Multi-meteorological factor mode forecast temperature correction method based on deep learning

Abstract

The invention discloses a multi-meteorological factor mode forecast temperature correction method based on deep learning, and relates to the technical field of meteorological data processing, and the method comprises the steps: collecting multi-meteorological factor data and observation temperature field data, and outputting a meteorological chart structure; inputting the meteorological chart structure into a graph convolutional network for deep feature extraction, and outputting a multi-meteorological factor feature tensor; performing global and local space-time characteristic analysis on the multi-meteorological-factor characteristic tensor through a dynamic gating fusion network, and outputting a fusion characteristic tensor; inputting the fusion feature tensor into a full-connection neural network for nonlinear transformation and dimension mapping to form a temperature correction model of deep learning; performing different model parameter disturbances on the temperature correction model to generate an accuracy index; inputting into a quality detection model for training, and judging the quality of each temperature correction model. According to the invention, through dual-mechanism cooperation of meteorological chart structure construction and dynamic gating fusion, deep feature extraction and adaptive weight fusion of multiple meteorological factors are realized.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, and in particular to a method for correcting temperature forecasts based on a multi-meteorological factor model using deep learning. Background Technology

[0002] In recent years, meteorological forecast temperature correction technology has gradually evolved towards multi-factor collaboration and deep learning integration. Mainstream solutions are based on multi-meteorological factor (such as air pressure, humidity, and wind field) data output from numerical models, employing convolutional neural networks (CNNs) or long short-term memory networks (LSTMs) for feature extraction and error correction. Among these, multi-source data fusion frameworks are gradually incorporating attention mechanisms to optimize factor weight allocation, graph neural networks (GNNs) are being applied to model the spatial relationships of meteorological elements, and generative adversarial networks (GANs) are being explored for temperature field distribution reconstruction. End-to-end training paradigms significantly improve correction efficiency. Deep learning models, through their nonlinear mapping capabilities, are gradually replacing traditional statistical correction methods, providing a technological foundation for high-precision temperature forecasts.

[0003] However, existing technologies have limitations in modeling factor coupling mechanisms and efficiently evaluating model quality. Traditional multi-factor processing methods often rely on static feature splicing or simple weighted fusion, failing to explicitly construct the physical relationship topology between meteorological factors. This results in insufficient model learning ability for nonlinear interactions between factors, a lack of adaptive adjustment mechanisms for factor contributions under dynamic weather scenarios, and model evaluation entirely dependent on computationally intensive posterior validation. Furthermore, the lack of a rapid quality screening paradigm based on meta-features makes it difficult to balance the timeliness of operational deployment with the reliability of model selection. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a deep learning-based method for correcting temperature forecasts using multi-meteorological factor models, addressing limitations in factor coupling mechanism modeling and efficient model quality evaluation.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides a deep learning-based method for temperature correction in multi-meteorological factor model forecasts. The method includes: collecting multi-meteorological factor data and observed temperature field data, and preprocessing them; treating each meteorological factor in the multi-meteorological factor data as a factor node, treating the coupling relationships between meteorological factors as factor association edges, and assigning weights to these association edges; integrating the factor nodes, factor association edges, and association edge weights to output a meteorological map structure; inputting the meteorological map structure into a graph convolutional network for deep feature extraction, and outputting a multi-meteorological factor feature tensor; performing global and local spatiotemporal characteristic analysis on the multi-meteorological factor feature tensor through a dynamic gating fusion network to generate dynamic weight coefficients; and weighted fusion of the dynamic weight coefficients and the multi-meteorological factor feature tensor. Output a fused feature tensor; input the fused feature tensor into a fully connected neural network for nonlinear transformation and dimensionality mapping, outputting corrected temperature field data; compare the corrected temperature field data with the observed temperature field data, outputting temperature correction values; based on the temperature correction values, optimize the network parameters of the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network using the backpropagation algorithm to form a deep learning temperature correction model; apply different model parameter perturbations to the temperature correction model, outputting a set of correction models; perform quality verification on each temperature correction model in the set, generating an accuracy index; use the accuracy index and the model meta-features of each temperature correction model as a training set, input them into a quality detection model for training, and judge the quality of each temperature correction model.

[0008] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method described in this invention, the specific steps of inputting the meteorological map structure into a graph convolutional network for deep feature extraction and outputting a multi-meteorological factor feature tensor are as follows:

[0009] The meteorological map structure is input into a graph convolutional network, and the neighborhood information of the factor nodes is aggregated through the message passing mechanism of the graph convolutional network to output the node feature tensor.

[0010] The node feature tensor is transformed and nonlinearly activated to output a multi-meteorological factor feature tensor.

[0011] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the step of performing global and local spatiotemporal characteristic analysis on the feature tensors of multiple meteorological factors through a dynamic gating fusion network to generate dynamic weight coefficients includes the following specific steps:

[0012] A dynamic gated fusion network is used to perform spatiotemporal convolution on the feature tensors of multiple meteorological factors to output local spatiotemporal features.

[0013] Global pooling is performed on the feature tensors of multiple meteorological factors to output global statistical properties;

[0014] Local spatiotemporal features and global statistical properties are concatenated to generate a comprehensive feature vector;

[0015] The comprehensive feature vector is linearly transformed and normalized to generate dynamic weight coefficients.

[0016] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method described in this invention, the specific steps of weighted fusion of dynamic weight coefficients and multi-meteorological factor feature tensors to output a fused feature tensor are as follows:

[0017] The dynamic weight coefficients and the feature tensors of multiple meteorological factors are weighted by coefficients, and the weighted feature tensor is output.

[0018] The weighted feature tensors are classified and aggregated along the feature dimensions of meteorological factors, and the aggregated feature tensors are output.

[0019] Adjust the dimensions and representation of the aggregated feature tensor to output the fused feature tensor.

