Pavement temperature prediction method based on numerical model fusion and deep learning

Abstract

The present invention discloses a road surface temperature prediction method based on numerical model fusion deep learning, including making training sets and verification sets based on multi-factor forecast data and observation data; integrating and forming a unified-scale regional high-resolution multi-substrate feature library based on regional high-resolution multi-substrate data; constructing a deep learning road surface temperature prediction model based on an attention mechanism graph convolutional neural network, using a Gaussian weighted loss function to carry out model training; using the trained model, for the target time, based on the numerical weather forecast model multi-factor forecast data and regional high-resolution multi-substrate data, generating a road surface temperature prediction result based on numerical model fusion deep learning. The present invention effectively solves the sample imbalance problem of road surface temperature at different scales, enhances the model's prediction ability for road surface temperature extremes, and improves the accuracy of highway road surface temperature prediction based on numerical model fusion deep learning, and has extremely strong application value.

Description

Technical Field

[0001] The present invention relates to the field of pavement temperature prediction, and in particular to a pavement temperature prediction method based on numerical model fusion and deep learning. Background Art

[0002] In modern transportation systems, the safety and efficiency of highways are crucial. Pavement temperature is one of the important factors affecting the safe operation of highways. Especially in cold winter or extremely hot weather, changes in pavement temperature may cause icing or softening of asphalt, leading to traffic accidents. In order to prevent and deal with these problems, it is particularly important to accurately predict pavement temperature.

[0003] Traditional pavement temperature prediction methods mainly rely on physical models and empirical formulas, and are calculated based on meteorological factors (such as temperature, humidity, wind speed, etc.) and the thermodynamic properties of pavement materials. However, these models are difficult to fully and accurately reflect the changes in pavement temperature in complex actual environments, especially in dealing with nonlinear and multi-scale temporal and spatial dependencies. In addition, these methods usually require a large amount of field data and parameter adjustments, are complex to operate, and have limited application scope.

[0004] In recent years, with the development of artificial intelligence and big data technology, pavement temperature prediction methods based on deep learning have gradually attracted attention. Through large-scale data training, deep learning methods can automatically learn complex nonlinear relationships and feature expressions, significantly improving prediction accuracy. However, simply relying on deep learning models for pavement temperature prediction often ignores explicit knowledge and experience in physical models, which may lead to insufficient generalization of the model. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a pavement temperature prediction method based on numerical model fusion deep learning, which can significantly improve the accuracy and reliability of pavement temperature prediction.

[0006] Technical solution: The road surface temperature prediction method based on numerical model fusion deep learning of the present invention comprises the following steps:

[0007] The training set and validation set are produced based on the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station;

[0008] Based on the regional high-resolution multivariate underlying surface data, a unified scale regional high-resolution multivariate underlying surface feature library is formed;

[0009] A deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network is constructed, and the Gaussian weighted loss function is used to carry out model training; the model includes: input module, graph learning module, graph convolution module, local time convolution module, global time convolution module, attention module and output module. The input module is used to integrate the multi-factor forecast data based on the numerical weather forecast model, the observation data of the traffic meteorological automatic monitoring station and the regional high-resolution multivariate underlying surface feature library to obtain the input feature matrix. The graph learning module converts the input feature matrix into a graph structure to capture the temporal and spatial features between the input features. The graph convolution module is used to aggregate the input features of the target point with the graph structure. The features of the adjacency matrix of the neighboring points around the target point in the structure are extracted to extract the spatial dependency relationship between the nodes in the graph structure; the local time convolution module is used to convolve the local time series data in the feature matrix output by the graph convolution module, and extract the local spatiotemporal features of the input features; the global time convolution module is used to convolve the global time series data in the feature matrix output by the graph convolution module, and extract the overall spatiotemporal features of the input features; the attention module is used to capture the temporal and spatial dependency relationship between the local spatiotemporal features and the overall spatiotemporal features; the output module is used to map the feature matrix output by the attention module to the target variable, and generate the final road surface temperature prediction result based on the numerical model fusion deep learning;

[0010] Using the trained model, for the target time, based on the multi-factor forecast data of the numerical weather forecast model and the regional high-resolution multivariate underlying surface data, the road surface temperature forecast results based on the numerical model fusion deep learning are generated.

