A deep learning-based multi-meteorological factor mode temperature prediction correction method

Abstract

The application discloses a kind of based on deep learning multi-meteorological factor mode forecast temperature revision method, it is related to meteorological data processing technical field, including, acquisition multi-meteorological factor data and observation temperature field data, output meteorological chart structure;Meteorological chart structure is input graph convolution network and carries out depth feature extraction, outputs multi-meteorological factor feature tensor;By dynamic gate fusion network, global and local spatio-temporal characteristics analysis is carried out to multi-meteorological factor feature tensor, and output fusion feature tensor;Fusion feature tensor is input full connection neural network and carries out nonlinear transformation and dimension mapping, forms the temperature revision model of deep learning;Different model parameter disturbance is carried out to temperature revision model, and accuracy index is generated;Input to quality detection model and train, judge the quality of each temperature revision model.The application is realized by meteorological chart structure construction and dynamic gate fusion double mechanism cooperation, and multi-meteorological factor depth feature extraction and adaptive weight fusion.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological data processing, and in particular to a multi-meteorological factor mode forecast temperature correction method based on deep learning. BACKGROUND

[0002] In recent years, meteorological forecast temperature correction technology has gradually evolved towards multi-factor synergy and deep learning integration. The mainstream scheme is based on multi-meteorological factor (such as pressure, humidity, wind field) data output by numerical mode, and uses convolutional neural network (CNN) or long short-term memory network (LSTM) for feature extraction and error correction. Among them, the multi-source data fusion framework gradually introduces an attention mechanism to optimize factor weight distribution, graph neural network (GNN) is preliminarily applied to spatial relationship modeling of meteorological elements, and generative adversarial network (GAN) is tried to be used for temperature field distribution reconstruction. The end-to-end training paradigm significantly improves the correction efficiency. Deep learning models gradually replace traditional statistical correction methods through nonlinear mapping capabilities, providing a technical foundation for high-precision temperature prediction.

[0003] However, the existing technology has limitations in factor coupling mechanism modeling and model quality efficient evaluation. Traditional multi-factor processing methods mostly rely on static feature splicing or simple weighted fusion, and fail to explicitly construct the physical correlation topology between meteorological factors, resulting in insufficient learning ability of the model for the nonlinear interaction between factors, lack of adaptive adjustment mechanism for factor contribution degree in dynamic weather scenarios, and complete dependence on computationally intensive post-validation for model evaluation, lack of rapid quality screening paradigm based on meta-features, making it difficult to balance the timeliness and model optimization reliability of business deployment. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the present application provides a multi-meteorological factor mode forecast temperature correction method based on deep learning to solve the problem of limitations in factor coupling mechanism modeling and model quality efficient evaluation.

[0006] To solve the above technical problems, the present application provides the following technical solutions:

[0007] The application provides a multi-weather factor mode forecast temperature correction method based on deep learning, which comprises the following steps: collecting multi-weather factor data and observation temperature field data and performing pretreatment; regarding each weather factor in the multi-weather factor data as a factor node, regarding the coupling relationship between each weather factor as a factor correlation edge, and giving the factor correlation edge a correlation edge weight; integrating the factor node, the factor correlation edge and the correlation edge weight, and outputting a weather map structure; inputting the weather map structure into a graph convolution network to perform deep feature extraction, and outputting a multi-weather factor feature tensor; performing global and local spatiotemporal characteristic analysis on the multi-weather factor feature tensor through a dynamic gating fusion network, and generating a dynamic weight coefficient; performing weighted fusion on the dynamic weight coefficient and the multi-weather factor feature tensor, and outputting a fusion feature tensor; inputting the fusion feature tensor into a fully connected neural network to perform nonlinear transformation and dimension mapping, and outputting corrected temperature field data; comparing the corrected temperature field data with the observation temperature field data, and outputting a temperature correction value; based on the temperature correction value, performing network parameter optimization on the graph convolution network, the dynamic gating fusion network and the fully connected neural network through a back propagation algorithm, and forming a deep learning temperature correction model; performing different model parameter perturbations on the temperature correction model, and outputting a correction model set; performing quality verification on each temperature correction model in the correction model set, and generating an accuracy index; inputting the accuracy index and the model feature of each temperature correction model as a training set into a quality detection model for training, and judging the quality of each temperature correction model.

[0008] As a preferred scheme of the multi-weather factor mode forecast temperature correction method based on deep learning, wherein: the step of inputting the weather map structure into the graph convolution network to perform deep feature extraction and output a multi-weather factor feature tensor comprises the following steps:

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

[0010] The node feature tensor is subjected to feature transformation and nonlinear activation, and a multi-weather factor feature tensor is output.

[0011] As a preferred scheme of the multi-weather factor mode forecast temperature correction method based on deep learning, wherein: the step of performing global and local spatiotemporal characteristic analysis on the multi-weather factor feature tensor through the dynamic gating fusion network and generating a dynamic weight coefficient comprises the following steps:

[0012] The multi-weather factor feature tensor is subjected to spatiotemporal convolution through the dynamic gating fusion network, and a local spatiotemporal feature is output.

[0013] The multi-weather factor feature tensor is subjected to global pooling, and a global statistical characteristic is output.

[0014] The local space-time features and the global statistical characteristics are spliced to generate a comprehensive feature vector;

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

[0016] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor mode prediction temperature of the application, wherein: the dynamic weight coefficient and the multi-weather factor feature tensor are weighted and fused to output a fused feature tensor, and the specific steps are,

[0017] The dynamic weight coefficient and the multi-weather factor feature tensor are weighted and fused to output a fused feature tensor;

[0018] The weighted feature tensor is classified and aggregated along the feature dimension of the weather factor to output an aggregated feature tensor;

[0019] The aggregated feature tensor is adjusted in dimension and representation to output a fused feature tensor.

