CN-121997251-A - Road icing space-time prediction method based on graphic neural network and selective state space model

CN121997251ACN 121997251 ACN121997251 ACN 121997251ACN-121997251-A

Abstract

The invention belongs to the technical field of highway icing prediction, and particularly discloses a road icing space-time prediction method based on a graphic neural network and a selective state space model, which aims at complex climatic characteristics of a Chuan-xi Gao Yuande Ma Gaosu road section, firstly, multi-source meteorological data along the line, including air temperature, humidity, rainfall, wind speed, road surface temperature and the like, are collected; the method comprises the steps of generating an icing label according to multiple meteorological parameter thresholds and logic judgment by using an innovative method based on meteorological experience criteria, solving the problem of data annotation when no direct icing record exists, introducing longitude and latitude information of an observation point into data as spatial features, modeling spatial correlation of meteorological features among different detection positions by using a graph neural network, and learning and predicting the meteorological features by using a selective state spatial model modeling meteorological feature. The method has the advantages of high prediction precision and strong capability of adapting to complex meteorological environment of the plateau, and can be used for monitoring and safety management of the icing risk of the expressway in the plateau area.

Inventors

ZHANG XIAOBO
LI QIANG
YANG HAN
GUO YUTIAN
WU SIRUI

Assignees

西南交通大学

Dates

Publication Date: 20260508
Application Date: 20251226

Claims (9)