[0020] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the specific steps of inputting the fused feature tensor into a fully connected neural network for nonlinear transformation and dimensionality mapping, and outputting corrected temperature field data, are as follows:

[0021] The fused feature tensor is input into a fully connected neural network. The first fully connected layer of the fully connected neural network performs a linear transformation on the fused feature tensor to generate an intermediate feature representation.

[0022] The intermediate feature representation is nonlinearly transformed by the ReLU activation function to generate a nonlinear feature representation.

[0023] The nonlinear feature representation is mapped to the temperature field dimension through a second fully connected layer, and the corrected temperature field data is output.

[0024] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method described in this invention, the specific steps of comparing the corrected temperature field data with the observed temperature field data and outputting the temperature correction value are as follows:

[0025] The corrected temperature field data is spatially and temporally aligned with the observed temperature field data;

[0026] The aligned and corrected temperature field data are compared with the observed temperature field data element by element, and the temperature correction value is output.

[0027] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the following steps are taken: Based on the temperature correction value, the network parameters of the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network are optimized through the backpropagation algorithm to form a deep learning temperature correction model.

[0028] Perform root mean square error statistics on each temperature correction value and output the loss function value;

[0029] The loss function value is used as the starting point of the backpropagation algorithm, and the gradient information of the graph convolutional network, dynamic gated fusion network and fully connected neural network corresponding to the loss function value is solved by the chain rule.

[0030] The Adam optimizer is used to iteratively update the network parameters of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks according to gradient information, and the updated network parameters are output.

[0031] The updated network parameters are loaded into the architecture of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks for optimization and adjustment, and then integrated to form a deep learning temperature correction model.

[0032] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the specific steps of perturbing the temperature correction model with different model parameters and outputting a correction model set are as follows:

[0033] Perform parameter sensitivity analysis on the temperature correction model to determine the range of network parameters that need to be perturbed;

[0034] Based on the range of network parameters, set multiple different combinations of parameter configurations;

[0035] The temperature correction model is initialized by configuring parameters in different combinations, and the correction model set is output.

[0036] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the step of quality verification of each temperature correction model in the correction model set and generating an accuracy index refers to using multi-meteorological factor data and observed temperature field data as training set and validation set respectively, using the training set to perform temperature correction prediction on each temperature correction model in the correction model set, comparing the temperature correction prediction results with the validation set, and performing quantitative analysis on the difference comparison results to generate an accuracy index.

[0037] As a preferred embodiment of the deep learning-based multi-meteorological factor model forecast temperature correction method of the present invention, the specific steps are as follows: The accuracy index and the model meta-features of each temperature correction model are used as a training set and input into a quality detection model for training to determine the quality of each temperature correction model.

[0038] A quality detection model was built based on the scikit-learn machine learning framework.

[0039] The model meta-features of each temperature correction model are standardized to generate standardized feature vectors;

[0040] The standardized feature vectors are paired with the accuracy index to form the detection training samples;

[0041] The quality detection model is trained using the test training samples under supervised learning. The trained quality detection model is then used to perform quality detection on the temperature correction model, and the quality judgment result is output.

[0042] The beneficial effects of this invention are as follows: Through the synergistic effect of meteorological map structure construction and dynamic gating fusion, deep feature extraction and adaptive weight fusion of multiple meteorological factors are achieved. By integrating meteorological factors as factor nodes and constructing associated edge weights into a meteorological map structure, and utilizing the message passing mechanism of graph convolutional networks to aggregate neighborhood information, explicit modeling of complex coupling relationships among multiple factors is achieved, providing spatially correlated feature representations for temperature correction. The dynamic gating fusion network performs spatiotemporal convolution and global pooling on the feature tensors of multiple meteorological factors, achieving adaptive calibration of the contribution of multiple factors and improving the discriminative ability of feature fusion. Through an end-to-end network optimization and quality verification system, a high-precision and highly generalizable temperature correction model is formed. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a flowchart of a method for correcting temperature forecasts using a deep learning-based multi-meteorological factor model.

[0045] Figure 2 This is a flowchart of deep feature extraction for graph convolutional networks.

[0046] Figure 3 A flowchart for feature fusion in a dynamic gating fusion network.

[0047] Figure 4 A flowchart for training the quality verification and quality inspection model. Detailed Implementation

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0049] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0050] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0051] Reference Figures 1-4 This is one embodiment of the present invention, which provides a method for correcting temperature forecasts based on a deep learning-based multi-meteorological factor model, comprising the following steps:

[0052] S1. Collect multi-meteorological factor data and observed temperature field data, and perform preprocessing; take each meteorological factor in the multi-meteorological factor data as factor nodes, take the coupling relationship between each meteorological factor as factor association edge, and assign association edge weight to the factor association edge; integrate factor nodes, factor association edges and association edge weights to output meteorological map structure.

[0053] Specifically, data on multiple meteorological factors, including temperature, humidity, air pressure, wind speed, and cloud cover, are collected from various data sources such as numerical weather prediction models, ground meteorological observation stations, and satellite remote sensing. Simultaneously, actual observed temperature field data within the corresponding spatiotemporal range are also collected. The corresponding spatiotemporal range refers to the area covered by the multi-meteorological factor data output by the numerical weather prediction model and the actual observed temperature field data within the same time period and geographical region. The collected multi-meteorological factor data and observed temperature field data undergo data cleaning and normalization. Data cleaning includes checking and processing missing values ​​and outliers that clearly exceed physically reasonable ranges in the multi-meteorological factor data. "Significantly exceeding the physically reasonable range" refers to meteorological factor values ​​that violate basic atmospheric physics or historical statistical patterns, such as surface temperatures exceeding ±60℃, relative humidity greater than 100% or negative, and sea level pressure below 870 hPa or above 1080 hPa. Normalization processing maps the numerical ranges of all multi-meteorological factor data and observed temperature field data to a uniform range of zero to one through linear transformation, eliminating differences in dimensions and numerical ranges between different meteorological factors and providing standardized data input for subsequent calculations. The preprocessing outputs clean and standardized multi-meteorological factor data and standardized observed temperature field data with uniform numerical ranges.

