Road icing space-time prediction method based on graph neural network and selective state space model
By combining graph neural networks and selective state-space models, a multi-meteorological feature map and spatiotemporal prediction model were constructed, which solved the problems of instability and data scarcity in road icing prediction in the western Sichuan plateau region. This enabled high-precision and rapid icing risk prediction, supporting real-time decision-making in traffic management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SOUTHWEST JIAOTONG UNIV
- Filing Date
- 2025-12-26
- Publication Date
- 2026-05-08
AI Technical Summary
Existing methods for predicting road icing suffer from problems such as instability, weak generalization ability, and high latency in high-altitude and complex terrain areas. In particular, in the western Sichuan plateau region, the low frequency and uneven spatiotemporal distribution of icing events lead to scarce data samples, making it difficult to obtain accurate labels. Traditional methods are ill-suited to the complex and ever-changing meteorological environment.
By employing a method based on graph neural networks and selective state space models, a multi-meteorological feature map is constructed, spatial correlation features are extracted using graph convolution, and time dimension modeling is performed using selective state space models, thereby achieving rapid and high-precision prediction of road icing status in the next time step.
It improves the accuracy and adaptability of prediction under complex plateau climate conditions, can quickly and in real time capture changes in road surface temperature, provide high-precision prediction of icing risk, support traffic management departments to take timely preventive measures, and improve the effectiveness and operability of early warning.
Smart Images
Figure CN121997251A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of icing prediction technology for plateau highways, and in particular to a spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models. Background Technology
[0002] The Sichuan-Western Plateau is characterized by its high altitude, complex terrain, large diurnal temperature range, drastic temperature fluctuations, and extremely variable climate. Winters are marked by frequent alternating rain and snow and sudden temperature drops, making road icing a particularly prominent issue. Due to the unique climate and topography of the plateau region, icing not only affects vehicle safety but can also lead to serious consequences such as road closures and traffic disruptions, posing a significant threat to the safe operation and emergency management of highways. Currently, most road icing prediction methods rely on existing icing monitoring records as model training samples. However, in practical applications, the low frequency and uneven spatiotemporal distribution of icing events result in scarce icing data samples and difficulty in obtaining accurate annotations, leading to insufficient model generalization performance and decreased prediction accuracy. Furthermore, traditional methods are mostly based on static thresholds or linear regression models, which fail to fully exploit the dynamic correlation characteristics of meteorological data in both time series and spatial distribution, resulting in poor adaptability to the complex and variable meteorological environment of the Sichuan-Western Plateau. Therefore, researching a road icing prediction method that can overcome the difficulties in icing data annotation and improve the model's adaptability and prediction accuracy under complex plateau climate conditions is of significant practical importance. Summary of the Invention
[0003] This invention discloses a spatiotemporal prediction method for road icing based on graph neural networks and a selective state-space model. It aims to address the problems of instability, weak generalization ability, and high latency in existing road icing prediction technologies, particularly in high-altitude complex terrain, multi-meteorological element coupling, and rapid temporal response. This invention achieves rapid and high-precision prediction of road icing conditions in the next time step by constructing feature maps of multiple meteorological elements, introducing graph neural networks for spatial coupling feature extraction, and combining a new generation of selective state-space deep models for temporal dimension modeling.
[0004] To achieve the above objectives, the present invention provides the following technical solution: a spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models, comprising the following steps: S1. Collect multi-source meteorological parameter data and road network structure data along the highway, and perform preprocessing; S2. Set thresholds for various meteorological parameters in the collected meteorological parameter data, and label the meteorological parameter data in combination with preset logical judgment rules to generate an icing label sequence, which is used as a supervised learning target for subsequent model training. S3. Construct a feature space graph structure, map the preprocessed meteorological features to graph nodes, use graph convolution in graph neural network to extract the spatial correlation of meteorological features, and output spatial coding features that represent the dependency relationship between multiple meteorological points. S4. Input the spatial coding features into a selective state-space model for time series modeling, utilize the dynamic evolution capability of the state-space model to extract the changes of meteorological parameters in the time dimension, capture the long-term dependence of meteorological time series, and generate a time series feature vector. S5. The spatial coding features in step S3 and the temporal feature vector in step S4 are nonlinearly fused, and then combined with a prediction output layer of a fully connected neural network to construct a road icing prediction model. In the training phase, the icing labels generated in step S2 are used for supervised learning to train and optimize the model. In the prediction phase, the model is inferred based on the current meteorological data to obtain the icing probability, and finally the icing risk prediction of the target road in the next time step is completed.
