by introducing a bandwidth-aware collaborative update mechanism, ECCIS can adaptively adjust the model synchronization frequency in weak network and low-computing-power environments, achieving a balance between real-time performance and accuracy. ECCIS ensures efficient collaboration in a spatially distributed environment, significantly improving the dynamic response capability of meteorological risk prediction. However, ECCIS still belongs to a static inference layer structure; model parameter updates rely on the periodic computation of the central node and lack real-time self-learning capabilities in response to meteorological changes and inspection feedback. Therefore, Part 5 introduces a temporal hierarchical and policy feedback mechanism based on ECCIS to achieve dynamic model evolution and policy implementation.

[0203] 5. Predictive – Inspection Layered Feedback and Visualization Strategy Generation Mechanism

[0204] This section, building upon the Edge-Center Collaborative Inference Framework (ECCIS) in Part 4, further constructs a Prediction-Inspection Hierarchical Feedback and Visualization (PI-HFV) mechanism. This mechanism aims to enable the prediction model to possess continuous self-learning, cross-scale adaptation, and policy implementation capabilities through short-cycle parameter correction and long-cycle hyperparameter updates, forming an intelligent evolutionary closed loop of "prediction—decision—feedback—relearning."

[0205] 5.1 Design of a Multi-Time-Scale Hierarchical Feedback Mechanism

[0206] PI-HFV constructs a two-layer feedback structure in the time dimension, consisting of short-cycle (edge ​​nodes) and long-cycle (central nodes), specifically as follows: Figure 6 As shown, edge nodes collect meteorological and inspection data in real time and perform incremental fine-tuning to achieve rapid self-correction; the central node optimizes hyperparameters and retrains the model based on historical data and multi-source feedback to ensure continuous generalization of seasonality and regionality. The system consists of a central node, edge nodes, and a policy generation module. It forms a closed loop through uplink inspection feedback and downlink parameter updates, enabling self-learning and dynamic evolution across multiple time scales.

[0207] (1) Short-cycle edge feedback

[0208] Edge nodes are deployed along rural roads to continuously collect meteorological monitoring and inspection feedback. When a prediction deviation is detected to exceed a threshold, a lightweight incremental learning process is triggered, and the weights are adjusted according to equation (7). Perform a partial update. (The rest of the text is missing.) These are the actual measured values ​​from on-site inspections.

[0209] (7)

[0210] in, The road node risk prediction results are output from the ECCIS module in Part 4. The deviation between the measured values ​​obtained from inspections or meteorological sensors forms the basis for local correction. Let the prediction error loss function be... This represents the gradient of the loss function with respect to the model parameters. The learning rate is used to control the step size of each incremental update.

[0211] This process can be completed in real time under low computing power conditions, ensuring that local forecasts are sensitive to sudden weather events.

[0212] (2) Long-cycle central feedback

[0213] The central node aggregates the biased summaries uploaded by multiple edge nodes, combines them with historical meteorological and traffic data, and searches for hyperparameter combinations by solving the optimization problem defined in equation (8) through Bayesian optimization. This data is periodically distributed to edge devices, thereby enabling unified optimization and long-term adaptation of cross-regional models.

[0214] (8)

[0215] 5.2 Model Self-Learning and Parameter Evolution Mechanism

[0216] To enable cross-cycle evolution, the system maintains a global parameter cache at the central node. It records the optimal model state under different time windows. Within a short period, edge nodes update local parameters and upload summaries. Over a long period, the central node utilizes... Aggregate parameter gradients, perform global retraining and knowledge distillation, and update the main model. .

[0217] This "local increment-global distillation" mechanism ensures the continuous portability of the model across different climatic seasons and geographical regions, avoiding the performance drift problem of traditional static models.

[0218] 5.3 Risk Mapping and Strategy Generation Module

[0219] PI-HFV not only focuses on the self-learning of model performance but also emphasizes the interpretability and executability of prediction results. The system will predict the output... Transformed into a geospatial risk layer defined by equation (9) And automatically generate the management strategy defined by equation (10). .in For space mapping functions, It is a rule- and case-based strategy generator.

[0220] (9)

[0221] (10)

[0222] At the application level, the blue area represents the risk of water accumulation, the red area represents the risk of icing, and the gray area represents the risk of low visibility. At the same time, the system outputs differentiated maintenance strategies such as de-icing, drainage, and traffic restrictions based on the "Technical Specifications for Rural Road Maintenance" and historical case database. The feedback data after the strategy is implemented then enters the short-cycle correction layer to complete the closed loop of decision-making, feedback and optimization.