[0054] Each meteorological factor in the preprocessed standardized multi-meteorological factor data, such as temperature, humidity, and air pressure, is defined as an independent factor node. Each factor node is mathematically represented using a factor eigenvector, which contains the standardized values ​​of the meteorological factor at multiple spatial locations and time steps. The Pearson correlation coefficient between the factor eigenvectors of every two different meteorological factors is calculated. This coefficient is evaluated by measuring the consistency of the changing trends of each meteorological factor at multiple spatial locations and time steps. The Pearson correlation coefficient is used as the strength of the statistical coupling relationship between the meteorological factors. Each statistical coupling relationship is defined as a factor association edge connecting two factor nodes. Each factor association edge is assigned a weight, the specific value of which is the absolute value of the Pearson correlation coefficient, representing the strength of the coupling relationship between the corresponding two meteorological factors.

[0055] All defined factor nodes, all factor-related edges, and the weights of each factor-related edge are integrated. The integration method follows the mathematical norms of graph theory, treating factor nodes as vertices of the graph, factor-related edges as edges connecting vertices, and the weights of the related edges as attributes of the corresponding edges, thereby constructing a complete and weighted meteorological map structure. The meteorological map structure fully expresses the coupling relationship network between all individual meteorological factors and groups of meteorological factors in multi-meteorological factor data.

[0056] S2. Input the meteorological map structure into the graph convolutional network for deep feature extraction and output a multi-meteorological factor feature tensor.

[0057] The meteorological map structure is input into a graph convolutional network, and the neighborhood information of the factor nodes is aggregated through the message passing mechanism of the graph convolutional network to output the node feature tensor.

[0058] Specifically, for each factor node, the graph convolutional network traverses all factor-related edges connected to the factor node, collects factor feature vectors of neighboring factor nodes, and performs weighted summarization of neighborhood information according to the corresponding edge weights. The weighted summarization process is achieved by linearly combining the factor feature vectors of neighboring factor nodes, and the weight value is the edge weight of the factor-related edge. The edge weight emphasizes the contribution of neighboring factor nodes with strong coupling relationships to the current factor node. The weighted summarization information is combined with the factor feature vector of the current factor node to generate an updated node feature representation. The graph convolutional network performs message passing and information aggregation through multiple convolutional iterations. Each iteration further integrates a wider range of neighborhood information, thereby capturing the complex spatial and correlational dependencies between meteorological factors. After each convolutional operation, an intermediate representation containing the updated features of all factor nodes is generated. After processing by all convolutional layers, a node feature tensor is output. The node feature tensor retains the feature information of each factor node in terms of spatial location and time step, while incorporating the coupling relationship between meteorological factor groups, providing rich deep feature representations for subsequent feature transformations.

[0059] Furthermore, the graph convolutional network is constructed based on the topological relationship of the meteorological map structure. The network architecture consists of multiple layers of graph convolutional layers connected sequentially. Each layer of the graph convolutional network includes a message passing mechanism and a feature transformation mechanism. The message passing mechanism obtains the information aggregation information between each factor node and its neighboring factor nodes based on the factor association edges and their weights in the input meteorological map structure. The feature transformation mechanism performs dimensional mapping and nonlinear transformation on the aggregated node features through a nonlinear activation function. The multiple graph convolutional layers iteratively execute neighborhood information aggregation and feature transformation to gradually expand the feature perception of each factor node, thereby deeply mining the complex spatial coupling and nonlinear dependency relationships between multiple meteorological factors and outputting a node feature tensor that integrates global neighborhood information as the result of deep feature extraction of multiple meteorological factors.

[0060] Perform feature transformation and nonlinear activation on the node feature tensor to output a multi-meteorological factor feature tensor;

[0061] Specifically, feature transformation is performed on the node feature tensor. A linear transformation adjusts the dimension and reconstructs the features of each factor node's feature vector. This linear transformation, achieved through matrix multiplication, maps each factor feature vector of the node feature tensor to the feature space, enhancing feature expressiveness and adapting to subsequent processing requirements. The linearly transformed node feature tensor retains the feature information of all factor nodes. A nonlinear activation function is applied to the linearly transformed node feature tensor, using the ReLU activation function to nonlinearly process each element of each factor feature vector. Specifically, the ReLU activation function sets negative values ​​in the factor feature vector to zero while keeping positive values ​​unchanged, thus introducing nonlinear characteristics and enhancing the node feature tensor's ability to model complex meteorological factor relationships. The nonlinearly activated node feature tensor integrates the transformation features of each factor node, forming a more expressive feature representation. The node feature tensors, after feature transformation and nonlinear activation, are then formatted into a unified format, outputting a multi-meteorological factor feature tensor. This multi-meteorological factor feature tensor contains the deep features of all meteorological factors in terms of spatial location and time step, fusing the coupling relationships and nonlinear characteristics between factors, providing standardized input data for the subsequent spatiotemporal characteristic analysis of the dynamic gating fusion network.

[0062] S3. Perform global and local spatiotemporal characteristic analysis on the feature tensor of multiple meteorological factors through a dynamic gating fusion network to generate dynamic weight coefficients; then perform weighted fusion of the dynamic weight coefficients and the feature tensor of multiple meteorological factors to output the fused feature tensor.

[0063] A dynamic gated fusion network is used to perform spatiotemporal convolution on the feature tensors of multiple meteorological factors to output local spatiotemporal features.