[0020] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor mode prediction temperature of the application, wherein: the fused feature tensor is input into a fully connected neural network for nonlinear transformation and dimension mapping to output corrected temperature field data, and the specific steps are,

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

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

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

[0024] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor mode prediction temperature of the application, wherein: the corrected temperature field data is compared with the observed temperature field data to output a temperature correction value, and the specific steps are,

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

[0026] The aligned corrected temperature field data and the observed temperature field data are element-by-element difference statistics to output a temperature correction value.

[0027] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor model prediction temperature of the application, wherein: the temperature correction value is used as the starting point of the back propagation algorithm, and the gradient information of the graph convolution network, the dynamic gate fusion network and the fully connected neural network corresponding to the loss function value is solved by the chain rule;

[0028] The mean square error of each temperature correction value is calculated, and the loss function value is outputted;

[0029] The loss function value is used as the starting point of the back propagation algorithm, and the gradient information of the graph convolution network, the dynamic gate fusion network and the fully connected neural network corresponding to the loss function value is solved by the chain rule;

[0030] The network parameters of the graph convolution network, the dynamic gate fusion network and the fully connected neural network are iteratively updated according to the gradient information by using the Adam optimizer, and the updated network parameters are outputted;

[0031] The updated network parameters are loaded into the architecture of the graph convolution network, the dynamic gate fusion network and the fully connected neural network for optimization and adjustment, and a deep learning temperature correction model is formed by integration.

[0032] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor model prediction temperature of the application, wherein: the temperature correction model is disturbed by different model parameters, and a correction model set is outputted, and the specific steps are as follows,

[0033] The parameter sensitivity of the temperature correction model is analyzed to determine the network parameter range that needs to be disturbed;

[0034] Based on the network parameter range, a plurality of different parameter configuration combinations are set;

[0035] The temperature correction model is initialized by the parameter configuration combination respectively, and a correction model set is outputted.

[0036] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor model prediction temperature of the application, wherein: the quality of each temperature correction model in the correction model set is verified, and the accuracy index is generated by using the multi-weather factor data and the observed temperature field data as the training set and the verification set respectively, using the training set to correct the temperature of each temperature correction model in the correction model set, comparing the results of the temperature correction prediction with the verification set, and quantitatively analyzing the difference value comparison result to generate the accuracy index.

[0037] As a preferred scheme of the temperature correction method based on the deep learning-based multi-weather factor model prediction of the application, wherein: the accuracy index and the model element characteristics of each temperature correction model are input into the quality detection model for training to determine the quality of each temperature correction model, and the specific steps are,

[0038] The quality detection model is constructed based on the scikit-learn machine learning framework;

[0039] The model element characteristics of each temperature correction model are standardized to generate a standardized feature vector;

[0040] The standardized feature vector and the accuracy index are paired with feature labels to form a detection training sample;

[0041] The quality detection model is supervised learning trained by using the detection training sample, and the quality detection model is used to detect the quality of the temperature correction model, and the quality judgment result is output.

[0042] The application has the following advantages: through the cooperation of meteorological map structure construction and dynamic gate fusion double mechanism, the deep feature extraction and adaptive weight fusion of multi-weather factors are realized. The meteorological factors are taken as factor nodes and the associated edge weight is integrated into a meteorological map structure, the message passing mechanism of the graph convolution network is used to aggregate the neighborhood information, the explicit modeling of the complex coupling relationship between multiple factors is realized, the feature representation rich in spatial correlation is provided for temperature correction, the dynamic gate fusion network performs spatio-temporal convolution and global pooling on the multi-weather factor feature tensor, realizes the adaptive calibration of the contribution degree of multiple factors, improves the discriminant ability of feature fusion, and through the end-to-end network optimization and quality verification system, a high-precision and generalization-capable temperature correction model is formed. BRIEF DESCRIPTION OF DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0044] Fig. 1 The flowchart of the temperature correction method based on the deep learning-based multi-weather factor model prediction.

[0045] Fig. 2 The flowchart of the deep feature extraction of the graph convolution network.

[0046] Fig. 3 The flowchart of the feature fusion of the dynamic gate fusion network.

[0047] Fig. 4 Flow chart for quality verification and quality detection model training. DETAILED DESCRIPTION

[0048] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0049] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail as not to unnecessarily obscure aspects of the present application.

[0050] Secondly, the "one embodiment" or "an embodiment" referred to herein means that a specific feature, structure, or characteristic described in connection with that embodiment can be included in at least one implementation of the present application. The appearances of the phrase "in one embodiment" or "in an embodiment" in various places in the specification are not necessarily referring to the same embodiment, nor are separate or alternative embodiments mutually exclusive of other embodiments.

[0051] Reference Signs List Figs. 1-4 For one embodiment of the present application, the embodiment provides a deep learning-based multi-meteorological factor mode temperature correction method, comprising the following steps:

[0052] S1, collect multi-meteorological factor data and observed 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 correlation edge, and the factor correlation edge is given a correlation edge weight; the factor node, the factor correlation edge and the correlation edge weight are integrated, and a meteorological graph structure is output.