[0011] Furthermore, when preparing the training set and the validation set, it is necessary to remove the outliers in the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station, fill in the missing values, and then perform standardization respectively.

[0012] Furthermore, the method for filling missing values ​​is the central difference method.

[0013] Furthermore, the method for integrating and forming a unified-scale regional high-resolution multivariate underlying surface feature library is: normalizing the regional high-resolution multivariate underlying surface data separately to obtain multivariate unified-scale regional high-resolution underlying surface features, and then integrating them to obtain a unified-scale regional high-resolution multivariate underlying surface feature library.

[0014] Furthermore, a deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network is constructed, specifically:

[0015] Input module: Acquire and integrate standardized numerical weather forecast model multi-factor forecast data , standardized traffic weather automatic monitoring station observation data and a unified scale regional high-resolution multivariate underlying surface feature database , forming the input feature matrix ;

[0016] Graph learning module: The specific forms of the initial node feature matrix, adjacency matrix and degree matrix in the graph structure are:

[0017] ,

[0018] ,

[0019] ,

[0020] in, is the initial node feature matrix, and Respectively represent nodes and nodes , For Node and nodes The corresponding adjacency matrix The elements in For the node The number of connected edges, also known as the degree matrix The Line Column elements, finally forming an adjacency matrix Sum degree matrix ;

[0021] Graph convolution module: contains multiple graph convolution layers, multiple residual connection layers and multiple jump connection layers; performs convolution operations on graph structure data, enhances gradient flow through residual connections, and retains global information through jump connections; is used to aggregate the input features of the target point and the features of the adjacency matrix of the neighboring points around the target point in the graph structure, and extract the spatial dependency features between nodes; its specific form is:

[0022] ,

[0023] in, is the identity matrix, The adjacency matrix after adding the self-loop is made so that each node contains its own information during the graph convolution process; for The degree matrix represents the number of connections of each node; For the The weight matrix of the graph convolution layer learns the linear transformation of node features; is the activation function; and Respectively Layer and The output feature matrix of the layer graph convolution; after After layer iteration, the output feature matrix of the graph convolution module is obtained ;

[0024] Local temporal convolution module: contains a convolution layer and a pooling layer; the convolution layer outputs the feature matrix of the graph convolution module The local time series data in the convolution operation is performed, and the pooling layer downsamples the convolutional features to extract the local spatiotemporal features of the input features, that is, the short-term time-dependent features; its specific form is:

[0025] ,

[0026] ,

[0027] in, represents a one-dimensional convolution operation, represents the pooling operation, is the local temporal convolution kernel size, Output feature matrix for the convolutional layer, Output feature matrix for local temporal convolution module, i.e. local spatiotemporal features;

[0028] Global temporal convolution module: contains a convolution layer; outputs the feature matrix through the graph convolution module The global time series data in is convolved to extract the overall spatiotemporal features of the input features, that is, the long-term time-dependent features; its specific form is:

[0029] ,

[0030] in, is the global temporal convolution kernel size, Output feature matrix for the global temporal convolution module, i.e., overall spatiotemporal features;

[0031] Attention module: It includes an attention mechanism layer and a feedforward neural network layer. The attention mechanism layer calculates the correlation between local spatiotemporal features and overall spatiotemporal features, and performs weighted average of local spatiotemporal features and overall spatiotemporal features. The feedforward neural network layer further refines the processing features and introduces activation functions to enhance the model's expression ability and nonlinear feature extraction ability, capturing the temporal and spatial dependencies between local spatiotemporal features and overall spatiotemporal features. Its specific form is:

[0032] ,

[0033] ,

[0034] ,

[0035] in, Let the query matrix be , The key matrix is ​​set to , is the key matrix The transpose of For the The value matrix of the indexes; is the index variable for the query matrix, represents the index of the key matrix and the value matrix, is the index to sum over all key matrices; is the dimension of the key, used to scale the dot product attention score; is the attention weight, indicating the index Index The intensity of attention; is the weighted feature matrix, through the attention weight pair value matrix Take the weighted average to get; is the weight matrix of the feedforward neural network, is the bias term of the feedforward neural network; Output feature matrix for attention module;

[0036] Output module: contains a fully connected layer, which maps the output feature matrix of the attention module to the target variable, and is used to generate the final road surface temperature prediction result based on the numerical model fusion deep learning; its specific form is:

[0037] ,

[0038] in, is the road surface temperature prediction result, is the weight matrix of the fully connected layer, is the bias term of the fully connected layer.