[0005] Preferably, in step S1, the meteorological parameter data includes air temperature, dew point temperature, air humidity, precipitation, wind speed, air pressure, dew point temperature, saturated water vapor pressure, latitude and longitude, and road surface temperature; the road network structure data specifically includes the network structure between the geographical locations of the monitoring stations; the preprocessing includes data cleaning, outlier handling, missing value handling, and data normalization.
[0006] Preferably, in step S2, threshold values for various meteorological parameters are set for the collected meteorological parameter data, and the meteorological parameter data is labeled according to preset logical judgment rules, specifically including: For precipitation scenarios, the freezing condition is: when there is effective precipitation, i.e., the precipitation exceeds the preset precipitation threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, freezing is directly determined. For scenarios with no precipitation and high humidity, the icing condition is: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the condensation condition is met at the same time, that is, when the air humidity is greater than or equal to the preset threshold for high humidity, it is determined to be icing. For potential icing scenarios with low temperature and high humidity, the icing conditions are as follows: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the air humidity is greater than or equal to the humidity threshold, but the road surface temperature is greater than the dew point temperature, it is determined to be icing. If none of the icing conditions are met in any of the scenarios, then it is determined that no ice will form.
[0007] Preferably, the road icing prediction model employs a joint modeling approach using a graph neural network (GNN) and a selective state-space model (Mamba), specifically including: Construct a feature map with meteorological features as nodes, and generate an adjacency matrix of the feature map based on the statistical correlation between meteorological factors; High-dimensional correlation features between multi-source meteorological parameters are extracted using graph convolutional layers and graph attention layers. The graph structure output is used as the input to the selective state-space model to achieve joint learning of spatial features and temporal state-space representation; By using the gating structure of a selective state-space model, important meteorological conditions can be selectively memorized and updated.
[0008] Preferably, the feature map construction method includes the following: Each meteorological parameter is treated as a node to construct a "feature node graph"; the correlation strength between parameters is calculated based on the numerical change relationship of each meteorological parameter in the historical time series, and this is used as the basis for edge connection between nodes. A fixed, learnable feature map structure is constructed to reflect the interaction mechanism between meteorological elements. A graph neural network model is used to perform convolution and propagation calculations on the feature node map to extract high-dimensional interaction features between meteorological elements, providing a more physically meaningful representation for subsequent time series modeling.
[0009] Preferably, the prediction output layer adopts a fully connected neural network structure to receive the fused feature vector generated by the nonlinear fusion of spatial coding features and temporal feature vectors, and converts it into the icing probability of the next time step, specifically including the following: The number of neurons in the input layer is the same as the dimension of the fused feature vector. It is used to receive the fused feature vector generated by the nonlinear fusion of spatial encoded features and temporal feature vectors. The feature vector contains information in the spatial dimension and the evolution law and local change trend of historical meteorological parameters in the temporal dimension. The hidden layer is a multi-layer fully connected structure, used to further learn the nonlinear mapping relationship between road surface temperature and meteorological characteristics; The hidden layers employ the ReLU activation function to enhance the network's non-linear expressive power; The output layer has 1 neuron and is used to output the predicted probability of road icing in the next time step.
[0010] Preferably, the road icing prediction model, during the training phase, specifically includes: Training samples are generated using a time-sliding window with a window length of L and a prediction step size of 1 or more steps. The cross-entropy loss function is used as the training objective. The Adam optimizer is used for model optimization. Dropou and EarlyStopping are added during training to suppress overfitting and improve generalization performance. Gradient clipping is added when updating model weights to improve training stability.