[0223] 5.4 Closed-loop optimization and system evolution effects

[0224] PI-HFV forms a complete closed loop through the synergy between upper and lower layers:

[0225] (1) Edge nodes perform prediction and local correction;

[0226] (2) Periodic optimization and parameter distribution of the central node;

[0227] (3) The strategy generation module translates the prediction results into actionable measures;

[0228] (4) The execution feedback drives the model to learn iteratively again.

[0229] This mechanism enables the cyclical co-evolution of the model, data, and strategy, significantly improving the system's resilience and long-term maintainability under extreme weather events.

[0230] The PI-HFV mechanism, building upon the ECCIS framework in Part 4, introduces time stratification and policy feedback dimensions, enabling the weather-driven rural road risk prediction system to move from "static reasoning" to "dynamic evolution." Through the combined action of short-cycle edge parameter correction, long-cycle central hyperparameter adjustment, and risk visualization and policy generation modules, the system achieves a self-evolving closed loop of "prediction-decision-execution-relearning," providing sustainable optimization technical support for intelligent road maintenance and emergency management.

[0231] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0232] This embodiment also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0233] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0234] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0235] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0236] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0237] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A weather-driven method for predicting rural road risks, characterized in that, Includes the following steps: Step S1: Based on the input road network information and meteorological grid information, construct a road-meteorological dual-map structure, including a road map, a meteorological map, and their cross-domain correlation matrix. Based on the cross-domain attention propagation mechanism, perform weighted perception on the dynamic impact of meteorological grid points to fuse road and meteorological information and generate a road-meteorological fusion feature vector. Step S2: Based on the generated road-meteorological fusion feature vector, the road physical attribute vector is introduced. By fusing the physical prior feature modulation, feature interaction fusion and nonlinear modulation are performed to obtain the feature vector after physical constraint modulation. Step S3: Input the modulated feature vector into the edge-center collaborative reasoning framework for risk prediction, and achieve joint prediction of local fast response and global fine optimization through collaborative computation between edge nodes and center nodes; wherein, edge nodes perform lightweight local reasoning, center nodes perform multi-source data fusion and global optimization, and maintain parameter consistency through a collaborative update mechanism, forming a two-layer reasoning output from local prediction to global optimization. Step S4: Based on the risk prediction results, update the model parameters through a prediction-inspection hierarchical feedback mechanism, and generate a visualized risk map and management strategy.

2. The weather-driven rural road risk prediction method according to claim 1, characterized in that, Step S1 specifically includes: Construct a road map based on the input road network information. Road map Based on road topology, adjacent road nodes are connected by edges to describe the topological structure and traffic elements of road nodes; where the node set... In this context, each node represents a road node, i.e., a road segment; the edge set... An edge in the middle represents the connection relationship between adjacent road nodes; Meteorological maps are constructed based on the input meteorological grid information. Weather map Based on meteorological nodes, adjacent meteorological nodes are connected by edges to reflect the spatial proximity and dynamic wind field evolution between meteorological grid points; among them, the node set In this context, a node represents a meteorological grid point, and the set of edges... An edge in the middle represents the spatial connection between adjacent meteorological grid points; To establish a cross-domain mapping between road nodes in the road map and meteorological grid points in the meteorological map, a cross-domain correlation matrix is ​​constructed. ; Middle elements , Representing meteorological grid points Located at road node Within their respective spheres of influence, the two entities have a cross-domain adjacency relationship. This indicates that the two are not directly related; Only the cross-domain connection structure is depicted, without carrying propagation strength information; the propagation strength of cross-domain information is determined by the subsequent physical propagation weights. Attention coefficient Dynamically determined; Combined with road map Weather map and cross-domain association matrix It forms a road-meteorological dual-map structure; Design road map Middle Road Node Characteristic Weather map China Meteorological Grid Characteristic The attention coefficient is calculated according to formula (1). Used to characterize meteorological grid points For road nodes Semantic relevance: (1) in, It is a linear transformation matrix. For learnable attention vectors, for Defined cross-domain neighborhood, Road nodes Meteorological grids in cross-domain neighbors Features This represents the natural exponential function. represents the linear rectified activation function with leakage, and || represents the concatenation operation; Integrating physical propagation weights into attention mechanisms The attention coefficient after physical enhancement is obtained. ; Then, the meteorological enhancement characteristics of the road nodes are calculated according to formula (2): (2) in, Indicates the activation function; This is a weighted fusion result of meteorological elements and dynamic wind field at road nodes; the obtained... This is the road-meteorological fusion feature vector.