[0039] Furthermore, the Gaussian weighted loss function is:

[0040] ,

[0041] ,

[0042] in, is the Gaussian weighted loss function, is the total number of samples used to calculate the loss function in model training, and They are respectively used to calculate the loss function in model training. The observed and predicted values ​​of samples; and are the mean and standard deviation of the observed samples used to calculate the loss function in model training; is the mean square error weight function.

[0043] The formula for iterative optimization of model parameters during model training is:

[0044] ,

[0045] in, and are the model parameters before and after iterative optimization, is the learning rate, is the Gaussian weighted loss function, Represents the gradient of the loss function with respect to the model parameters.

[0046] The road surface temperature prediction system based on numerical model fusion deep learning of the present invention comprises:

[0047] The data set production unit is used to produce training sets and validation sets based on the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station; and based on the regional high-resolution multivariate underlying surface data, integrate and form a unified-scale regional high-resolution multivariate underlying surface feature library;

[0048] The model building and training unit is used to build a deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network, and adopts a Gaussian weighted loss function to carry out model training; the model includes: input module, graph learning module, graph convolution module, local time convolution module, global time convolution module, attention module and output module. The input module is used to integrate the multi-factor forecast data based on the numerical weather forecast model, the observation data of the traffic meteorological automatic monitoring station and the regional high-resolution multivariate underlying surface feature library to obtain the input feature matrix. The graph learning module converts the input feature matrix into a graph structure to capture the temporal and spatial features between the input features. The graph convolution module is used to aggregate the input of the target point. The input features are combined with the features of the adjacency matrix of the neighboring points around the target point in the graph structure to extract the spatial dependency between the nodes in the graph structure; the local time convolution module is used to convolve the local time series data in the feature matrix output by the graph convolution module and extract the local spatiotemporal features of the input features; the global time convolution module is used to convolve the global time series data in the feature matrix output by the graph convolution module and extract the overall spatiotemporal features of the input features; the attention module is used to capture the temporal and spatial dependency between the local spatiotemporal features and the overall spatiotemporal features; the output module is used to map the feature matrix output by the attention module to the target variable to generate the final road surface temperature prediction result based on the numerical model fusion deep learning;

[0049] The forecasting unit is used to use the trained model to generate road surface temperature forecast results based on numerical model fusion deep learning for the target time based on the multi-factor forecast data of the numerical weather forecast model and the regional high-resolution multivariate underlying surface data.

[0050] An electronic device for storing and executing the method, comprising:

[0051] A memory storing executable program code;

[0052] a processor coupled to the memory;

[0053] The processor calls the executable program code stored in the memory to execute the steps of the road surface temperature prediction method based on numerical model fusion deep learning.

[0054] A computer-readable storage medium for storing and executing the method, wherein the computer-readable storage medium stores computer instructions, and when the computer instructions are called, they are used to execute the steps of the road surface temperature prediction method based on numerical model fusion deep learning.

[0055] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are: (1) by integrating the multi-factor forecast data of the numerical weather forecast model, the observation data of the traffic meteorological automatic monitoring station and the regional high-resolution multivariate underlying surface data, the advantages of different data sources are fully utilized, and the complex temporal and spatial dependencies between the road surface temperature and various forecast factors are effectively captured. Compared with the traditional methods, it has significant advantages in feature extraction and physical meaning, and improves the accuracy and reliability of the road surface temperature forecast; (2) the deep learning road surface temperature forecast model constructed based on the attention mechanism graph convolutional neural network can better handle high-dimensional and unstructured Multi-source data, among which the introduction of the graph convolution module effectively captures the spatial dependency of the road network structure and the atmospheric environment, while the introduction of the attention module further strengthens the model's focus on key features, improving the model's generalization ability and prediction effect; (3) Aiming at the Gaussian distribution characteristics of road surface temperature, the established Gaussian weighted loss function effectively solves the problem of uneven temperature samples of different scales in road surface temperature prediction, ensuring that the model can pay more attention to the prediction of temperature extremes during training, enhancing the model's ability to predict road surface temperature extremes, and significantly improving the prediction accuracy under extreme weather conditions, which has extremely strong practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 is a flow chart of the method of the present invention;