[0011] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: a spatiotemporal prediction system for road icing based on graph neural networks and selective state-space models, comprising the following modules: The data acquisition and processing module collects multi-source meteorological parameter data and road network structure data along the highway and performs preprocessing. The icing label generation module sets thresholds for various meteorological parameters in the collected meteorological parameter data and labels the meteorological parameter data in combination with preset logical judgment rules to generate an icing label sequence. The temporal feature extraction module constructs a feature space graph structure, maps preprocessed meteorological features to graph nodes, and uses graph convolution in graph neural networks to extract spatial correlations of meteorological features, outputting spatially encoded features that represent the dependencies between multiple meteorological points. The spatially encoded features are then input into a selective state-space model for temporal modeling. The dynamic evolution capability of the state-space model is used to extract changes in meteorological parameters over time, capture long-term dependencies in meteorological time series, and generate temporal feature vectors. The model prediction module uses a fully connected neural network as the output layer to build a complete prediction model, trains and predicts the icing probability of the next time step, and assesses the icing risk. It nonlinearly fuses spatial coding features and temporal feature vectors, and then combines them with a prediction output layer of a fully connected neural network to build a road icing prediction model. In the training phase, it uses generated icing labels for supervised learning to train and optimize the model. In the prediction phase, it performs model inference based on current meteorological data to obtain the icing probability, and finally completes the prediction of the icing risk of the target road in the next time step.
[0012] On the other hand, to achieve the above objectives, the present invention also provides the following technical solution: a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the aforementioned spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models.
[0013] The beneficial effects of this invention are: 1) Innovative combination of advanced sequence models and rigorous meteorological logic: This invention utilizes an efficient time-series forecasting model based on a selective state-space model and an innovative method of meteorological experience-based judgment. This solves the problem of traditional methods struggling to balance accuracy and efficiency when processing complex, high-frequency meteorological sequence data from the Sichuan-Western Plateau. The selective state-space deep network algorithm can efficiently capture road surface temperature changes at the next time step, providing high-precision dynamic input for subsequent judgments and laying a solid foundation for accurate prediction.
[0014] 2) Achieves accurate multi-factor moisture determination in complex environments: This invention introduces dew point temperature as a key criterion for determining moisture sources and designs scenario-based logics for high humidity due to condensation, extreme low temperatures, and potential risks of low temperature and high humidity. This solves the problem of identifying freezing phenomena caused by moisture sources such as frost and dew condensation in the western Sichuan plateau region, where direct precipitation data is scarce or monitoring points are sparse, significantly improving the robustness and accuracy of the prediction method.
[0015] 3) Balancing long-term dependency capture and computational efficiency: This invention employs a selective state-space model as the core prediction model, overcoming the bottlenecks of high computational complexity and slow convergence speed of traditional self-attention mechanisms or recurrent neural network models when processing long-term series data. The selective state-space model efficiently models the long-range dependencies of the sequence with linear time complexity, enabling the system to perform fast, real-time road surface temperature prediction, which is particularly suitable for the immediate early warning needs of highways.
[0016] 4) The prediction results are directly used for decision-making, making it highly practical: This invention directly outputs the icing / non-icing state of the road in the next time step, with clear and explicit conclusions. Through logical judgment and threshold mechanisms, the judgment logic is made more consistent with the actual road conditions, which can guide traffic management departments to take timely and targeted prevention and de-icing measures, improving the effectiveness and operability of the early warning, and providing strong support for the safe operation of the expressway in western Sichuan. Attached Figure Description
[0017] Figure 1 This is a schematic diagram of the steps in the spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models. Figure 2 This is a schematic diagram of the Mamba model structure; Figure 3 This is a schematic diagram of a road icing spatiotemporal prediction system module based on graph neural networks and selective state-space models. In the diagram, 110 is the data acquisition and processing module; 120 is the icing label generation module; 130 is the time series feature extraction module; and 140 is the model prediction module. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] This invention provides a technical solution: a spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models, such as... Figure 1 As shown, it includes the following steps: S1. Collect multi-source meteorological parameter data and road network structure data along the Sichuan-Western Plateau Expressway and perform preprocessing.
[0020] The meteorological parameter data includes air temperature, dew point temperature, air humidity, rainfall, wind speed, air pressure, latitude and longitude, and road surface temperature; the preprocessing includes data cleaning, outlier handling, missing value handling, and data normalization.
[0021] In this embodiment, geographical location information and multiple meteorological parameter data of the Dema section of the Sichuan-Western Plateau Expressway are collected. The collected data includes three elements: time element, geographical element, and meteorological element. For the time element, the time scale of the collected data is at the second level to ensure that rapid changes in meteorological conditions can be captured, providing a basis for subsequent predictions at the next time step. For the geographical element, the latitude and longitude information of the observation point or road section is collected to identify the specific geographical location of the observation point. For the meteorological element, various meteorological parameters are collected, mainly including: Air temperature: Record the ambient air temperature to help calculate the dew point temperature and assess the effects of low temperatures.