[0064] Specifically, a dynamic gating fusion network is constructed based on a deep learning framework, using multi-meteorological factor feature tensors and observed temperature field data as training data for training. The dynamic gating fusion network performs spatiotemporal convolution on the multi-meteorological factor feature tensors to extract local spatiotemporal features. Spatiotemporal convolution, by applying convolution kernels in the time and spatial dimensions, captures the dynamic changes of the multi-meteorological factor feature tensors over time and the local patterns in the spatial grid, respectively. Specifically, the spatiotemporal convolution kernel slides along the time step to extract the feature changes of each meteorological factor at consecutive time points, and simultaneously slides along the spatial location to capture the local correlations between adjacent spatial grid points. The size of the convolution kernel... The stride is set according to the temporal and spatial resolution of the multi-meteorological factor feature tensor to ensure sufficient temporal and spatial coverage and fully explore local spatiotemporal patterns. Spatiotemporal convolution generates intermediate feature representations reflecting local spatiotemporal dynamics by weighted summation of each feature vector of the multi-meteorological factor feature tensor. The dynamic gating fusion network further introduces a gating mechanism to selectively filter the intermediate feature representations of the convolution operation, enhancing the focus on key spatiotemporal features. The gating mechanism retains local features that contribute significantly to temperature correction by obtaining the weight of each feature vector, while suppressing the influence of irrelevant or noisy features. After processing by spatiotemporal convolution and gating mechanism, local spatiotemporal features are output.

[0065] Global pooling is performed on the feature tensors of multiple meteorological factors to output global statistical properties;

[0066] Specifically, global pooling aggregates all feature vectors of a multi-meteorological factor feature tensor across spatial and temporal dimensions to generate global statistical properties. The aggregation employs average pooling, averaging the feature vectors of each meteorological factor across all spatial locations and time steps to generate a statistical feature vector reflecting the global mean. Variance pooling is then used to calculate the variance of the feature vectors of each meteorological factor across all spatial locations and time steps, generating a statistical feature vector reflecting global variability. The mean and variance represent the central tendency and dispersion of the multi-meteorological factor feature tensor across the entire spatiotemporal range, respectively, together constituting the global statistical properties. The global pooling operation reduces the dimensionality of the multi-meteorological factor feature tensor into a low-dimensional statistical feature representation, and after global pooling, the global statistical properties are output.

[0067] Local spatiotemporal features and global statistical properties are concatenated to generate a comprehensive feature vector;

[0068] Specifically, the feature vectors of local spatiotemporal features and global statistical characteristics are sequentially connected to form a comprehensive feature vector, with the feature vectors of local spatiotemporal features first and the feature vectors of global statistical characteristics second. The splicing process ensures that the dimensions of local spatiotemporal features and global statistical characteristics are aligned. This is achieved by standardizing and adjusting the dimensions of local spatiotemporal features and global statistical characteristics before splicing, ensuring that they have the same feature length. The spliced ​​comprehensive feature vector integrates the dynamic details of local spatiotemporal features and the overall patterns of global statistical characteristics, forming a comprehensive representation of the feature tensor of multiple meteorological factors.

[0069] A linear transformation and normalization are performed on the comprehensive feature vector to generate dynamic weight coefficients;

[0070] The linear transformation maps the dimension of the comprehensive feature vector through matrix multiplication, converting the high-dimensional comprehensive feature vector into a low-dimensional representation suitable for generating weight coefficients. The weights of the linear transformation are initialized based on the temporal and spatial characteristics of the multi-meteorological factor feature tensor to ensure that the comprehensive feature vector after linear transformation highlights the feature components most important for temperature correction. The comprehensive feature vector after linear transformation retains the core information and is normalized. Specifically, the softmax function is used to process each element of the comprehensive feature vector. The softmax function maps the values ​​of the comprehensive feature vector to the interval between 0 and 1 and ensures that the sum of all elements is 1, thereby generating a set of normalized dynamic weight coefficients. The dynamic weight coefficients reflect the relative importance of different feature components in the comprehensive feature vector to temperature correction and can dynamically adjust the contribution ratio of subsequent feature fusion. The normalization operation enhances the stability and interpretability of the weight coefficients.

[0071] The dynamic weight coefficients and the feature tensors of multiple meteorological factors are weighted by coefficients, and the weighted feature tensor is output.

[0072] Specifically, the dynamic weight coefficients are aligned with the feature dimensions of the multi-meteorological factor feature tensor through the dimension repetition method, ensuring that each weight value corresponds to a feature component; the weighting process is implemented through dot product, applying the dynamic weight coefficients to each feature vector of the multi-meteorological factor feature tensor to generate a weighted feature tensor; the coefficient weighting retains the original structure of the multi-meteorological factor feature tensor, only adjusting the magnitude of the feature values ​​to reflect the importance of the dynamic weight coefficient allocation;

[0073] The formula for calculating the weighted feature tensor is as follows:

[0074] ;

[0075] in, This represents the weighted feature tensor obtained by weighting the feature tensors of multiple meteorological factors using coefficients. Indicates dynamic weighting coefficients. Represents a multi-meteorological factor characteristic tensor. This represents the element-wise multiplication operator.

[0076] The weighted feature tensors are classified and aggregated along the feature dimensions of meteorological factors, and the aggregated feature tensors are output.

[0077] Specifically, based on the type of meteorological factors (such as temperature, humidity, and air pressure), the weighted feature tensor is grouped according to factor category, with each group corresponding to a set of feature vectors for the meteorological factor. Each set of feature vectors is then aggregated using a weighted summation method, fusing and summing the feature vectors of the same meteorological factor across all spatial locations and time steps. The weights for fusion and summation are determined based on the values ​​of each feature component in the weighted feature tensor to ensure that the aggregation result reflects the contribution of key features. After classification and aggregation, an aggregated feature vector is generated for each group of meteorological factors, containing a comprehensive feature representation of the meteorological factor within the spatiotemporal range. Finally, the aggregated feature vectors of all meteorological factors are sorted to generate an aggregated feature tensor.

[0078] Adjust the dimensions and representation of the aggregated feature tensor to output the fused feature tensor;

[0079] Specifically, a linear transformation is used to map the dimensions of the aggregated feature tensor, adjusting its feature dimensions to a fixed size to match the input dimension requirements of a fully connected neural network. Representation adjustment is then performed on the dimension-adjusted aggregated feature tensor, standardizing the mean and variance of the feature vectors in each aggregated feature tensor through normalization. The feature values ​​are then rescaled to a distribution with a mean of zero and a variance of 1, enhancing the stability and generalization ability of the features. The feature tensors, after dimension and representation adjustments, are then organized into a unified format, outputting a fused feature tensor. This fused feature tensor contains optimized feature representations of all meteorological factors, with fixed dimensions and stable representation.