[0057] Figure 2 Schematic diagram of the deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network;

[0058] Figure 3 Schematic diagram of graph convolution module;

[0059] Figure 4 Optimize the diagram for model training. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0061] The present invention combines numerical weather forecast models, multivariate underlying surface features and deep learning models, fully utilizing the explicit knowledge of numerical weather forecast models and the powerful feature extraction capabilities of deep learning to construct a road surface temperature prediction model that integrates multi-source data and has a highly nonlinear expression capability, which helps to enhance the model's ability to predict extreme road surface temperature, thereby greatly improving the accuracy and reliability of road surface temperature forecasts. Figure 1 As shown, the road surface temperature prediction method based on numerical model fusion deep learning of the present invention specifically includes the following steps:

[0062] S1. Obtain the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station, and perform data quality control and standardization preprocessing; divide them into training set and validation set according to a certain ratio; the specific steps are as follows:

[0063] S11. Obtain multi-factor forecast data of numerical weather forecast model and observation data of traffic meteorological automatic monitoring station; the multi-factor forecast data of numerical weather forecast model includes temperature, air pressure, humidity, wind speed, wind direction and precipitation forecast data output by numerical weather forecast model, and the observation data of traffic meteorological automatic monitoring station includes road surface temperature, air pressure, humidity, wind speed, wind direction and precipitation observation data collected by meteorological automatic monitoring stations along highways.

[0064] S12, perform quality control and standardization preprocessing on the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station obtained in S11; the quality control first removes abnormal values ​​(such as abnormal outliers), and then fills the missing values ​​with the central difference method; the standardization preprocessing is used to eliminate the influence of different variable dimensions, and its calculation method is:

[0065] ,

[0066] in, is the total number of samples of the original numerical weather forecast model multi-factor forecast data or the traffic meteorological automatic monitoring station observation data, For the The original numerical weather forecast model multi-factor forecast data or traffic meteorological automatic monitoring station observation data of samples, It is standardized multi-factor forecast data of numerical weather forecast model or standardized observation data of traffic meteorological automatic monitoring station.

[0067] S13. Divide the standardized multi-factor forecast data of numerical weather forecast model and the standardized observation data of traffic meteorological automatic monitoring station into training set and validation set according to a certain ratio. The common division ratio is 80% training set and 20% validation set.

[0068] S2. Obtain regional high-resolution multivariate underlying surface data, including altitude, terrain relief, surface roughness, vegetation coverage, water coverage, and hard cover, and perform normalization preprocessing to integrate and form a regional high-resolution multivariate underlying surface feature library of a unified scale; the specific steps are as follows:

[0069] S21. Obtain regional high-resolution multivariate underlying surface data, including regional high-resolution altitude, terrain relief, surface roughness, vegetation coverage, water coverage and hard cover, all of which are two-dimensional grid matrices;

[0070] S22. Perform normalization preprocessing on the regional high-resolution multivariate underlying surface data to form regional high-resolution underlying surface features of uniform scale; the calculation method is:

[0071] ,

[0072] in, is the original underlying surface data, and are the maximum and minimum values ​​of the original underlying surface data, The unified-scale regional high-resolution underlying surface features are integrated to form a unified-scale regional high-resolution multiple underlying surface feature library.