[0022] Air humidity: Record the relative humidity of the environment to help determine the conditions for water vapor condensation, i.e. the source of moisture.
[0023] Dew point temperature: The temperature at which air is cooled to saturation under conditions where the water vapor content and air pressure remain unchanged. It is used to help determine the conditions for water vapor condensation.
[0024] Precipitation refers to the amount of precipitation. It involves recording daily precipitation and analyzing its potential contribution to freezing, especially in low-temperature environments.
[0025] The wind speed is the average wind speed at a height of 10 meters. The daily average wind speed at a height of 10 meters is recorded, taking into account the influence of wind speed on ice formation and duration, as well as its effect on the dispersion of ice and snow.
[0026] The air pressure is the surface air pressure, which is recorded daily. Changes in air pressure can affect the probability of icing. Low-pressure systems are usually accompanied by more precipitation and lower temperatures.
[0027] By comprehensively collecting the above detailed meteorological parameters, we can more accurately capture and analyze various meteorological factors affecting icing. To ensure the consistency and accuracy of the data, the collected data undergoes the following preprocessing: data cleaning, outlier handling, missing value handling, and data normalization. This ensures the consistency and accuracy of the data and transforms it into a time series format suitable for processing by a selective state-space model.
[0028] The preprocessed data is constructed into time series data in chronological order to eliminate the influence of outliers and maintain data continuity. Then, latitude and longitude locations are used as spatial features and fused with time series features to establish a spatiotemporal correlation mapping relationship, thereby improving the prediction accuracy and generalization ability of the model in the complex meteorological environment of the plateau. Finally, a meteorological parameter dataset is obtained.
[0029] S2. Set thresholds for various meteorological parameters in the collected meteorological data, and label the meteorological parameter data according to preset logical judgment rules to generate an icing label sequence, which will be used as a supervised learning target for subsequent model training. Implement a series of core thresholds to clearly define the key meteorological conditions for icing, and use dew point temperature as the basis for judging the source of moisture. The judgment logic is evaluated according to the following scenarios.
[0030] For precipitation scenarios, the freezing condition is: when there is effective precipitation, i.e., the precipitation exceeds the preset precipitation threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, freezing is directly determined. For scenarios with no precipitation and high humidity, the icing condition is: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the condensation condition is met at the same time, that is, when the road surface temperature is less than or equal to the road surface temperature and the air humidity is greater than or equal to the preset threshold for high humidity, it is determined to be icing. For potential icing scenarios with low temperature and high humidity, the icing conditions are as follows: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the air humidity is greater than or equal to the humidity threshold, but the road surface temperature is greater than the dew point temperature, it is determined to be icing. If none of the icing conditions are met in any of the scenarios, then it is determined that no ice will form.
[0031] S3. Construct a feature space graph structure, map the preprocessed meteorological features to graph nodes, use graph convolution in graph neural networks to extract the spatial correlation of meteorological features, and output spatial coding features that represent the dependencies between multiple meteorological points.
[0032] To fully utilize the spatial correlation characteristics between monitoring points, this embodiment introduces a Graph Neural Network (GNN) as a spatial modeling module to achieve efficient interactive propagation of meteorological conditions between different monitoring stations along the road. This module can seamlessly interface with the temporal modeling module of this invention to achieve dynamic supplementation of spatially relevant information.
[0033] Various monitoring points along the highway are considered as nodes in the diagram. Each node contains one piece of meteorological data. To reflect the true spatial relationships between monitoring points, this invention does not simply use Euclidean distance, but rather adopts an adjacency definition method oriented towards road structure, including but not limited to: (1) Adjacency relationship based on road topology If two monitoring stations are directly connected on the road or located on the same continuous road segment, they are designated as adjacent nodes. Such edges can be viewed as "actual topological edges" of the road, reflecting the physical continuity of the road segment.
[0034] (2) Weighted edges based on geographical distance For geographical distance For the nearest monitoring stations, construct weighted edges:
[0035] Among them Distance attenuation coefficient. This weighting method captures the meteorological correlation between nearby monitoring points.