[0080] S4. Input the fused feature tensor into the fully connected neural network for nonlinear transformation and dimension mapping, and output corrected temperature field data; compare the corrected temperature field data with the observed temperature field data to output the temperature correction value; based on the temperature correction value, optimize the network parameters of the graph convolutional network, the dynamic gated fusion network and the fully connected neural network through the backpropagation algorithm to form a deep learning temperature correction model.

[0081] The fused feature tensor is input into a fully connected neural network. The first fully connected layer of the fully connected neural network performs a linear transformation on the fused feature tensor to generate an intermediate feature representation.

[0082] Specifically, the first fully connected layer of the fully connected neural network receives the fused feature tensor, performs a linear transformation, and maps each feature vector in the fused feature tensor to a new feature space through matrix multiplication. The weight parameters used in the transformation process are configured according to the dimension of the fused feature tensor during initialization to ensure that the input features can be effectively transformed. The linear transformation performs a row linear combination of all feature components in the fused feature tensor, preserving the correlation structure between the original features while enhancing the expressive power of the features. The transformed result forms a new feature vector, which is used as an intermediate feature representation. The intermediate feature representation has a higher level of abstraction than the fused feature tensor and can better support the subsequent nonlinear modeling process.

[0083] The intermediate feature representation is nonlinearly transformed by the ReLU activation function to generate a nonlinear feature representation.

[0084] The ReLU activation function keeps positive elements in the intermediate feature representation unchanged and sets negative elements to zero, breaking the linear constraints imposed by the linear transformation and introducing nonlinear characteristics. This enables fully connected neural networks to fit complex meteorological relationships. The nonlinear transformation process, without changing the dimension of the intermediate feature representation, selects and enhances feature responses that contribute to temperature correction and suppresses the expression of weakly correlated or redundant features. The transformed feature vector set constitutes the nonlinear feature representation. The nonlinear feature representation inherits the spatial and temporal organization of the intermediate feature representation in structure, but its numerical distribution is more concentrated in the positive value region, enhancing the sparsity and discriminativeness of the features. As the output of the nonlinear transformation, the nonlinear feature representation provides input data with nonlinear expressive capabilities for subsequent dimensionality mapping.

[0085] The temperature field dimension is mapped to the nonlinear feature representation through the second fully connected layer, and the corrected temperature field data is output.

[0086] Specifically, the nonlinear feature representation is fed into the second fully connected layer, which performs temperature field dimension mapping, transforming the nonlinear feature representation from the current feature space to an output space consistent with the target temperature field. The temperature field dimension mapping is achieved through linear transformation. The weights and bias parameters used in the transformation process are set according to the number of spatial grids and the time step of the target temperature field to ensure that the output dimension is completely matched with the structure of the observed temperature field data. The second fully connected layer combines each feature vector in the nonlinear feature representation to generate temperature prediction values ​​corresponding to the spatial location and time step. All temperature prediction values ​​are arranged in the original spatiotemporal order to form complete corrected temperature field data.

[0087] The corrected temperature field data is spatially and temporally aligned with the observed temperature field data;

[0088] Specifically, spatial alignment is achieved by matching the geographic coordinate grids of the corrected temperature field data and the observed temperature field data, ensuring that each spatial location in the corrected temperature field data precisely corresponds to the corresponding location in the observed temperature field data, without changing the resolution and coverage of the two data sets during the alignment process. Temporal alignment is completed by synchronizing the time step sequences of the corrected and observed temperature field data, ensuring that each time point in the corrected temperature field data is consistent with the corresponding time point in the observed temperature field data, with the start time, end time, and time interval of the time series perfectly matched. After the alignment operation is completed, the corrected temperature field data and the observed temperature field data achieve point-to-point correspondence on the spatial grid and time series, forming a data pair with consistent structure.

[0089] The aligned and corrected temperature field data are compared with the observed temperature field data element by element, and the temperature correction value is output.

[0090] Specifically, difference statistics calculate the difference between the corrected temperature value and the observed temperature value at each spatial grid point in the corrected temperature field data and the observed temperature field data, generating a set of difference data containing all the difference results. The difference results reflect the deviation between the predicted output and the actual observation. Positive values ​​indicate that the corrected temperature is too high, and negative values ​​indicate that the corrected temperature is too low. The absolute value of the difference indicates the degree of deviation. All differences are organized in the order of the observed temperature field data to form temperature correction values.

[0091] Perform root mean square error statistics on each temperature correction value and output the loss function value;

[0092] Mean squared error (MSE) statistics perform a global error assessment on temperature correction values. The statistical process squares all temperature correction values ​​to eliminate the cancellation of positive and negative differences and amplify the impact of large errors. The average of the squared differences is calculated across all spatial locations and time steps, ensuring comprehensive error assessment by covering all data points throughout the entire spatiotemporal range. The MSE result is a loss function value, reflecting the overall prediction accuracy; a smaller value indicates better prediction correction. The loss function value guides the adjustment direction of network parameters, directly affecting the generation of gradient information and the magnitude of parameter updates.

[0093] The formula for calculating the loss function value is as follows:

[0094] ;

[0095] in, This represents the loss function value calculated from the mean square error of each temperature correction value. This represents the number of temperature corrections across all spatial locations and time steps. Index representing temperature correction values, Indicates the first Temperature correction value.

[0096] The loss function value is used as the starting point of the backpropagation algorithm, and the gradient information of the graph convolutional network, dynamic gated fusion network and fully connected neural network corresponding to the loss function value is solved by the chain rule.