[0053] Specifically, from various data sources such as numerical weather prediction models, ground meteorological observation stations, and satellite remote sensing, multi-meteorological factor data including temperature, humidity, air pressure, wind speed, and cloud cover are collected, and actual observed temperature field data in the corresponding spatio-temporal range are also collected. The corresponding spatio-temporal range is the same time period and the same geographical area covered by the multi-meteorological factor data output by the exponential numerical weather prediction model and the actual observed temperature field data. The collected multi-meteorological factor data and observed temperature field data are subjected to data cleaning and normalization processing. Data cleaning processing includes checking and processing missing values and abnormal values that are obviously beyond the physical reasonable range in the multi-meteorological factor data. Being obviously beyond the physical reasonable range means that the meteorological factor values violate basic atmospheric physics common sense or historical statistical rules, such as ground temperature exceeding ±60℃, relative humidity greater than 100% or negative, sea level air pressure lower than 870 hundred pascals or higher than 1080 hundred pascals, etc. Normalization processing maps the numerical range of all multi-meteorological factor data and observed temperature field data into the interval of zero to one through linear transformation, eliminating the differences in dimension and numerical range between different meteorological factors, and providing standardized data input for subsequent calculations. The preprocessing process outputs clean and uniformly numerical standardized multi-meteorological factor data and standardized observed temperature field data.

[0054] Each meteorological factor in the standardized multi-meteorological factor data obtained after preprocessing, such as temperature, humidity, and air pressure, is defined as an independent factor node. Each factor node is mathematically represented using a factor feature vector, which contains the standardized values of the meteorological factor at multiple spatial locations and multiple time steps. The Pearson correlation coefficient between the factor feature vectors of any two different meteorological factors is calculated by measuring the consistency of the change trend 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 two meteorological factors. Each statistical coupling relationship is defined as a factor association edge connecting two factor nodes. Each factor association edge is assigned an association edge weight, and the specific value of the association edge weight 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, factor association edges, and association edge weights corresponding to each factor association edge are integrated. The integration is done according to the mathematical specifications of graph theory, with factor nodes as vertices and factor association edges as edges connecting vertices, and association edge weights as attributes of corresponding edges, thereby constructing a complete and weighted meteorological graph structure. The meteorological graph structure completely expresses the coupling relationship network between all individual meteorological factors and meteorological factor groups in the multi-meteorological factor data.

[0056] S2, input the weather map structure into a graph convolution network to perform deep feature extraction, and output a multi-weather factor feature tensor.

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

[0058] Specifically, for each factor node, the graph convolution network traverses all factor association edges connected to the factor node, collects factor feature vectors of neighborhood factor nodes, and performs weighted aggregation on neighborhood information according to the corresponding association edge weight; the weighted aggregation process is realized by linear combination of the factor feature vectors of the neighborhood factor nodes, and the weight value is the association edge weight of the factor association edge, which emphasizes the contribution of neighborhood factor nodes with strong coupling relationship to the current factor node; the information after weighted aggregation is combined with the factor feature vector of the current factor node to generate an updated node feature representation; the graph convolution network performs message passing and information aggregation through multi-layer convolution iteration, further fusing neighborhood information in a wider range in each iteration, thereby capturing the complex spatial and association dependency relationships between weather factors; after each convolution operation, an intermediate representation containing the updated features of all factor nodes is generated, and after processing by all convolution layers, a node feature tensor is output; the node feature tensor retains the feature information of each factor node in the spatial position and time step, while incorporating the coupling relationship between the weather factor group, providing rich deep feature representation for subsequent feature transformation;

[0059] Further, the construction of the graph convolution network is based on the topological relationship of the weather map structure, and the network architecture is sequentially connected by multiple graph convolution layers, each graph convolution layer of the graph convolution 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 the neighborhood factor nodes according to the factor association edges and the association edge weights in the input weather map structure, and the feature transformation mechanism performs dimension mapping and nonlinear transformation on the aggregated node features through a nonlinear activation function; the multiple graph convolution layers iteratively perform neighborhood information aggregation and feature transformation, gradually expanding the feature perception of each factor node, thereby deeply mining the complex spatial coupling and nonlinear dependency relationships between multiple weather factors, outputting a node feature tensor that integrates global neighborhood information as a result of deep feature extraction of multiple weather factors.

[0060] Performing feature transformation and nonlinear activation on the node feature tensor, and outputting a multi-weather factor feature tensor;

[0061] Specifically, the feature transformation is performed on the node feature tensor, and the dimension adjustment and feature reconstruction are performed on the factor feature vector of each factor node through linear transformation; the linear transformation is realized through matrix multiplication, each factor feature vector of the node feature tensor is mapped to the feature space, the feature expression ability is enhanced, and the requirements of subsequent processing are adapted; the node feature tensor after linear transformation still retains the feature information of all factor nodes, a nonlinear activation function is applied to the node feature tensor after linear transformation, a ReLU activation function is used to perform nonlinear processing on each element of each factor feature vector, that is, the ReLU activation function sets the negative value in the factor feature vector to zero and keeps the positive value unchanged, thereby introducing nonlinear characteristics and enhancing the modeling ability of the node feature tensor on complex meteorological factor relationships; the node feature tensor after nonlinear activation integrates the transformed features of each factor node to form a more expressive feature representation, the node feature tensor after feature transformation and nonlinear activation processing is arranged in a unified format, and a multi-meteorological factor feature tensor is output; the multi-meteorological factor feature tensor contains the deep features of all meteorological factors in the spatial position and time step, and fuses the coupling relationship and nonlinear characteristics between factors, thereby providing standardized input data for subsequent dynamic gate fusion network spatiotemporal characteristic analysis.

[0062] S3, performing global and local spatiotemporal characteristic analysis on the multi-meteorological factor feature tensor through the dynamic gate fusion network to generate a dynamic weight coefficient; performing weighted fusion on the dynamic weight coefficient and the multi-meteorological factor feature tensor to output a fusion feature tensor.