[0036] The training and prediction process of the graph neural network is as follows: (1) Obtain meteorological data from monitoring stations and construct node features. (2) Establish an adjacency matrix based on road topology, distance between monitoring stations, etc. (3) Perform multiple spatial feature propagation through GNN to obtain spatial representation.
[0037] S4. Input the spatial coding features into a selective state-space model for time series modeling, utilize the dynamic evolution capability of the state-space model to extract the changes of meteorological parameters in the time dimension, capture the long-term dependence of meteorological time series, and generate a time series feature vector.
[0038] The model constructed in this embodiment adopts a selective state-space model architecture, which is used to efficiently extract features and predict time series data from meteorological time series data. It includes a time series feature embedding layer, a selective state-space core layer (SSMBlocks), and a prediction output layer, which are connected in sequence. The output of the previous layer serves as the input of the next layer, which is used to realize the whole process modeling of meteorological data from time series feature extraction to temperature prediction. The temporal feature embedding layer is used to receive and process meteorological parameter data divided by a time sliding window. It performs feature encoding and sequence initialization on the input multidimensional meteorological time series (including temperature, humidity, rainfall, wind speed, air pressure and latitude and longitude location features, etc.), and converts the meteorological sequence into a high-dimensional feature vector representation to capture the nonlinear coupling relationship between meteorological elements, which serves as the input to the selective state space core layer. The Selective State Space Core Layer is the core module of the model. It models the time dependence of meteorological parameters based on the Selective State Space Structure. Through dynamic state updates and selective gating mechanisms, it efficiently captures long-term time-series dependence features and local meteorological change trends. This layer outputs a road surface state feature vector containing past and current meteorological evolution information. The formula for calculating the state-space model is as follows:
[0039]
[0040] In the formula, For input signal, In hidden state, For output signal; These are the state matrix, input matrix, output matrix, and jump connection matrix, respectively.
[0041] Discretization using zero-order hold (ZOH) is performed, and the calculation formula is as follows:
[0042]
[0043]
[0044] in For time step.
[0045] Mamba's core innovation lies in enabling the discretization of parameters. Depends on the input matrix This achieves a "selective" mechanism.
[0046] Furthermore, the selective state space core layer comprises L stacked Mamba blocks, such as Figure 2 As shown, any Mamba block includes: Selective State-Space Model Module: Used for efficient time series modeling, capturing long-distance dependencies of multi-source meteorological data in the time dimension; Linear mapping layer: used to map input features to a high-dimensional representation space to enhance feature representation capabilities and enable the model to learn more detailed and complex meteorological feature relationships; Selection mechanism: used to adjust the parameter weights of the state space model according to the dynamic changes of input features, thereby selectively focusing on key historical moment information or forgetting redundant information, so as to improve the model's temporal modeling efficiency and generalization performance. Core Mamba operation: Models the input time series through discretized selective state-space equations, establishes the dynamic transmission process of meteorological parameters changing over time, and realizes dependency capture and continuous state update across time steps; Convolutional layers: used to extract local contextual features within a time sliding window, capturing minute changes in meteorological data on short time scales, such as sudden increases in rainfall, dramatic changes in humidity, or rapid drops in temperature, to enhance the model's sensitivity to extreme climate change. Normalization layer and activation function: The normalization layer is used to balance the differences in feature distribution of different batches of meteorological data and improve the convergence speed of model training; the activation function is used to enhance the nonlinear expression ability of the network, so that the model can better fit the complex and changeable meteorological environment characteristics of the western Sichuan plateau.