[0097] Specifically, the backpropagation algorithm starts with the loss function value and propagates the error signal backward layer by layer from the output to the input. The chain rule is used to decompose the partial derivatives of the loss function value with respect to the parameters of each layer in the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network. The gradient of the loss function value with respect to the parameters of the second fully connected layer is calculated. Then, the gradients of the nonlinear feature representation, intermediate feature representation, fused feature tensor, aggregated feature tensor, weighted feature tensor, multi-meteorological factor feature tensor, and node feature tensor are calculated sequentially forward until the initial parameters of the graph convolutional network are traced back. The gradient information reflects the sensitivity of the loss function value to each network parameter and reflects the direction and magnitude of the network parameter update. The gradient calculation process is strictly executed in reverse according to the forward propagation path of the network to ensure that the partial derivatives of each step correspond to the forward operation. The gradient information includes the partial derivatives of all trainable parameters in the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network, providing a basis for subsequent parameter updates.

[0098] The Adam optimizer is used to iteratively update the network parameters of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks according to gradient information, and the updated network parameters are output.

[0099] Specifically, the Adam optimizer receives gradient information and uses it to optimize the network parameters of graph convolutional networks, dynamically gated fusion networks, and fully connected neural networks. The gradient optimization process updates the convolutional weights in the graph convolutional network, the kernel parameters and gate weights in the dynamically gated fusion network, and the connection weights and bias terms in the fully connected neural network. The update operation is performed in each training iteration step, with the parameter adjustment direction opposite to the gradient direction. The adjustment magnitude is determined by the learning rate and the gradient magnitude. For example, when the gradient value of the fully connected layer weights is large, the Adam optimizer automatically reduces the learning rate of the network parameters based on the moving average of the squared historical gradients to prevent excessively large update steps that could cause oscillations. Conversely, for network parameters with smaller gradients, the learning rate is increased to accelerate convergence. The Adam optimizer uses an adaptive mechanism to balance the update speed of different parameters, avoiding gradient explosion or vanishing problems and improving convergence stability. The iterative update process continues until the loss function value converges, at which point the updated network parameters are output.

[0100] The updated network parameters are loaded into the architecture of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks for optimization and adjustment, and then integrated to form a deep learning temperature correction model.

[0101] Specifically, the updated network parameters are reloaded into the corresponding layers of the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network, replacing the original parameter values ​​and completing the optimization and adjustment of the network architecture. The parameter loading process ensures that the shape and dimension of the parameters in each layer match the network structure, avoiding parameter misalignment or loss. The optimized graph convolutional network, the dynamic gated fusion network, and the fully connected neural network work together to form a complete processing flow, from meteorological map structure input to corrected temperature field data output, forming a deep learning temperature correction model with stronger temperature correction capabilities and the ability to more accurately capture the complex relationship between multiple meteorological factors and temperature.

[0102] S5. Perturb the model parameters of the temperature correction model with different parameters and output the correction model set; perform quality verification on each temperature correction model in the correction model set and generate an accuracy index; use the accuracy index and the model meta-features of each temperature correction model as the training set, input them into the quality detection model for training, and judge the quality of each temperature correction model.

[0103] Perform parameter sensitivity analysis on the temperature correction model to determine the range of network parameters that need to be perturbed;

[0104] Specifically, a parameter sensitivity analysis is performed on the network parameters within the temperature correction model. This analysis assesses the impact of individual or local network parameters on the prediction results by perturbing single or local network parameters within the model and observing the deviations in the spatial distribution, regional extreme value changes, and temporal trends of the corrected temperature field data output by the model. The magnitude of the impact is determined based on whether the overall temperature field structure undergoes significant distortion, whether the regional temperature gradient fluctuates, and whether temporal continuity is disrupted. If the perturbation causes temperature deviations or abrupt changes in the temperature field, the single or local network parameter is considered to have a high impact and is considered a sensitive parameter; if the output changes smoothly and maintains the original spatiotemporal pattern, the impact is considered low. The sensitivity of the network parameters is determined based on the significance of the impact. Highly sensitive network parameters cause significant fluctuations in temperature correction values ​​after perturbation, while low-sensitivity network parameters are less affected. Based on the results of parameter sensitivity analysis, the categories of sensitive parameters affecting the output of the temperature correction model are selected, and perturbation ranges are determined for the network parameters. The perturbation range is set with the network parameter values ​​as the center and upper and lower floating boundaries are set. For example, a perturbation range of ±15% is set for the convolution weights corresponding to the strongly coupled factor-related edges in the graph convolutional network, a perturbation range of ±20% is set for the gate weights in the dynamic gated fusion network, and a perturbation range of ±10% is set for the connection weights of the second fully connected layer in the fully connected neural network. This ensures that the perturbed parameters are still within the physically interpretable and numerically stable range. After the parameter sensitivity analysis is completed, the range of network parameters that need to be perturbed is output.

[0105] Based on the range of network parameters, set multiple different combinations of parameter configurations;

[0106] Specifically, the network parameter range serves as the basis for parameter configuration and is used to generate multiple sets of differentiated parameter configuration combinations. Each set of parameter configuration combinations selects several network parameters from the network parameter range, and randomly samples or divides them according to a preset perturbation range to generate multiple different parameter configuration combinations. The preset perturbation range is set based on the results of parameter sensitivity analysis, combined with the importance of the network layer where the network parameter is located, the historical statistics of the gradient magnitude, and the numerical stability requirements, with the value of the network parameter as the center and the percentage range of fluctuation above and below.

[0107] The temperature correction model is initialized with model copies by combining parameter configurations, and the correction model set is output.