[0073] S3. Construct a deep learning road surface temperature prediction model based on an attention mechanism graph convolutional neural network that integrates multi-factor forecast data of numerical weather forecast models, observation data from traffic meteorological automatic monitoring stations, and regional high-resolution multivariate underlying surface data, and use a Gaussian weighted loss function to train the model; wherein, the deep learning road surface temperature prediction model based on an attention mechanism graph convolutional neural network includes an input module, a graph learning module, a graph convolution module, a local time convolution module, a global time convolution module, an attention module, and an output module; the input module is used to obtain and integrate data in S1 and S2 to obtain an input feature matrix; the graph learning module converts the input feature matrix into a graph structure to capture the temporal and spatial features between input features; The convolution module is used to aggregate the input features of the target point and the features of the adjacency matrix of the neighboring points around the target point in the graph structure, and extract the spatial dependency between the nodes in the graph structure; the local time convolution module is used to convolve the local time series data in the output feature matrix of the graph convolution module, and extract the local spatiotemporal features of the input features; the global time convolution module is used to convolve the global time series data in the output feature matrix of the graph convolution module, and extract the overall spatiotemporal features of the input features; the attention module is used to capture the temporal and spatial dependencies between the local spatiotemporal features and the overall spatiotemporal features, and the output module is used to map the output feature matrix of the attention module to the target variable, and generate the final road surface temperature prediction result based on the numerical model fusion deep learning; Figure 2 As shown, the specific steps are as follows:

[0074] S31. Construct a deep learning road surface temperature prediction model based on an attention mechanism graph convolutional neural network that integrates multi-factor forecast data of numerical weather forecast models, observation data of traffic meteorological automatic monitoring stations, and regional high-resolution multivariate underlying surface data. The model includes an input module, a graph learning module, a graph convolution module, a local time convolution module, a global time convolution module, an attention module, and an output module. The specific form is:

[0075] Input module: Obtain and integrate standardized numerical weather prediction model multi-factor forecast data obtained from steps S1 and S2 , standardized traffic weather automatic monitoring station observation data and a unified scale regional high-resolution multivariate underlying surface feature database , forming the input feature matrix .

[0076] Graph learning module: converts the input feature matrix into a graph structure to capture the temporal and spatial features between the input features, and calculates the initial node feature matrix, adjacency matrix and degree matrix, so that the subsequent graph convolution operation can capture the spatial dependency of the data; its specific form is:

[0077] ,

[0078] ,

[0079] ,

[0080] in, is the initial node feature matrix, and Respectively represent nodes and nodes , For Node and nodes The corresponding adjacency matrix The elements in For the node The number of connected edges, also known as the degree matrix The Line Column elements, finally forming an adjacency matrix Sum degree matrix .

[0081] Graph convolution module: Figure 3 As shown in the figure, it contains multiple graph convolution layers, multiple residual connection layers and multiple jump connection layers; it performs convolution operations on the graph structure data (i.e., the initial node feature matrix, adjacency matrix and degree matrix obtained by the graph learning module), enhances the gradient flow through residual connections, and retains global information through jump connections; it is used to aggregate the input features of the target point and the features of the adjacency matrix of the neighboring points around the target point in the graph structure, and extract the spatial dependency features between nodes; its specific form is:

[0082] ,

[0083] in, is the identity matrix, To add the adjacency matrix after self-loop, each node contains its own complete information during the graph convolution process. Self-loop refers to adding an edge between the node of the graph and itself in the graph convolution module, that is, a node is connected to itself, so that when the node performs the convolution operation, it will not only aggregate features from adjacent nodes, but also retain its own feature information. Adding self-loops can ensure that the node will not lose its original information during the feature aggregation process; for The degree matrix represents the number of connections of each node; For the The weight matrix of the graph convolution layer learns the linear transformation of node features; is the activation function; and Respectively Layer and The output feature matrix of the layer graph convolution; after After layer iteration, the output feature matrix of the graph convolution module can be obtained .

[0084] Local temporal convolution module: contains a convolution layer and a pooling layer; the convolution layer outputs the feature matrix of the graph convolution module The local time series data in the convolution operation is performed, and the local time series data is the time series data (no more than one week); the pooling layer downsamples the convolutional features to extract the local spatiotemporal features of the input features, that is, the short-term time-dependent features; its specific form is:

[0085] ,

[0086] ,

[0087] in, represents a one-dimensional convolution operation, represents the pooling operation, Output feature matrix for graph convolution module, is the local temporal convolution kernel size, Output feature matrix for the convolutional layer, The local temporal convolution module outputs a feature matrix, i.e., local spatiotemporal features.