[0047] S5. Spatial coding features and temporal feature vectors are nonlinearly fused, and then combined with a prediction output layer of a fully connected neural network to construct a road icing prediction model including GNN and Mamba models. During the training phase, the icing labels generated in step S2 are used for supervised learning to train and optimize the model. In the prediction phase, the model inference is performed based on the current meteorological data to obtain the icing probability, and finally the icing risk result of the target road in the next time step is determined. The prediction output layer is used to receive spatiotemporal feature information from the fused system and, through linear mapping and fully connected layers, outputs the predicted road icing probability value for the next time step. The number of neurons in the input layer is the same as the dimension of the road surface state feature vector. It is used to receive the temporal feature information output from the core layer of the selective state space. The feature vector contains the evolution law and local change trend of historical meteorological parameters in the time dimension. The hidden layers are multi-layer fully connected structures used to further learn the nonlinear mapping relationship between road surface temperature and meteorological characteristics; the hidden layers use the ReLU activation function to enhance the nonlinear expressive power of the network and prevent gradient vanishing, thereby improving the convergence stability and prediction accuracy of the model under extreme climatic conditions; The formula for calculating the ReLU activation function is as follows:
[0048] The output layer has 1 neuron and is used to output the predicted probability of road icing in the next time step. The prediction output layer is jointly optimized with GNN and Mamba through end-to-end training, enabling the model to capture both long-term time-series dependencies and short-term temperature fluctuation features, thus ensuring the continuity and robustness of the prediction results under the complex climate environment of the western Sichuan plateau.
[0049] To achieve accurate prediction of road icing conditions, this invention designs and employs a loss function system suitable for binary classification during the model training phase. This implementation will provide a detailed explanation of the composition of the loss function, its optimization objective, and the resulting technical effects. The output is processed by a softmax function to obtain the predicted probability distribution. The formula for the Softmax function is as follows:
[0050] For real labels The cross-entropy loss function is defined as follows:
[0051] in This refers to the batch sample size. This represents the model's predicted probability of the true class.
[0052] The labeled dataset was stratified and divided into training, validation and test sets in a 7:2:1 ratio. An early stopping strategy was used to prevent overfitting during training. The training set was used to update parameters, the validation set was used to tune hyperparameters, and the test set was used to evaluate performance. Mean squared error and absolute error were used for comprehensive evaluation.
[0053] Based on the same inventive concept as the above-described method embodiments, this application also provides a spatiotemporal prediction system for road icing based on graph neural networks and selective state-space models. This system can achieve the functions provided by the above-described method embodiments, such as... Figure 3 As shown, the system includes the following modules: The data acquisition and processing module 110 collects multi-source meteorological parameter data and road network structure data along the highway and performs preprocessing. The icing label generation module 120 sets threshold values for various meteorological parameters in the collected meteorological parameter data and labels the meteorological parameter data in combination with preset logical judgment rules to generate an icing label sequence. The temporal feature extraction module 130 constructs a feature space graph structure, maps the preprocessed meteorological features to graph nodes, uses graph convolution in the graph neural network to extract the spatial correlation of meteorological features, and outputs spatial coding features representing the dependency relationship between multiple meteorological points; inputs the spatial coding features into a selective state space model for temporal modeling, uses the dynamic evolution capability of the state space model to extract the changes of meteorological parameters in the time dimension, captures the long-term dependency relationship of meteorological time series, and generates a temporal feature vector; Model prediction module 140 uses a fully connected neural network as the output layer to build a complete prediction model, trains and predicts the icing probability of the next time step, and assesses the icing risk; it nonlinearly fuses spatial coding features and temporal feature vectors, and then combines them with a prediction output layer of a fully connected neural network to build a road icing prediction model; during the training phase, it uses generated icing labels for supervised learning to train and optimize the model; during the prediction phase, it performs model inference based on current meteorological data to obtain the icing probability, and finally completes the prediction of the icing risk of the target road in the next time step.
[0054] Based on the same inventive concept as the above method embodiments, this application also provides an electronic device, the device including: a processor; and a memory for storing one or more programs; When the one or more programs are executed by the processor, the processor performs the spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models.
[0055] The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models includes the following steps: Collect multi-source meteorological parameter data and road network structure data along the highway, and perform preprocessing; Thresholds for various meteorological parameters are set for the collected meteorological parameter data, and the meteorological parameter data is labeled in combination with preset logical judgment rules to generate an icing label sequence, which is used as a supervised learning target for subsequent model training. Construct a feature space graph structure, map the preprocessed meteorological features to graph nodes, use graph convolution in graph neural network to extract the spatial correlation of meteorological features, and output spatial coding features that represent the dependency relationship between multiple meteorological points. Spatial coding features are input into a selective state-space model for time series modeling. The dynamic evolution capability of the state-space model is used to extract the changes of meteorological parameters in the time dimension, capture the long-term dependence of meteorological time series, and generate time series feature vectors. A road icing prediction model is constructed by nonlinearly fusing spatial coding features and temporal feature vectors, and then combining them with a prediction output layer of a fully connected neural network. During the training phase, supervised learning is performed using generated icing labels to train and optimize the model. During the prediction phase, the model inference is performed based on current meteorological data to obtain the icing probability, and finally the icing risk prediction of the target road in the next time step is completed.