3. The meteorological-driven rural road risk prediction method according to claim 2, characterized in that, The physical propagation weight The calculation method is as follows: Spatial proximity weights are defined to characterize spatial proximity constraints, reflecting the geographical attenuation characteristics of the influence of meteorological grid points on road nodes. The calculation formula is shown in equation (3): (3) in, Representing meteorological grid points With road nodes Euclidean distance, To prevent division by zero errors, This is the distance attenuation factor; The directional consistency weight is defined to characterize the directional consistency constraint, and its calculation formula is shown in equation (4): (4) in, The direction angle of the line connecting the road node and the meteorological grid point. For local wind direction, This is an adjustment coefficient used to control the sensitivity to wind direction differences; Then, calculate the physical propagation weights according to formula (5): (5) in, This is the balance coefficient.

4. The weather-driven rural road risk prediction method according to claim 1, characterized in that, In step S2, the implementation method of the feature modulation mechanism that integrates physical priors is as follows: The generated road-meteorological fusion feature vector With the introduced road physical attribute vector Each inputs a separate feature extraction branch; The feature interaction module dynamically fuses the features extracted from the two branches, employing an attention weighting mechanism to capture the coupling relationship between meteorological elements and road attributes. Then, a nonlinear mapping function is used to map the fused features, generating a feature vector modulated by physical constraints. ; Its expression is shown in equation (6): (6) in, Indicates feature-level interactive operations. It is a nonlinear mapping function. , , These are the weighting coefficients.

5. The weather-driven rural road risk prediction method according to claim 1, characterized in that, In step S3, the edge-center collaborative reasoning framework includes edge nodes and center nodes; The edge nodes are deployed with lightweight temporal convolutional networks for local risk prediction; the lightweight temporal convolutional networks undergo model pruning and parameter quantization; the model pruning compresses the network structure by removing neurons and connections with low prediction contributions; the parameter quantization reduces model storage and transmission costs by converting floating-point parameters into low-ratio point-to-point representations. The central node performs the following tasks: Multi-source data fusion: It integrates numerical weather forecasts, radar observation data, and prediction results from multiple edge nodes, and bridges the differences in spatiotemporal resolution between different data sources through feature alignment and spatiotemporal interpolation. Global model optimization: A deep fusion model is used to jointly model multi-dimensional meteorological elements, and error compensation is achieved through statistical regression or residual correction; Parameter synchronization and knowledge transfer: The optimized parameter set is compressed and distributed to each edge node to realize model weight update and knowledge transfer; The collaborative update mechanism achieves the following tasks: Adaptive task allocation: When network latency is high, switch to edge-independent operation mode to ensure the real-time performance of local predictions; when network conditions are good, switch to central collaborative optimization mode to achieve global synchronization and joint training. Results Upload and Parameter Feedback: Edge nodes periodically upload summarized prediction results, while central nodes send parameter update packages to complete bidirectional model synchronization; Drift detection and update triggering: A local drift detection module is embedded in the edge node to monitor prediction deviations and actively trigger parameter synchronization requests.

6. The weather-driven rural road risk prediction method according to claim 1, characterized in that, In step S4, the prediction-inspection hierarchical feedback mechanism includes: 1) Construct a multi-timescale hierarchical feedback mechanism: Construct a two-layer feedback structure with short and long cycles in the time dimension; edge nodes collect meteorological and inspection data in real time and perform incremental fine-tuning to achieve rapid self-correction; the central node performs hyperparameter optimization and model retraining based on historical and multi-source feedback to achieve continuous generalization of seasonality and regionality; a closed loop is formed through the uplink inspection feedback of edge nodes and the downlink parameter update of central nodes to achieve self-learning and dynamic evolution at multiple time scales; 2) Construct a model self-learning and parameter evolution mechanism: Maintain a global parameter cache at the central node. Record the optimal model state under different time windows; within a short period, edge nodes update local parameters and upload summaries. Over a long period, the central node utilizes... Aggregate parameter gradients, perform global retraining and knowledge distillation, and update the main model. ; 3) Construct a risk mapping and strategy generation module: This module generates the output risk prediction results. Transform into a geographic risk layer And automatically generate management policies. : in, For space mapping functions, A rule- and case-based strategy generator; 4) Achieving a complete closed loop of closed-loop optimization and system evolution through collaboration between upper and lower layers, including: Edge nodes perform prediction and local correction; Periodic optimization and parameter distribution of the central node; The strategy generation module translates the prediction results into actionable measures; The feedback from the execution further drives the model's iterative learning.

7. The weather-driven rural road risk prediction method according to claim 6, characterized in that, The dual-layer feedback structure includes: Short-cycle edge feedback: Edge nodes trigger incremental updates of local parameters based on the deviation between real-time inspection data and prediction results; Long-term central feedback: The central node performs hyperparameter optimization and global model retraining based on the bias summary and historical data uploaded by multiple edge nodes.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images