[0108] Each set of parameter configurations is used to initialize a copy of the temperature correction model. During initialization, the network architecture of the temperature correction model remains unchanged, and only the network parameters that need to be perturbed are replaced with the specific values ​​in the corresponding parameter configuration combination. The remaining unperturbed parameters retain their original trained values. Each copy of the temperature correction model inherits the complete structure and forward inference process of the temperature correction model, but exhibits different prediction characteristics due to parameter differences. All parameter configuration combinations are applied sequentially to generate multiple copies of the temperature correction model with parameter differences. Each copy of the temperature correction model can independently perform the temperature correction task. After all the copies of the temperature correction model are initialized, the sets of copies are organized into a correction model set. The correction model set contains multiple temperature correction models with perturbed parameters, and each temperature correction model corresponds to a unique set of parameter configurations.

[0109] The quality validation of each temperature correction model in the correction model set and the generation of accuracy index involve using multi-meteorological factor data and observed temperature field data as training and validation sets, respectively. The training set is used to perform temperature correction predictions on each temperature correction model in the correction model set. The results of the temperature correction predictions are compared with the results of the validation set, and the results of the difference comparison are quantitatively analyzed. The quantitative analysis refers to evaluating the spatial and temporal distribution characteristics of the deviation between the predicted output of each temperature correction model and the observed temperature field data, judging the degree of deviation of the deviation from the true value. That is, a small deviation and a concentrated distribution indicate that the correction result is close to the actual observation result. The corresponding accuracy index is assigned according to the level of deviation. The higher the accuracy index, the better the correction effect of the temperature correction model.

[0110] A quality detection model was built based on the scikit-learn machine learning framework.

[0111] Specifically, the quality detection model is built within the scikit-learn machine learning framework, selecting support vector machines or random forests as the basic learning algorithms. The input to the quality detection model is the meta-features of the temperature correction model, and the output is the corresponding quality judgment result. The construction process of the quality detection model includes data preparation, algorithm selection, hyperparameter setting, and training process configuration. In the data preparation stage, structural information and training process information of each temperature correction model in the correction model set are collected, and quantifiable meta-features are extracted, such as the number of network layers, total number of parameters, number of training epochs, and loss function descent rate. The basic learning algorithm is selected based on the characteristics of the task to determine the classification or regression mode. The structure and training process of the quality detection model are defined within the scikit-learn framework.

[0112] The model meta-features of each temperature correction model are standardized to generate standardized feature vectors;

[0113] Specifically, the meta-features of the temperature correction model refer to quantifiable indicators that describe the structural properties and dynamics of the training process of the temperature correction model itself, and are used to characterize the intrinsic characteristics of the temperature correction model. The meta-features of the temperature correction model are obtained by parsing the network architecture configuration and training log information of each temperature correction model in the correction model set, including the number of network layers, total number of parameters, activation function type, learning rate setting, number of training epochs, and gradient change magnitude, etc.

[0114] Furthermore, the model meta-features of each temperature correction model differ in value and dimension. To eliminate the impact of these differences on the training of the quality detection model, the model meta-features of the temperature correction model need to be standardized. The standardization process uses the Z-score method to calculate the mean and standard deviation of the model meta-features for each feature dimension on the correction model set, transforming the feature value distribution into a standard normal distribution with a mean of zero and a variance of one. The processed model meta-features constitute a standardized feature vector, with each temperature correction model corresponding to a standardized feature vector. The vector dimension is consistent with the number of model meta-features. The standardized feature vector retains the relative relationships of the original model meta-features while improving numerical stability.

[0115] The standardized feature vectors are paired with the accuracy index to form the detection training samples;

[0116] Specifically, standardized feature vectors are paired one-to-one with the accuracy indices of the corresponding temperature correction models. Each standardized feature vector serves as an input feature, and the corresponding accuracy index serves as the target label, forming the detection training samples. The pairing process ensures data consistency and avoids misalignment or omission. All temperature correction models in the correction model set participate in the pairing, forming a complete sample set. The detection training samples are organized in a unified format, with clear input feature and target label structures, suitable for supervised learning tasks. The sample set covers temperature correction models with different performance levels, ensuring that the quality detection model can learn the mapping relationship between model meta-features and accuracy indices. After the detection training samples are formed, they serve as training data for the quality detection model, driving the model to learn the correlation between model features and performance.

[0117] The quality detection model is trained using the test training samples under supervised learning. The trained quality detection model is then used to perform quality detection on the temperature correction model, and the quality judgment result is output.

[0118] The training samples are input into the quality detection model for supervised learning training. The training process optimizes the model's parameters by minimizing the error between the predicted accuracy index and the true accuracy index. In each training round, the quality detection model receives standardized feature vectors, performs feature fusion and nonlinear transformation on these vectors, outputs the predicted accuracy index, calculates the loss between the predicted and true accuracy indices, and updates the model's weights accordingly until convergence. After training, the quality detection model has the ability to predict the performance of temperature correction models based on their meta-features. The trained model is then used to perform quality checks on new or existing temperature correction models, outputting the corresponding accuracy index. The accuracy index is used to determine the quality of the temperature correction model, forming a quality assessment result. This result can be used for temperature correction model selection, deployment decisions, or further optimization guidance.

[0119] This embodiment also provides a computer device applicable to the temperature correction method for multi-meteorological factor model forecasts based on deep learning, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the temperature correction method for multi-meteorological factor model forecasts based on deep learning as proposed in the above embodiment.