[0056] Based on the same inventive concept as the above-described method embodiments, this application also provides a computer-readable storage medium having a computer program stored thereon, the computer program being implemented when executed by a processor. A spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models.
[0057] The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models includes the following steps: Collect multi-source meteorological parameter data and road network structure data along the highway, and perform preprocessing; Thresholds for various meteorological parameters are set for the collected meteorological parameter data, and the meteorological parameter data is labeled in combination with preset logical judgment rules to generate an icing label sequence, which is used as a supervised learning target for subsequent model training. Construct a feature space graph structure, map the preprocessed meteorological features to graph nodes, use graph convolution in graph neural network to extract the spatial correlation of meteorological features, and output spatial coding features that represent the dependency relationship between multiple meteorological points. Spatial coding features are input into a selective state-space model for time series modeling. The dynamic evolution capability of the state-space model is used to extract the changes of meteorological parameters in the time dimension, capture the long-term dependence of meteorological time series, and generate time series feature vectors. A road icing prediction model is constructed by nonlinearly fusing spatial coding features and temporal feature vectors, and then combining them with a prediction output layer of a fully connected neural network. During the training phase, supervised learning is performed using generated icing labels to train and optimize the model. During the prediction phase, the model inference is performed based on current meteorological data to obtain the icing probability, and finally the icing risk prediction of the target road in the next time step is completed.
[0058] This invention uses logical judgment and threshold mechanisms to make the judgment logic more closely match the actual road conditions, which can guide traffic management departments to take timely and targeted prevention and de-icing measures, improve the effectiveness and operability of early warning, and provide strong protection for the safe operation of the Sichuan-Western Plateau Expressway.
[0059] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0060] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0061] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0062] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.
[0063] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0064] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."
[0065] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models, characterized in that, Includes the following steps: S1. Collect multi-source meteorological parameter data and road network structure data along the highway, and perform preprocessing; S2. Set thresholds for various meteorological parameters in the collected meteorological parameter data, and label the meteorological parameter data in combination with preset logical judgment rules to generate an icing label sequence. S3. Construct a feature space graph structure, map the preprocessed meteorological features to graph nodes, use graph convolution in graph neural network to extract the spatial correlation of meteorological features, and output spatial coding features that represent the dependency relationship between multiple meteorological points. S4. Input the spatial coding features into a selective state-space model for time series modeling, utilize the dynamic evolution capability of the state-space model to extract the changes of meteorological parameters in the time dimension, capture the long-term dependence of meteorological time series, and generate a time series feature vector. S5. The spatial coding features in step S3 and the temporal feature vector in step S4 are nonlinearly fused, and then combined with a prediction output layer of a fully connected neural network to construct a road icing prediction model. In the training phase, the icing labels generated in step S2 are used for supervised learning to train and optimize the model. In the prediction phase, the model is inferred based on the current meteorological data to obtain the icing probability, and finally the icing risk prediction of the target road in the next time step is completed.
2. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 1, characterized in that: In step S1, the meteorological parameter data includes air temperature, dew point temperature, air humidity, precipitation, wind speed, air pressure, dew point temperature, saturated water vapor pressure, latitude and longitude, and road surface temperature; the road network structure data specifically includes the network structure between the geographical locations of the monitoring stations; the preprocessing includes data cleaning, outlier handling, missing value handling, and data normalization.
3. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 1, characterized in that, In step S2, threshold values for various meteorological parameters are set for the collected meteorological parameter data, and the meteorological parameter data is labeled according to preset logical judgment rules, specifically including: For precipitation scenarios, the freezing condition is: when there is effective precipitation, i.e., the precipitation exceeds the preset precipitation threshold, and the surface temperature is less than or equal to the preset surface temperature threshold, freezing is directly determined. For scenarios with no precipitation and high humidity, the icing condition is: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the condensation condition is met at the same time, that is, when the air humidity is greater than or equal to the preset threshold for high humidity, it is determined to be icing. For potential icing scenarios with low temperature and high humidity, the icing conditions are as follows: when the road surface temperature is less than or equal to the preset threshold for road surface temperature and the air humidity is greater than or equal to the humidity threshold, but the road surface temperature is greater than the dew point temperature, it is determined to be icing. If none of the icing conditions are met in any of the scenarios, then it is determined that no ice will form.
4. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 1, characterized in that, The road icing prediction model employs a joint modeling approach using a graph neural network (GNN) and a selective state-space model (Mamba), specifically including: Construct a feature map with meteorological features as nodes, and generate an adjacency matrix of the feature map based on the statistical correlation between meteorological factors; High-dimensional correlation features between multi-source meteorological parameters are extracted using graph convolutional layers and graph attention layers. The graph structure output is used as the input to the selective state-space model to achieve joint learning of spatial features and temporal state-space representation; By using the gating structure of a selective state-space model, important meteorological conditions can be selectively memorized and updated.
5. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 4, characterized in that, The feature map construction methods include the following: Each meteorological parameter is treated as a node to construct a "feature node graph"; the correlation strength between parameters is calculated based on the numerical change relationship of each meteorological parameter in the historical time series, and this is used as the basis for edge connection between nodes. A fixed, learnable feature map structure is constructed to reflect the interaction mechanism between meteorological elements. A graph neural network model is used to perform convolution and propagation calculations on the feature node map to extract high-dimensional interaction features between meteorological elements, providing a more physically meaningful representation for subsequent time series modeling.
6. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 1, characterized in that: The prediction output layer adopts a fully connected neural network structure to receive the fused feature vector generated by the nonlinear fusion of spatial coding features and temporal feature vectors, and converts it into the icing probability of the next time step, specifically including the following: The number of neurons in the input layer is the same as the dimension of the fused feature vector. It is used to receive the fused feature vector generated by the nonlinear fusion of spatial encoded features and temporal feature vectors. The feature vector contains information in the spatial dimension and the evolution law and local change trend of historical meteorological parameters in the temporal dimension. The hidden layer is a multi-layer fully connected structure, used to further learn the nonlinear mapping relationship between road surface temperature and meteorological characteristics; The hidden layers employ the ReLU activation function to enhance the network's non-linear expressive power; The output layer has 1 neuron and is used to output the predicted probability of road icing in the next time step.
7. The spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to claim 1, characterized in that, The road icing prediction model, during its training phase, specifically includes: Training samples are generated using a time-sliding window with a window length of L and a prediction step size of 1 or more steps. The cross-entropy loss function is used as the training objective. The Adam optimizer is used for model optimization. Dropou and EarlyStopping are added during training to suppress overfitting and improve generalization performance. Gradient clipping is added when updating model weights to improve training stability.
8. A prediction system for a spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models according to any one of claims 1-7, characterized in that: Includes the following modules: The data acquisition and processing module (110) collects multi-source meteorological parameter data and road network structure data along the expressway and performs preprocessing. The icing label generation module (120) sets threshold values for various meteorological parameters in the collected meteorological parameter data and labels the meteorological parameter data in combination with preset logical judgment rules to generate an icing label sequence. The temporal feature extraction module (130) constructs a feature space graph structure, maps the preprocessed meteorological features to graph nodes, uses graph convolution in the graph neural network to extract the spatial correlation of meteorological features, and outputs spatial coding features that represent the dependency relationship between multiple meteorological points; inputs the spatial coding features into the selective state space model for temporal modeling, uses the dynamic evolution capability of the state space model to extract the changes of meteorological parameters in the time dimension, captures the long-term dependency relationship of meteorological time series, and generates temporal feature vectors; The model prediction module (140) uses a fully connected neural network as the output layer to build a complete prediction model, trains and predicts the icing probability of the next time step, and assesses the icing risk; it nonlinearly fuses spatial coding features with temporal feature vectors, and then combines them with a prediction output layer of a fully connected neural network to build a road icing prediction model; during the training phase, it uses the generated icing labels for supervised learning to train and optimize the model. During the forecasting phase, based on current meteorological data, model reasoning is performed to obtain the probability of icing, and finally the icing risk prediction of the target road in the next time step is completed.
9. A computer-readable storage medium, characterized in that: It stores a computer program, which, when executed by a processor, implements the spatiotemporal prediction method for road icing based on graph neural networks and selective state-space models as described in any one of claims 1-7.