[0063] Performing spatiotemporal convolution on the multi-meteorological factor feature tensor through the dynamic gate fusion network to output a local spatiotemporal feature;

[0064] Specifically, a dynamic gated fusion network is constructed based on a deep learning framework, and a multi-meteorological factor feature tensor and an observed temperature field data are used as training data for dynamic gated fusion network training; the dynamic gated fusion network performs spatio-temporal convolution on the multi-meteorological factor feature tensor to extract local spatio-temporal features; the spatio-temporal convolution captures the dynamic changes of the multi-meteorological factor feature tensor in the time series and the local patterns in the spatial grid by applying convolution kernels in the time dimension and the spatial dimension, respectively; specifically, the spatio-temporal 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 position to capture the local correlation between adjacent spatial grid points; the size and stride of the convolution kernel are set according to the time and spatial resolution of the multi-meteorological factor feature tensor to ensure that sufficient time and spatial range is covered and local spatio-temporal patterns are fully mined; the spatio-temporal convolution generates an intermediate feature representation reflecting local spatio-temporal dynamics by weighted summation of each feature vector of the multi-meteorological factor feature tensor; the dynamic gated fusion network further introduces a gating mechanism to selectively filter the intermediate feature representation of the convolution operation and enhance the attention to key spatio-temporal features; the gating mechanism retains the local features that have a greater contribution to temperature correction by obtaining the weight of each feature vector, while suppressing the influence of irrelevant or noise features; the local spatio-temporal features are output after spatio-temporal convolution and gating mechanism processing.

[0065] Global pooling is performed on the multi-meteorological factor feature tensor to output global statistical characteristics;

[0066] Specifically, global pooling is performed by aggregating all feature vectors of the multi-meteorological factor feature tensor in the spatial and temporal dimensions to generate global statistical characteristics; the aggregation adopts an average pooling method to calculate the average value of the feature vector of each meteorological factor at all spatial positions and time steps to generate a statistical feature vector reflecting the global mean; a variance pooling method is used to calculate the variance of the feature vector of each meteorological factor at all spatial positions and time steps to generate a statistical feature vector reflecting the global variability; the mean and variance represent the central tendency and dispersion of the multi-meteorological factor feature tensor in the entire spatio-temporal range, respectively, and together constitute the global statistical characteristics; the global pooling operation compresses the high-dimensional data of the multi-meteorological factor feature tensor into low-dimensional statistical feature representation through dimension reduction processing; after global pooling processing, the global statistical characteristics are output.

[0067] The local spatio-temporal features and global statistical characteristics are concatenated to generate a comprehensive feature vector;

[0068] Specifically, the feature vectors of the local spatio-temporal features and the feature vectors of the global statistical characteristics are sequentially connected in the order of the feature vectors of the local spatio-temporal features first and the feature vectors of the global statistical characteristics last to form a comprehensive feature vector; the splicing process ensures that the dimensions of the local spatio-temporal features and the global statistical characteristics are aligned, and by adjusting the dimensions of the local spatio-temporal features and the global statistical characteristics before splicing, the local spatio-temporal features and the global statistical characteristics have the same feature length; the comprehensive feature vector after splicing integrates the dynamic details of the local spatio-temporal features and the overall law of the global statistical characteristics, forming a comprehensive representation of the multi-weather factor feature tensor;

[0069] The linear transformation is initialized according to the time and space characteristics of the multi-weather factor feature tensor to ensure that the comprehensive feature vector after the linear transformation can highlight the most important feature components for temperature correction; the comprehensive feature vector after the linear transformation retains the core information of the comprehensive feature vector, and the normalized comprehensive feature vector is obtained by processing each element of the comprehensive feature vector using a softmax function; the softmax function maps the values of the comprehensive feature vector to the interval of 0 to 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 proportion of subsequent feature fusion; the normalization operation enhances the stability and interpretability of the weight coefficients.

[0070] The linear transformation is initialized according to the time and space characteristics of the multi-weather factor feature tensor to ensure that the comprehensive feature vector after the linear transformation can highlight the most important feature components for temperature correction; the comprehensive feature vector after the linear transformation retains the core information of the comprehensive feature vector, and the normalized comprehensive feature vector is obtained by processing each element of the comprehensive feature vector using a softmax function; the softmax function maps the values of the comprehensive feature vector to the interval of 0 to 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 proportion of subsequent feature fusion; the normalization operation enhances the stability and interpretability of the weight coefficients.

[0071] The dynamic weight coefficients and the multi-weather factor feature tensor are coefficient-weighted to output a weighted feature tensor;

[0072] Specifically, the feature dimensions of the dynamic weight coefficients and the multi-weather factor feature tensor are aligned through a dimension repetition method to ensure that each weight value corresponds to a feature component; the weighting process is implemented through dot multiplication to apply the dynamic weight coefficients to each feature vector of the multi-weather factor feature tensor to generate a weighted feature tensor after weighting; the coefficient weighting retains the original structure of the multi-weather factor feature tensor and only adjusts the size of the feature values to reflect the importance of the dynamic weight coefficients;

[0073] The weighted feature tensor calculation formula is,

[0074] ;

[0075] wherein, represents the weighted feature tensor after the coefficient weighting of the multi-weather factor feature tensor, a dynamic weight coefficient, a multi-weather factor feature tensor, an element-level multiplication operator.

[0076] The weighted feature tensor is classified and aggregated along the feature dimension of the weather factor, and an aggregated feature tensor is output.

[0077] Specifically, according to the type of the weather factor (such as temperature, humidity, and air pressure, etc.), the weighted feature tensor is grouped according to the factor category, and each group corresponds to a set of feature vectors of the weather factor. Each set of feature vectors is aggregated, and the same weather factor at all spatial positions and time steps is fused and summarized by weighted summation. The weight of the fusion and summary is determined according to the value of each feature component in the weighted feature tensor, so as to ensure that the aggregation result reflects the contribution of the key features. After the classification and aggregation are completed, an aggregated feature vector is generated for each group of weather factors, and the aggregated feature vector contains the comprehensive feature representation of the weather factor in the spatio-temporal range. The aggregated feature vectors of all weather factors are sorted to generate an aggregated feature tensor.