[0120] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0121] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the deep learning-based multi-meteorological factor model forecast temperature correction method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0122] In summary, this invention achieves deep feature extraction and adaptive weight fusion of multiple meteorological factors through a dual mechanism of meteorological map structure construction and dynamic gating fusion. By integrating meteorological factors as factor nodes and constructing associated edge weights into a meteorological map structure, and utilizing the message passing mechanism of graph convolutional networks to aggregate neighborhood information, it achieves explicit modeling of complex coupling relationships among multiple factors, providing spatially correlated feature representations for temperature correction. The dynamic gating fusion network performs spatiotemporal convolution and global pooling on the feature tensors of multiple meteorological factors, achieving adaptive calibration of the contribution of multiple factors and improving the discriminative ability of feature fusion. Through an end-to-end network optimization and quality verification system, a high-precision and highly generalizable temperature correction model is formed.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for correcting temperature forecasts based on deep learning multi-meteorological factor models, characterized in that: include, Collect data on multiple meteorological factors and observe temperature field data, and perform preprocessing; Each meteorological factor in the multi-meteorological factor data is taken as a factor node, the coupling relationship between each meteorological factor is taken as a factor association edge, and the association edge is assigned a weight. The factor nodes, factor association edges, and association edge weights are integrated to output the meteorological map structure. The meteorological map structure is input into a graph convolutional network for deep feature extraction, and the output is a multi-meteorological factor feature tensor. A dynamic gating fusion network is used to analyze the global and local spatiotemporal characteristics of the feature tensor of multiple meteorological factors to generate dynamic weight coefficients. The dynamic weight coefficients and the feature tensor of multiple meteorological factors are then weighted and fused to output the fused feature tensor. The fused feature tensor is input into a fully connected neural network for nonlinear transformation and dimension mapping, outputting corrected temperature field data; the corrected temperature field data is compared with the observed temperature field data to output the temperature correction value; based on the temperature correction value, the network parameters of the graph convolutional network, the dynamic gated fusion network, and the fully connected neural network are optimized through the backpropagation algorithm to form a deep learning temperature correction model; Different model parameter perturbations are applied to the temperature correction model to output a set of correction models. The quality of each temperature correction model in the set is verified to generate an accuracy index. The accuracy index and the model meta-features of each temperature correction model are used as a training set and input into the quality detection model for training to judge the quality of each temperature correction model.

2. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The specific steps involve inputting the meteorological map structure into a convolutional network for deep feature extraction and outputting a multi-meteorological factor feature tensor. The meteorological map structure is input into a graph convolutional network, and the neighborhood information of the factor nodes is aggregated through the message passing mechanism of the graph convolutional network to output the node feature tensor. The node feature tensor is transformed and nonlinearly activated to output a multi-meteorological factor feature tensor.

3. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The step involves performing global and local spatiotemporal characteristic analysis on the feature tensors of multiple meteorological factors using a dynamic gated fusion network to generate dynamic weight coefficients. The specific steps are as follows: A dynamic gated fusion network is used to perform spatiotemporal convolution on the feature tensors of multiple meteorological factors to output local spatiotemporal features. Global pooling is performed on the feature tensors of multiple meteorological factors to output global statistical properties; Local spatiotemporal features and global statistical properties are concatenated to generate a comprehensive feature vector; The comprehensive feature vector is linearly transformed and normalized to generate dynamic weight coefficients.

4. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The specific steps for weighted fusion of dynamic weight coefficients and multi-meteorological factor feature tensors to output a fused feature tensor are as follows: The dynamic weight coefficients and the feature tensors of multiple meteorological factors are weighted by coefficients, and the weighted feature tensor is output. The weighted feature tensors are classified and aggregated along the feature dimensions of meteorological factors, and the aggregated feature tensors are output. Adjust the dimensions and representation of the aggregated feature tensor to output the fused feature tensor.

5. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The specific steps involve inputting the fused feature tensor into a fully connected neural network, performing nonlinear transformation and dimension mapping, and outputting corrected temperature field data. The fused feature tensor is input into a fully connected neural network. The first fully connected layer of the fully connected neural network performs a linear transformation on the fused feature tensor to generate an intermediate feature representation. The intermediate feature representation is nonlinearly transformed by the ReLU activation function to generate a nonlinear feature representation. The nonlinear feature representation is mapped to the temperature field dimension through a second fully connected layer, and the corrected temperature field data is output.

6. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The specific steps for comparing the corrected temperature field data with the observed temperature field data and outputting the temperature correction value are as follows: The corrected temperature field data is spatially and temporally aligned with the observed temperature field data; The aligned and corrected temperature field data are compared with the observed temperature field data element by element, and the temperature correction value is output.

7. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The process involves optimizing the network parameters of a graph convolutional network, a dynamically gated fusion network, and a fully connected neural network based on temperature correction values ​​using a backpropagation algorithm to form a deep learning temperature correction model. The specific steps are as follows: Perform root mean square error statistics on each temperature correction value and output the loss function value; The loss function value is used as the starting point of the backpropagation algorithm, and the gradient information of the graph convolutional network, dynamic gated fusion network and fully connected neural network corresponding to the loss function value is solved by the chain rule. The Adam optimizer is used to iteratively update the network parameters of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks according to gradient information, and the updated network parameters are output. The updated network parameters are loaded into the architecture of graph convolutional networks, dynamic gated fusion networks, and fully connected neural networks for optimization and adjustment, and then integrated to form a deep learning temperature correction model.

8. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The specific steps for perturbing the temperature correction model with different model parameters and outputting a correction model set are as follows: Perform parameter sensitivity analysis on the temperature correction model to determine the range of network parameters that need to be perturbed; Based on the range of network parameters, set multiple different combinations of parameter configurations; The temperature correction model is initialized by configuring parameters in different combinations, and the correction model set is output.

9. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The process of quality verification and accuracy index generation for each temperature correction model in the correction model set involves using multi-meteorological factor data and observed temperature field data as training and validation sets, respectively. The training set is used to perform temperature correction predictions on each temperature correction model in the correction model set. The results of the temperature correction predictions are compared with those of the validation set, and the results of the difference comparison are quantitatively analyzed to generate an accuracy index.

10. The temperature correction method for multi-meteorological factor model forecasts based on deep learning as described in claim 1, characterized in that: The steps involve using the accuracy index and the model meta-features of each temperature correction model as a training set, inputting them into the quality detection model for training, and judging the quality of each temperature correction model. A quality detection model was built based on the scikit-learn machine learning framework. The model meta-features of each temperature correction model are standardized to generate standardized feature vectors; The standardized feature vectors are paired with the accuracy index to form the detection training samples; The quality detection model is trained using the test training samples under supervised learning. The trained quality detection model is then used to perform quality detection on the temperature correction model, and the quality judgment result is output.

Images