1. The road icing space-time prediction method based on the graph neural network and the selective state space model is characterized by comprising the following steps of: s1, acquiring multisource meteorological parameter data and road network structure data along the expressway, and preprocessing; S2, setting various weather parameter thresholds for the collected weather parameter data, and marking the weather parameter data by combining with a preset logic judgment rule to generate an icing tag sequence; S3, constructing a characteristic space diagram structure, mapping the preprocessed meteorological features to diagram nodes, extracting spatial association of the meteorological features by using diagram convolution in a diagram neural network, and outputting spatial coding features representing dependency relations among multiple meteorological points; S4, inputting the space coding features into a selective state space model for time sequence modeling, extracting the change of meteorological parameters in the time dimension by utilizing the dynamic evolution capability of the state space model, capturing the long-term dependency of meteorological time sequence, and generating a time sequence feature vector; S5, carrying out nonlinear fusion on the spatial coding features in the step S3 and the time sequence feature vectors in the step S4, then combining a prediction output layer of a fully-connected neural network to construct a road icing prediction model, carrying out supervised learning by utilizing the icing labels generated in the step S2 in a training stage, carrying out training optimization on the model, carrying out model reasoning according to current meteorological data in the prediction stage to obtain icing probability, and finally completing icing risk prediction of a target road in the next time step.
2. The road icing space-time prediction method based on the graphic neural network and the selective state space model according to claim 1, wherein in the step S1, the meteorological parameter data comprise air temperature, dew point temperature, air humidity, precipitation, wind speed, air pressure, dew point temperature, saturated water vapor pressure, longitude and latitude and road surface temperature, the road network structure data specifically comprise network structures between geographical positions of measuring stations, and the preprocessing comprises data cleaning, outlier processing, missing value processing and data normalization.
3. The road icing space-time prediction method based on the graph neural network and the selective state space model according to claim 1, wherein in step S2, each weather parameter threshold is set for the collected weather parameter data, and the weather parameter data is labeled in combination with a preset logic judgment rule, specifically comprising: For a precipitation scene, the icing condition is that when effective precipitation exists, namely precipitation exceeds a precipitation preset threshold, and the surface temperature is less than or equal to the surface temperature preset threshold, the icing condition is directly judged to be icing; For a scene without precipitation and high humidity, the icing condition is that when the road surface temperature is less than or equal to a preset threshold value of the road surface temperature and simultaneously meets the condensing condition, namely the air humidity is greater than or equal to a preset threshold value of high humidity, the icing condition is judged to be icing; for a low-temperature high-humidity potential icing scene, the icing condition is that when the road surface temperature is less than or equal to a preset threshold value of the road surface temperature and the air humidity is greater than or equal to a humidity threshold value, the road surface temperature is greater than the dew point temperature, the road surface temperature is determined to be icing; And if the icing conditions of all the scenes are not satisfied, judging that the scenes are not frozen.
4. The road icing space-time prediction method based on the graph neural network and the selective state space model according to claim 1, wherein the road icing prediction model adopts the graph neural network GNN and the selective state space model Mamba to perform joint modeling, and specifically comprises the following steps: constructing a feature map taking meteorological features as nodes, and generating an adjacency matrix of the feature map according to statistical correlation among meteorological factors; extracting high-dimensional association features among multi-source meteorological parameters through a graph convolution layer and a graph annotation force layer; The output of the graph structure is used as the input of the selective state space model to realize the joint learning of the space characteristics and the time state space representation; and selectively memorizing and updating important meteorological states through a gating structure of the selective state space model.
5. The road icing space-time prediction method based on the graph neural network and the selective state space model according to claim 4, wherein the feature graph construction mode comprises the following steps: Calculating the correlation strength between the parameters according to the numerical value change relation of each meteorological parameter in the historical time sequence, and taking the correlation strength as the edge connection basis between the nodes; And rolling and spreading calculation is carried out on the characteristic node diagram by using a graph neural network model, thereby extracting high-dimensional interaction characteristics among the meteorological elements and providing a representation with more physical significance for subsequent time sequence modeling.
6. The road icing space-time prediction method based on the graphic neural network and the selective state space model according to claim 1, wherein the prediction output layer adopts a fully connected neural network structure and is used for receiving a fusion feature vector generated by nonlinear fusion of a space coding feature and a time sequence feature vector and converting the fusion feature vector into an icing probability of the next time step, and the method specifically comprises the following steps: The number of neurons of the input layer is the same as the dimension of the fused feature vector, and the neurons are used for receiving the fused feature vector generated by nonlinear fusion of the space coding feature and the time sequence feature vector, wherein the feature vector comprises information in the space dimension, and the evolution rule and the local variation trend of the historical meteorological parameters in the time dimension; the hidden layer is of a multi-layer full-connection structure and is used for further learning a nonlinear mapping relation between the road surface temperature and meteorological features; the hidden layer adopts a ReLU activation function to enhance the nonlinear expression capability of the network; The neuron number of the output layer is 1, and the neuron number is used for outputting the road icing probability predicted value of the next time step.
7. The road icing space-time prediction method based on the graph neural network and the selective state space model according to claim 1, wherein the road icing prediction model specifically comprises: Generating training samples by adopting a time sliding window, wherein the window length is L, the prediction step length is 1 or more steps, using a cross entropy loss function as a training target, adopting an Adam optimizer to perform model optimization, adding Dropou and Early Stopping in the training process to inhibit overfitting and improve generalization performance, and adding gradient clipping in the process of model weight updating to improve training stability.
8. A prediction system of a road icing spatiotemporal prediction method based on a graph neural network and a selective state space model according to any of claims 1-7, characterized by comprising the following modules: the data acquisition and processing module (110) acquires multisource meteorological parameter data and road network structure data along the expressway and performs preprocessing; The icing tag generation module (120) is used for setting various weather parameter thresholds for the acquired weather parameter data, marking the weather parameter data by combining with a preset logic judgment rule, and generating an icing tag sequence; The time sequence feature extraction module (130) is used for constructing a feature space diagram structure, mapping the preprocessed meteorological features to diagram nodes, extracting the spatial correlation of the meteorological features by using diagram convolution in a diagram neural network, and outputting spatial coding features representing the dependency relationship among multiple meteorological points; The model prediction module (140) uses the fully-connected neural network as an output layer to construct a complete prediction model, trains and predicts the icing probability of the next time step, evaluates the icing risk, carries out nonlinear fusion on the spatial coding features and the time sequence feature vectors, then combines the prediction output layer of the fully-connected neural network to construct a road icing prediction model, carries out supervised learning by using the generated icing labels in the training stage, carries out training optimization on the model, carries out model reasoning according to the current meteorological data in the prediction stage, and finally completes the icing risk prediction of the target road in the next time step.
9. A computer-readable storage medium, characterized in that a computer program is stored thereon, which computer program, when being executed by a processor, implements the road icing spatiotemporal prediction method based on a graph neural network and a selective state space model according to any of claims 1-7.