[0078] The aggregated feature tensor is dimensionally and representationally adjusted, and a fused feature tensor is output.

[0079] Specifically, the dimension of the aggregated feature tensor is mapped by linear transformation, and the feature dimension of the aggregated feature tensor is adjusted to a fixed size to match the input dimension requirement of the fully connected neural network. The representation of the aggregated feature tensor after dimension adjustment is adjusted, and the mean and variance of the feature vector in each aggregated feature tensor are standardized by normalization operation. The feature values are rescaled to a distribution with mean zero and variance one, enhancing the stability and generalization ability of the features. The feature tensor after dimension adjustment and representation adjustment is arranged in a unified format, and a fused feature tensor is output. The fused feature tensor contains the optimized feature representation of all weather factors, and the dimension is fixed and the representation is stable.

[0080] S4, input the fused feature tensor into the fully connected neural network for nonlinear transformation and dimension mapping, and output the corrected temperature field data; compare the corrected temperature field data with the observed temperature field data, and output the temperature correction value; based on the temperature correction value, optimize the network parameters of the graph convolution network, the dynamic gate fusion network and the fully connected neural network by the back propagation algorithm, and form a deep learning temperature correction model.

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

[0082] Specifically, the first fully connected layer of the full connection neural network receives the fusion feature tensor, performs linear transformation, and maps each feature vector in the fusion 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 fusion feature tensor during initialization, ensuring that the input features can be effectively converted. The linear transformation performs row linear combination on all feature components in the fusion feature tensor, preserving the correlation structure between the original features while enhancing the expression ability 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 fusion feature tensor and can better support the subsequent nonlinear modeling process.

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

[0084] The ReLU activation function keeps the positive value elements in the intermediate feature representation unchanged and sets the negative value elements to zero, breaking the linear constraint brought by the linear transformation and introducing nonlinear characteristics, enabling the full connection neural network to fit complex meteorological relationships. The nonlinear transformation process filters and enhances the feature responses that contribute to temperature correction without changing the dimension of the intermediate feature representation, suppressing 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 form of the intermediate feature representation in structure, but the numerical distribution is more concentrated in the positive value area, enhancing the sparsity and discriminability of the features. The nonlinear feature representation, as the output of the nonlinear transformation, provides input data with nonlinear expression ability for subsequent dimension mapping.

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

[0086] Specifically, the nonlinear feature representation is sent to the second fully connected layer, which performs temperature field dimension mapping to convert the nonlinear feature representation from the current feature space to the output space consistent with the target temperature field. The temperature field dimension mapping is realized through linear transformation. The weight and bias parameters used in the transformation process are set according to the number of spatial grids and time steps of the target temperature field, ensuring that the output dimension completely matches 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 position and time step. All temperature prediction values are arranged in the original spatio-temporal 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, the spatial alignment is achieved by matching the geographic coordinate grid of the corrected temperature field data and the observed temperature field data, ensuring that each spatial position in the corrected temperature field data corresponds to the corresponding position in the observed temperature field data accurately, and the resolution and coverage of the corrected temperature field data and the observed temperature field data are not changed during the alignment process; the time alignment is completed by synchronizing the time step sequence of the corrected temperature field data and the 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, and the start time, end time and time interval of the time sequence are completely matched; after the alignment operation is completed, the corrected temperature field data and the observed temperature field data are corresponded point by point on the spatial grid and the time sequence, forming a structured data pair.

[0089] The aligned corrected temperature field data and the observed temperature field data are element-by-element difference statistics, and the temperature correction value is output;

[0090] Specifically, the difference statistics are 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 difference results; the difference result reflects the deviation between the predicted output and the real observation, the positive value indicates that the corrected temperature is high, the negative value indicates that the corrected temperature is low, and the absolute value of the difference indicates the degree of deviation; all differences are organized in the order of the observed temperature field data, forming the temperature correction value;

[0091] The mean square error statistics of each temperature correction value are output, and the loss function value is output;

[0092] The mean square error statistics are a global error evaluation of the temperature correction value, and the square operation is performed on all temperature correction values in the statistical process, eliminating the mutual offset of positive and negative differences and strengthening the influence of large errors; the average value of the squared difference is calculated at all spatial positions and time steps, and the average value calculation covers all data points in the entire spatio-temporal range, ensuring the comprehensiveness of the error evaluation; the mean square error statistics result is the loss function value, which reflects the overall prediction accuracy, and the smaller the value is, the better the prediction correction effect is; the loss function value is used to guide the adjustment direction of the network parameters, and the size of the loss function value directly affects the generation of gradient information and the amplitude of parameter update;

[0093] The loss function value calculation formula is,

[0094] ;

[0095] Wherein, represents the loss function value of the mean square error statistics of each temperature correction value, represents the number of temperature correction values at all spatial positions and time steps, represents the index of the temperature correction value, represents the first temperature correction value, a temperature correction value.