[0088] Global temporal convolution module: contains a convolution layer; outputs the feature matrix through the graph convolution module The global time series data in the convolution operation is used to extract the overall spatiotemporal characteristics of the input features, that is, the long-term time-dependent characteristics; the global time series data is the time series data of the long-term time window (more than one week); its specific form is:

[0089] ,

[0090] in, represents a one-dimensional convolution operation, Output feature matrix for graph convolution module, is the global temporal convolution kernel size, The global temporal convolution module outputs the feature matrix, i.e., the overall spatiotemporal features.

[0091] Attention module: It includes an attention mechanism layer and a feedforward neural network layer. The attention mechanism layer calculates the correlation between local spatiotemporal features and overall spatiotemporal features, and performs weighted average of local spatiotemporal features and overall spatiotemporal features. The feedforward neural network layer further refines the processing features and introduces activation functions to enhance the model's expression ability and nonlinear feature extraction ability, capturing the temporal and spatial dependencies between local spatiotemporal features and overall spatiotemporal features. Its specific form is:

[0092] ,

[0093] ,

[0094] ,

[0095] in, Let the query matrix be , and are the key matrix and value matrix respectively, both set to ; is the key matrix The transpose of is the index variable for the query matrix, Represents the index of the key matrix and the value matrix, that is, the target index Other indices for attention calculation, is the index of the sum of all key matrices, where the index corresponds to the node in the graph structure; is the dimension of the key, used to scale the dot product attention score; is the attention weight, indicating the index Index The intensity of attention; is the weighted feature matrix, through the attention weight pair value matrix Taking weighted average, For the The value matrix of the indexes; is the weight matrix of the feedforward neural network, is the feedforward neural network bias term, is the activation function; Output feature matrix for attention module.

[0096] Output module: contains a fully connected layer, which maps the high-dimensional feature vector (i.e., the output feature matrix of the attention module) to the target variable (i.e., the road surface temperature), and is used to generate the final road surface temperature prediction result based on the numerical model fusion deep learning; its specific form is:

[0097] ,

[0098] in, is the road surface temperature prediction result, is the weight matrix of the fully connected layer, is the bias term of the fully connected layer.

[0099] S32. Establish a Gaussian weighted loss function, define the mean square error weight as the inverse function of the probability density function of the Gaussian distribution of the road surface temperature, so as to solve the problem of sample imbalance of road surface temperature at different scales; its specific form is:

[0100] ,

[0101] ,

[0102] in, is the Gaussian weighted loss function, is the total number of samples used to calculate the loss function in model training, and They are respectively used to calculate the loss function in model training. The observation value (i.e., the observation data of the traffic meteorological automatic monitoring station) and the forecast value (i.e., the highway pavement temperature forecast result based on the numerical model fusion deep learning) of each sample; and are the mean and standard deviation of the observed samples used to calculate the loss function in model training; is the mean square error weight function, which is designed to make the deviation from the mean The weight of samples farther away is larger, making the model more sensitive to errors far from the mean, thus paying more attention to these errors during the training process and enhancing the model's ability to predict extreme road surface temperatures.

[0103] S33, for the deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network constructed in S31, the Gaussian weighted loss function established in S32 is used, and the Adam optimizer is used to iteratively optimize the model parameters, and the model training is carried out, such as Figure 4 As shown; its specific form is:

[0104] ,

[0105] in, and are the model parameters before and after iterative optimization, is the learning rate, is the Gaussian weighted loss function, Represents the gradient of the loss function to the model parameters; multiple iterative optimizations make the model converge and improve model performance.

[0106] S4. Use the deep learning pavement temperature forecast model based on the attention mechanism graph convolutional neural network trained in step S3 to generate a pavement temperature forecast result based on the numerical model fusion deep learning based on the multi-factor forecast data of the numerical weather forecast model and the regional high-resolution multivariate underlying surface data for the target time.