Description

Road icing space-time prediction method based on graphic neural network and selective state space model Technical Field The invention relates to the technical field of highland highway icing prediction, in particular to a road icing space-time prediction method based on a graphic neural network and a selective state space model. Background The Sichuan plateau has high topography, complex topography, large day and night temperature difference, severe temperature change, abnormal and changeable climate conditions, and the phenomena of rain and snow alternation and air temperature abrupt drop frequently occur in winter, so that the road icing problem is particularly prominent. Due to special climates and topography factors in the plateau region, the icing not only affects the running safety of the vehicle, but also can cause serious consequences such as road closure, traffic interruption and the like, and the safety operation and emergency management of the expressway are obviously threatened. At present, most road icing prediction methods rely on the existing icing monitoring records as model training samples, but in practical application, icing events are low in occurrence frequency and uneven in space-time distribution, so that icing data samples are sparse and are difficult to obtain accurate labels, and further the model generalization performance is insufficient and the prediction accuracy is reduced. In addition, the traditional method is mostly based on a static threshold value or a linear regression model, so that dynamic association characteristics of meteorological data on time sequence and space distribution are difficult to fully mine, and the adaptability to complex and changeable meteorological environments in the Sichuan plateau region is poor. Therefore, the road icing prediction method capable of overcoming the difficulty in marking icing data and improving the adaptability and prediction accuracy of the model under the complex plateau climate condition has important practical significance. Disclosure of Invention The invention discloses a road icing space-time prediction method based on a graph neural network and a selective state space model, and aims to solve the problems of unstable prediction, weak generalization capability, high time delay and the like in the aspects of high-altitude complex terrain, multi-meteorological element coupling and quick time sequence response of the conventional road icing prediction technology. According to the method, a multi-meteorological-element feature map is constructed, a map neural network is introduced to extract space coupling features, and a new generation selective state space depth model is combined to conduct time dimension modeling, so that quick and high-precision prediction of the icing state of the road in the next time step is achieved. In order to achieve the purpose, the invention provides the following technical scheme that the road icing space-time prediction method based on the graph neural network and the selective state space model comprises the following steps: s1, acquiring multisource meteorological parameter data and road network structure data along the expressway, and preprocessing; S2, setting various weather parameter thresholds for the collected weather parameter data, and marking the weather parameter data by combining with a preset logic judgment rule to generate an icing tag sequence which is used as a supervision learning target for subsequent model training; S3, constructing a characteristic space diagram structure, mapping the preprocessed meteorological features to diagram nodes, extracting spatial association of the meteorological features by using diagram convolution in a diagram neural network, and outputting spatial coding features representing dependency relations among multiple meteorological points; S4, inputting the space coding features into a selective state space model for time sequence modeling, extracting the change of meteorological parameters in the time dimension by utilizing the dynamic evolution capability of the state space model, capturing the long-term dependency of meteorological time sequence, and generating a time sequence feature vector; S5, carrying out nonlinear fusion on the spatial coding features in the step S3 and the time sequence feature vectors in the step S4, then combining a prediction output layer of a fully-connected neural network to construct a road icing prediction model, carrying out supervised learning by utilizing the icing labels generated in the step S2 in a training stage, carrying out training optimization on the model, carrying out model reasoning according to current meteorological data in the prediction stage to obtain icing probability, and finally completing icing risk prediction of a target road in the next time step. Preferably, in step S1, the meteorological parameter data includes air temperature, dew point temperature, air humidity, precipitation, wind speed,