[0096] The loss function value is taken as the starting point of the back propagation algorithm, and the gradient information of the graph convolution network, the dynamic gate fusion network and the fully connected neural network corresponding to the loss function value is solved by the chain rule;

[0097] Specifically, the back propagation algorithm takes the loss function value as the starting point, and the error signal is transmitted back layer by layer from the output end to the input end. The chain rule is used to decompose the partial derivative of the loss function value to each layer network parameter in the graph convolution network, the dynamic gate fusion network and the fully connected neural network, and the gradient of the loss function value to the second fully connected layer parameter is counted. Then the gradients of the non-linear feature representation, the intermediate feature representation, the fusion feature tensor, the aggregated feature tensor, the weighted feature tensor, the multi-meteorological factor feature tensor and the node feature tensor are counted in turn, until the initial parameters of the graph convolution network are traced back. The gradient information reflects the sensitivity of the loss function value to each network parameter, and reflects the direction and amplitude of network parameter update. The gradient solving process is strictly executed in reverse according to the forward propagation path of the network, ensuring that the partial derivative of each step corresponds to the forward operation, and the gradient information contains the partial derivative of all trainable parameters in the graph convolution network, the dynamic gate fusion network and the fully connected neural network, providing a basis for subsequent parameter update.

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

[0099] Specifically, the Adam optimizer receives the gradient information and performs gradient optimization on the network parameters of the graph convolution network, the dynamic gate fusion network and the fully connected neural network according to the gradient information. The gradient optimization process updates the convolution weights in the graph convolution network, the convolution kernel parameters and the gate weights in the dynamic gate fusion network, and the connection weights and bias terms in the fully connected neural network, respectively. The update operation is performed in each training iteration step, and the parameter adjustment direction is opposite to the gradient direction. The adjustment amplitude is determined by the learning rate and the gradient size. For example, when the gradient value of the fully connected layer weight is large, the Adam optimizer will automatically reduce the learning rate of the network parameter based on the moving average of the historical gradient square, to prevent the update step from being too large and causing oscillation. For network parameters with small gradients, the learning rate is increased to speed up the convergence. The Adam optimizer balances the update speed of different parameters through the adaptive mechanism, avoids the problem of gradient explosion or disappearance, and improves the convergence stability. The iterative update process continues until the loss function value converges, and the updated network parameters are output.

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

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

[0102] S5, different model parameter perturbations are performed on the temperature correction model to output a set of correction models; the quality of each temperature correction model in the set of correction models is verified to generate an accuracy index; the accuracy index and the model feature of each temperature correction model are taken as a training set and input to a quality detection model for training to determine the quality of each temperature correction model.

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

[0104] Specifically, the parameter sensitivity analysis process is performed by perturbing a single or local network parameter in the temperature correction model and observing the deviation of the output corrected temperature field data in the temperature correction model in the spatial distribution form, regional extreme value change and time series trend to evaluate the influence degree of the single or local network parameter in the temperature correction model on the prediction result. The size of the influence degree is judged according to whether the overall structure of the temperature field is significantly distorted, whether the regional temperature gradient is fluctuated, and whether the time continuity is destroyed. If the perturbation leads to temperature deviation or mutation of the temperature field, it is considered that the single or local network parameter has high influence degree and belongs to a sensitive parameter. If the output changes gently and maintains the original spatio-temporal pattern, it is considered that the influence degree is low. The sensitivity of the network parameter is judged according to the significance of the influence degree. The network parameter with high sensitivity causes large fluctuations in the temperature correction value after perturbation, and the network parameter with low sensitivity has less influence after perturbation. Based on the analysis results of the parameter sensitivity analysis, the sensitive parameter categories that affect the output of the temperature correction model are screened, and the perturbation range of the network parameter is determined. The perturbation range is set with the network parameter value as the center and the upper and lower floating boundaries are set. For example, the convolution weight corresponding to the factor correlation edge with strong coupling relationship in the graph convolution network is set to a perturbation interval of ±15%, the gating weight in the dynamic gating fusion network is set to a perturbation interval of ±20%, and the connection weight of the second full connection layer in the full connection neural network is set to a perturbation interval of ±10%, so as to ensure that the perturbed parameters are still within the interval of physical interpretability and numerical stability. After the parameter sensitivity analysis is completed, the network parameter range that needs to be perturbed is output.

[0105] Based on the network parameter range, a plurality of different parameter configuration combinations are set.

[0106] Specifically, the network parameter range is used as the basis for parameter configuration to generate a plurality of different parameter configuration combinations. Each parameter configuration combination selects a plurality of network parameters from the network parameter range, randomly samples or divides according to a regular grid according to a preset perturbation range, and generates a plurality of different parameter configuration combinations. The preset perturbation range is set according to the results of the parameter sensitivity analysis, combined with the importance of the network layer where the network parameter is located, the historical statistics of the gradient amplitude and the numerical stability requirement, and the numerical value of the network parameter is set as the center to set the upper and lower floating percentage interval.

[0107] The parameter configuration combinations are used to initialize the model copies of the temperature correction model respectively, and the corrected model set is output.

[0108] Each parameter configuration combination is used to initialize a temperature correction model copy of the temperature correction model, and the initialization process keeps the network architecture of the temperature correction model unchanged, and only replaces the network parameters to be disturbed with specific values in the corresponding parameter configuration combination, and the remaining undisturbed parameters remain the original trained values; each temperature correction model copy inherits the complete structure and forward inference process of the temperature correction model, but shows different prediction characteristics due to parameter differences; all parameter configuration combinations are applied in turn to generate multiple temperature correction model copies with parameter differences, and each temperature correction model copy can independently perform a temperature correction task; after the initialization of all temperature correction model copies is completed, the temperature correction model copies are organized into a correction model set, and the correction model set includes multiple temperature correction models after parameter disturbance, and each temperature correction model corresponds to a unique set of parameter configurations.