Claims

1. A road surface temperature prediction method based on numerical model fusion deep learning, characterized in that: The following steps are involved: The training set and validation set are produced based on the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station; Based on the regional high-resolution multivariate underlying surface data, a unified scale regional high-resolution multivariate underlying surface feature library is formed; A deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network is constructed, and the Gaussian weighted loss function is used to carry out model training; the model includes: input module, graph learning module, graph convolution module, local time convolution module, global time convolution module, attention module and output module. The input module is used to integrate the multi-factor forecast data based on the numerical weather forecast model, the observation data of the traffic meteorological automatic monitoring station and the regional high-resolution multivariate underlying surface feature library to obtain the input feature matrix. The graph learning module converts the input feature matrix into a graph structure to capture the temporal and spatial features between the input features. The graph convolution module is used to aggregate the input features of the target point with the graph structure. The features of the adjacency matrix of the neighboring points around the target point in the structure are extracted to extract the spatial dependency relationship between the nodes in the graph structure; the local time convolution module is used to convolve the local time series data in the feature matrix output by the graph convolution module, and extract the local spatiotemporal features of the input features; the global time convolution module is used to convolve the global time series data in the feature matrix output by the graph convolution module, and extract the overall spatiotemporal features of the input features; the attention module is used to capture the temporal and spatial dependency relationship between the local spatiotemporal features and the overall spatiotemporal features; the output module is used to map the feature matrix output by the attention module to the target variable, and generate the final road surface temperature prediction result based on the numerical model fusion deep learning; Using the trained model, for the target time, based on the multi-factor forecast data of the numerical weather forecast model and the regional high-resolution multivariate underlying surface data, the road surface temperature forecast results based on the numerical model fusion deep learning are generated.

2. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 1 is characterized in that: When making the training set and the validation set, it is necessary to remove the outliers in the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station, fill in the missing values, and then perform standardization respectively.

3. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 2 is characterized in that: The method used to fill missing values ​​is the central difference method.

4. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 1 is characterized in that: The method for integrating and forming a unified-scale regional high-resolution multivariate underlying surface feature library is: normalize the regional high-resolution multivariate underlying surface data separately to obtain multivariate unified-scale regional high-resolution underlying surface features, and then integrate them to obtain a unified-scale regional high-resolution multivariate underlying surface feature library.

5. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 1 is characterized in that: The deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network is constructed as follows: Input module: Acquire and integrate standardized numerical weather forecast model multi-factor forecast data , standardized traffic weather automatic monitoring station observation data and a unified scale regional high-resolution multivariate underlying surface feature database , forming the input feature matrix ; Graph learning module: The specific forms of the initial node feature matrix, adjacency matrix and degree matrix in the graph structure are: ， ， ， in, is the initial node feature matrix, and Respectively represent nodes and nodes , For Node and nodes The corresponding adjacency matrix The elements in For the node The number of connected edges, also known as the degree matrix The Line Column elements, finally forming an adjacency matrix Sum degree matrix ; Graph convolution module: contains multiple graph convolution layers, multiple residual connection layers and multiple jump connection layers; performs convolution operations on graph structure data, enhances gradient flow through residual connections, and retains global information through jump connections; is used to aggregate the input features of the target point and the features of the adjacency matrix of the neighboring points around the target point in the graph structure, and extract the spatial dependency features between nodes; its specific form is: ， in, is the activation function, for The degree matrix represents the number of connections of each node. is the adjacency matrix after adding self-loops, is the identity matrix, For the The weight matrix of the graph convolution layer, and Respectively Layer and The output feature matrix of the layer graph convolution is After layer iteration, the output feature matrix of the graph convolution module is obtained ; Local temporal convolution module: contains a convolution layer and a pooling layer; the convolution layer outputs the feature matrix of the graph convolution module The local time series data in the convolution operation is performed, and the pooling layer downsamples the convolutional features to extract the local spatiotemporal features of the input features, that is, the short-term time-dependent features; its specific form is: ， ， in, Output feature matrix for the convolutional layer, represents a one-dimensional convolution operation, is the local temporal convolution kernel size, represents the pooling operation, Output feature matrix for local temporal convolution module, i.e. local spatiotemporal features; Global temporal convolution module: contains a convolution layer; outputs the feature matrix through the graph convolution module The global time series data in is convolved to extract the overall spatiotemporal features of the input features, that is, the long-term time-dependent features; its specific form is: ， in, is the global temporal convolution kernel size, Output feature matrix for the global temporal convolution module, i.e., overall spatiotemporal features; Attention module: It includes an attention mechanism layer and a feedforward neural network layer. The attention mechanism layer calculates the correlation between local spatiotemporal features and overall spatiotemporal features, and performs weighted average of local spatiotemporal features and overall spatiotemporal features. The feedforward neural network layer further refines the processing features and introduces activation functions to enhance the model's expression ability and nonlinear feature extraction ability, capturing the temporal and spatial dependencies between local spatiotemporal features and overall spatiotemporal features. Its specific form is: ， ， ， in, Let the query matrix be , is the key matrix The transpose of the key matrix Set as , is the dimension of the key, For the The value matrix of the indexes; is the index variable for the query matrix, represents the index of the key matrix and the value matrix, is the index to sum over all key matrices; is the attention weight, indicating the index Index The intensity of attention; is the weighted feature matrix; is the weight matrix of the feedforward neural network, is the bias term of the feedforward neural network; Output feature matrix for attention module; Output module: contains a fully connected layer, which maps the output feature matrix of the attention module to the target variable, and is used to generate the final road surface temperature prediction result based on the numerical model fusion deep learning; its specific form is: ， in, is the road surface temperature prediction result, is the weight matrix of the fully connected layer, is the bias term of the fully connected layer.

6. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 1 is characterized in that: The Gaussian weighted loss function is: ， ， in, is the Gaussian weighted loss function, is the total number of samples used to calculate the loss function in model training, is the mean square error weight function, and They are respectively used to calculate the loss function in model training. The observed and predicted values ​​of samples; and are the mean and standard deviation of the observed samples used to calculate the loss function in model training.

7. The road surface temperature prediction method based on numerical model fusion deep learning according to claim 1 is characterized in that: The formula for iterative optimization of model parameters during model training is: ， in, and are the model parameters before and after iterative optimization, is the learning rate, is the Gaussian weighted loss function, Represents the gradient of the loss function with respect to the model parameters.

8. The road surface temperature prediction system based on numerical model fusion deep learning is characterized by: include: The data set production unit is used to produce training sets and validation sets based on the multi-factor forecast data of the numerical weather forecast model and the observation data of the traffic meteorological automatic monitoring station; Based on the regional high-resolution multivariate underlying surface data, a unified-scale regional high-resolution multivariate underlying surface feature library is formed; The model building and training unit is used to build a deep learning road surface temperature prediction model based on the attention mechanism graph convolutional neural network, and adopts a Gaussian weighted loss function to carry out model training; the model includes: input module, graph learning module, graph convolution module, local time convolution module, global time convolution module, attention module and output module. The input module is used to integrate the multi-factor forecast data based on the numerical weather forecast model, the observation data of the traffic meteorological automatic monitoring station and the regional high-resolution multivariate underlying surface feature library to obtain the input feature matrix. The graph learning module converts the input feature matrix into a graph structure to capture the temporal and spatial features between the input features. The graph convolution module is used to aggregate the input of the target point. The input features are combined with the features of the adjacency matrix of the neighboring points around the target point in the graph structure to extract the spatial dependency between the nodes in the graph structure; the local time convolution module is used to convolve the local time series data in the feature matrix output by the graph convolution module and extract the local spatiotemporal features of the input features; the global time convolution module is used to convolve the global time series data in the feature matrix output by the graph convolution module and extract the overall spatiotemporal features of the input features; the attention module is used to capture the temporal and spatial dependency between the local spatiotemporal features and the overall spatiotemporal features; the output module is used to map the feature matrix output by the attention module to the target variable to generate the final road surface temperature prediction result based on the numerical model fusion deep learning; The forecasting unit is used to use the trained model to generate road surface temperature forecast results based on numerical model fusion deep learning for the target time based on the multi-factor forecast data of the numerical weather forecast model and the regional high-resolution multivariate underlying surface data.

9. An electronic device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the steps of the road surface temperature prediction method based on numerical model fusion deep learning as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, which, when called, are used to execute the steps of the road surface temperature prediction method based on numerical model fusion deep learning as described in any one of claims 1 to 7.

Images