[0109] The quality of each temperature correction model in the correction model set is verified, and the accuracy index is generated by taking the multi-weather factor data and the observed temperature field data as the training set and the verification set respectively, using the training set to perform temperature correction prediction on each temperature correction model in the correction model set, comparing the results of the temperature correction prediction with the verification set, and quantitatively analyzing the difference comparison results. Quantitative analysis refers to evaluating the distribution characteristics of the deviation between the prediction output of each temperature correction model and the observed temperature field data in space and time, judging the degree of deviation from the true value, i.e. the deviation is small and the distribution is concentrated, indicating that the correction result is close to the actual observation result. According to the degree of deviation, a corresponding accuracy index is given, and the higher the accuracy index, the better the temperature correction model correction effect;

[0110] A quality detection model is constructed based on the scikit-learn machine learning framework;

[0111] Specifically, the quality detection model is constructed under the scikit-learn machine learning framework, and a support vector machine or a random forest is selected as a basic learning algorithm. The input of the quality detection model is the model element features of the temperature correction model, and the output is the corresponding quality judgment result. The quality detection model construction process includes data preparation, algorithm selection, hyperparameter setting and training process configuration. In the data preparation stage, the structure information and training process information of each temperature correction model in the correction model set are collected, and quantifiable model element features such as the number of network layers, the total number of parameters, the number of training rounds and the loss function descent rate are extracted. The basic learning algorithm selection determines the classification or regression mode according to the task characteristics, and the structure and training process of the quality detection model are defined within the scikit-learn framework.

[0112] The model element features of each temperature correction model are standardized to generate a standardized feature vector;

[0113] Specifically, the model meta feature of the temperature correction model refers to a quantifiable index describing the structural properties and dynamic training process of the temperature correction model, used to represent the intrinsic characteristics of the temperature correction model. The model meta feature of the temperature correction model is obtained by analyzing the network architecture configuration and training log information of each temperature correction model in the correction model set, and specifically includes the number of network layers, the total number of parameters, the type of activation function, the learning rate setting, the number of training rounds, and the gradient change amplitude, etc.

[0114] Further, the model meta features of the temperature correction models differ in numerical value and dimension, and to eliminate the influence of the differences on the quality detection model training, the model meta features of the temperature correction models need to be standardized. The standardization processing adopts the Z-score method, and the mean and standard deviation of the model meta features on the correction model set are calculated for each feature dimension, so that the feature value distribution is converted 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, and each temperature correction model corresponds to a standardized feature vector. The vector dimension is consistent with the number of model meta features, and the standardized feature vector preserves the relative relationship of the original model meta features while improving the numerical stability.

[0115] The standardized feature vector is paired with the accuracy index as a feature label to form a detection training sample.

[0116] Specifically, the standardized feature vector is paired with the accuracy index of the corresponding temperature correction model on a one-to-one basis. Each standardized feature vector is used as an input feature, and the corresponding accuracy index is used as a target label to form a detection training sample. The pairing process ensures data consistency and avoids misplacement or omission. All temperature correction models in the correction model set participate in the pairing to form a complete sample set. The detection training sample is organized in a unified format, and the input feature and target label structure are clear, which is 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 the model meta features and the accuracy index. After the detection training sample is formed, it is used as the training data of the quality detection model to drive the model to learn the correlation between the model features and the performance.

[0117] The quality detection model is trained using the detection training sample through supervised learning, and the trained quality detection model is used to detect the quality of the temperature correction model and output a quality judgment result.

[0118] The detection training sample is input into the quality detection model, supervised learning training is performed, and the parameter of the quality detection model is optimized by minimizing the error between the predicted accuracy index and the true accuracy index in the training process; the quality detection model receives the standardized feature vector in each round of training, performs feature fusion and nonlinear transformation on the standardized feature vector, outputs the predicted accuracy index, calculates the loss value of the predicted accuracy index and the true accuracy index, and updates the weight of the quality detection model in reverse, until the quality detection model converges; after the training is completed, the quality detection model has the ability to predict the temperature correction model performance according to the model meta-feature, uses the trained quality detection model to detect the quality of a new or existing temperature correction model, outputs the corresponding accuracy index, judges the quality of the temperature correction model according to the accuracy index, forms a quality judgment result, and the quality judgment result can be used for temperature correction model screening, deployment decision or further optimization guidance.

[0119] The embodiment also provides a computer device suitable for the deep learning-based multi-weather factor mode prediction temperature correction method, which comprises 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 realize the deep learning-based multi-weather factor mode prediction temperature correction method proposed in the above embodiment.

[0120] The computer device can be a terminal, which comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be realized through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, the input device can be an external keyboard, touchpad or mouse, etc.

[0121] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement the method for correcting temperature prediction of a multi-weather-factor model based on deep learning proposed in the above embodiment; and the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0122] To sum up, the application realizes deep feature extraction and adaptive weight fusion of multi-weather factors by the fusion of the two mechanisms of meteorological map structure construction and dynamic gating.

[0123] It should be noted that the above embodiments are only used to illustrate the technical solutions of the application but not limit the application. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced equivalently without departing from the spirit and scope of the application, and all modifications or equivalent replacements should be covered in the scope of the claims of the application.

Claims

1. A method for temperature correction based on deep learning and multi-meteorological factor mode prediction, characterized in that: The application relates to a temperature correction method based on a deep learning model. Collecting and preprocessing multi-weather factor data and observed temperature field data; Taking each weather factor in the multi-weather factor data as a factor node, taking the coupling relationship between the weather factors as a factor correlation edge, and giving the factor correlation edge a correlation edge weight; Integrating the factor nodes, the factor correlation edges and the correlation edge weights to output a weather graph structure; Inputting the weather graph structure into a graph convolution network to perform deep feature extraction, and outputting a multi-weather factor feature tensor; Performing global and local spatiotemporal characteristic analysis on the multi-weather factor feature tensor through a dynamic gating fusion network to generate a dynamic weight coefficient; weighting and fusing the dynamic weight coefficient and the multi-weather factor feature tensor to output a fused feature tensor; Inputting the fused feature tensor into a fully connected neural network to perform nonlinear transformation and dimension mapping, and outputting corrected temperature field data; comparing the corrected temperature field data with the observed temperature field data to output a temperature correction value; based on the temperature correction value, performing network parameter optimization on the graph convolution network, the dynamic gating fusion network and the fully connected neural network through a back propagation algorithm to form a deep learning temperature correction model; Performing different model parameter perturbations on the temperature correction model to output a correction model set; performing quality verification on each temperature correction model in the correction model set to generate an accuracy index; inputting the accuracy index and the model feature of each temperature correction model as a training set into a quality detection model for training to determine the quality of each temperature correction model.

2. The deep learning-based multi-meteorological factor model forecast temperature revision method of claim 1, wherein: The specific steps of inputting the weather graph structure into the graph convolution network to perform deep feature extraction and output a multi-weather factor feature tensor are as follows: Inputting the weather graph structure into the graph convolution network, aggregating the neighborhood information of the factor nodes through the message passing mechanism of the graph convolution network, and outputting a node feature tensor; Performing feature transformation and nonlinear activation on the node feature tensor to output a multi-weather factor feature tensor. 3.The deep learning based multi-meteorological factor model prediction temperature revision method of claim 1, wherein: The specific steps of performing global and local spatiotemporal characteristic analysis on the multi-weather factor feature tensor through the dynamic gating fusion network to generate a dynamic weight coefficient are as follows: Performing spatiotemporal convolution on the multi-weather factor feature tensor through the dynamic gating fusion network to output local spatiotemporal features; Performing global pooling on the multi-weather factor feature tensor to output global statistical characteristics; Concatenating the local spatiotemporal features and the global statistical characteristics to generate a comprehensive feature vector; Performing linear transformation and normalization on the comprehensive feature vector to generate a dynamic weight coefficient.

4. The deep learning-based multi-meteorological factor model forecast temperature revision method of claim 1, wherein: The specific steps of weighting and fusing the dynamic weight coefficient and the multi-weather factor feature tensor to output a fused feature tensor are as follows: Coefficient weighting the dynamic weight coefficient and the multi-weather factor feature tensor to output a weighted feature tensor; Classifying and aggregating the weighted feature tensor along the feature dimension of the weather factor to output an aggregated feature tensor; Adjusting the dimension and representation of the aggregated feature tensor to output a fused feature tensor.

5. The deep learning-based multi-meteorological factor model forecast temperature correction method of claim 1, wherein: The specific steps of inputting the fused feature tensor into the fully connected neural network to perform nonlinear transformation and dimension mapping and output corrected temperature field data are as follows: Inputting the fused feature tensor into the fully connected neural network, and performing linear transformation on the fused feature tensor through a first fully connected layer of the fully connected neural network to generate an intermediate feature representation; The intermediate feature representation is nonlinearly transformed by a ReLU activation function to generate a nonlinear feature representation; The nonlinear feature representation is mapped to a temperature field dimension by a second fully connected layer to output corrected temperature field data.

6. The deep learning-based multi-meteorological factor model forecast temperature revision method of claim 1, wherein: The corrected temperature field data is compared with the observed temperature field data to output a temperature correction value, and the specific steps are, The corrected temperature field data is compared with the observed temperature field data to output a temperature correction value, and the specific steps are, The corrected temperature field data is compared with the observed temperature field data to output a temperature correction value, and the specific steps are, 7. The deep learning-based multi-meteorological factor model prediction temperature revision method of claim 1, wherein: The corrected temperature field data is compared with the observed temperature field data to output a temperature correction value, and the specific steps are, The temperature correction value is subjected to mean square error statistics to output a loss function value; The loss function value is used as the starting point of the back propagation algorithm, and the gradient information of the graph convolution network, the dynamic gate fusion network and the 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 the graph convolution network, the dynamic gate fusion network and the fully connected neural network according to the gradient information, and the updated network parameters are outputted; The updated network parameters are loaded into the architecture of the graph convolution network, the dynamic gate fusion network and the fully connected neural network for optimization and adjustment, and the deep learning temperature correction model is formed by integration.

8. The deep learning-based multi-meteorological factor model forecast temperature correction method of claim 1, wherein: The temperature correction model is subjected to different model parameter perturbations to output a correction model set, and the specific steps are, The temperature correction model is subjected to parameter sensitivity analysis to determine the network parameter range that needs to be perturbed; Based on the network parameter range, multiple different parameter configuration combinations are set; The temperature correction model is initialized by the parameter configuration combination to output a correction model set.

9. The deep learning-based multi-meteorological factor model prediction temperature revision method of claim 1, wherein: The quality of each temperature correction model in the correction model set is verified to generate an accuracy index, which means that the multi-weather factor data and the observed temperature field data are used as the training set and the verification set respectively, the training set is used to perform temperature correction prediction on each temperature correction model in the correction model set, the results of the temperature correction prediction are compared with the verification set, and the difference comparison results are quantitatively analyzed to generate an accuracy index.

10. The deep learning-based multi-meteorological factor model forecast temperature revision method of claim 1, wherein: The accuracy index and the model meta-feature of each temperature correction model are used as the training set and input into the quality detection model for training to determine the quality of each temperature correction model, and the specific steps are, A quality detection model is constructed based on the scikit-learn machine learning framework; The model meta-feature of each temperature correction model is standardized to generate a standardized feature vector; The standardized feature vector and the accuracy index are paired with feature labels to form a detection training sample; The quality detection model is supervised learning trained by using the detection training sample, and the quality of the temperature correction model is detected by using the trained quality detection model to output a quality judgment result.

Images