High-precision prediction method for dynamic selection and graph structure adaptive fusion of multi-source weather forecast oriented to turning weather

Abstract

The application relates to a high-precision prediction method for dynamic selection and graph structure adaptive fusion of multi-source weather forecasts for turning weather, relates to the technical field of meteorological information processing and intelligent prediction, and utilizes a graph convolution method to perform information transmission between modes, so that a nonlinear and adaptive fusion mechanism is realized. Through dynamic adjustment of edge weights and optimization of convolution weights, complementary information of different modes is fully utilized, and the prediction precision and local feature capturing capability are improved. Considering prediction uncertainty, historical performance and observation consistency, an interpretable confidence index is formed. For a low-confidence area, mode expansion or edge weight correction is automatically triggered, adaptive optimization of the prediction result is realized, and the robustness and reliability of the system are improved. Multi-source numerical mode output, real-time observation and turning weather characteristics are formed into a unified input system, comprehensive information support is provided for mode selection and fusion, data utilization efficiency is improved, and high-precision prediction can still be realized under complex weather conditions.

Description

Technical Field

[0001] This invention relates to the field of meteorological information processing and intelligent forecasting technology, specifically to a high-precision forecasting method for dynamic selection of multi-source meteorological forecasts and adaptive fusion of graph structures for transitional weather events. Background Technology

[0002] Current multi-source weather forecasts have the following shortcomings under transitional weather conditions: (1) Lack of adaptability in model selection: Traditional methods usually use fixed model combinations or manual experience to select models, which cannot be dynamically adjusted according to weather conditions and historical model performance. During transitional weather events, the performance of different models varies significantly, and fixed combinations lead to a decrease in prediction accuracy and local distortion.

[0003] (2) The fusion method lacks nonlinearity and structuring capabilities: Existing weighting methods are mostly linear superposition or simple parameter optimization, which fail to make full use of complementary information between models. They cannot effectively capture the local characteristics of extreme weather, resulting in insufficient accuracy and robustness of the fusion results.

[0004] (3) Lack of adaptive and interpretable confidence quantification: Existing fusion methods usually lack dynamic, uncertainty-driven confidence assessment mechanisms. They cannot adaptively extend or adjust based on low-confidence regions, thus reducing the reliability of predictions under extreme weather conditions.

[0005] (4) Insufficient structured utilization of multi-source information: The output of multi-source models, real-time observation and turning weather characteristics have not formed a unified structured input system, resulting in low information utilization efficiency. Summary of the Invention

[0006] To overcome the shortcomings of the above technologies, this invention provides a high-precision prediction method for dynamic selection and graph structure adaptive fusion of multi-source meteorological forecasts for transitional weather, which can dynamically select forecast models, adaptively adjust weighting parameters, and output confidence assessment results.

[0007] The technical solution adopted by this invention to overcome its technical problems is: A high-precision forecasting method for dynamic selection of multi-source meteorological forecasts and adaptive fusion of graph structure for transitional weather events includes: S1. Acquisition Mode Spatial grid point location Forecast time Forecast output results The forecast output results will be displayed. Normalization is performed to obtain the standardized pattern output. , , For pattern The total number of grid points, , Number of patterns; S2. Obtain the position of spatial grid points Forecast time Real-time observation data, real-time observation data Normalization is performed to obtain standardized observations. ; S3. Calculate the dynamic reliability of the model. ; S4. Calculate the pattern Spatial grid point location Forecast time Transitional Weather Index ; S5. Based on mode dynamic reliability and turning point weather index Obtain node feature vectors Normalized adaptive edge weights ; S6. Based on the normalized adaptive edge weights and node feature vectors Obtain the combined feature vector ; S7. Combine feature vectors Input into the GCN network, the GCN network's first Layer output features are , , The number of layers in the GCN network; S8. Features Residual enhancement is performed to obtain enhanced features. ; S9. Utilizing Enhanced Features Obtain the fusion feature tensor ; S10. Utilizing the fused feature tensor Obtain meteorological element prediction field .

[0008] Furthermore, in step S1, model data is obtained from the ECMWF reanalysis and forecast dataset and / or the GFS global numerical weather prediction dataset and / or the output data of the local high-resolution numerical weather prediction model. Spatial grid point location Forecast time Forecast output results .

[0009] Furthermore, in step S2, the spatial grid locations are obtained from the CLDAS dataset. Forecast time Real-time observation data .

[0010] Furthermore, in step S3, the formula is used... The dynamic reliability of the model is calculated. In the formula, This is the sensitivity coefficient. The value ranges from 0.5 to 1.0. The time decay coefficient, The value ranges from 0.8 to 0.95. The length of the sliding window. The value ranges from 24 to 72 hours. The number of time steps in the sliding window. , For the standardized model Spatial grid point location Forecast time The output, For the standardized grid point positions Forecast time Observations.

[0011] Furthermore, through the formula The turning point weather index was calculated. In the formula For pattern Forecast time Spatial grid point location The value after normalizing the maximum temperature gradient. For pattern Forecast time Spatial grid point location The value after normalizing the maximum humidity gradient. For pattern Forecast time Spatial grid point location The value after normalization of the maximum wind shear gradient. For pattern Forecast time Spatial grid point location The value after normalizing the maximum value of the precipitation intensity gradient; , , , All are weights. .

[0012] Furthermore, step S5 includes the following steps: S5-1. Through formula Calculate the first Dynamic mode reliability score for each mode , In the formula, , All are weights. , ; S5-2. Sort all dynamic modes of each model by reliability score from highest to lowest, and then select the top [models]. The patterns corresponding to the reliability scores of each dynamic pattern constitute a subset of candidate patterns. ,in, ; S5-3. Through formula Calculate the subset of candidate patterns The Middle Node feature vectors of each pattern , ,in For a subset of candidate patterns The Middle The pattern is in the forecast time. Standardized output venue ; S5-4. Through formula Calculate the subset of candidate patterns The Middle The pattern and the first Initial values ​​of edge weights for each pattern In the formula, The Pearson correlation coefficient is used. For a subset of candidate patterns The Middle The pattern is in the forecast time. Standardized output field By the forecast time Standardized output field The set, ; S5-5. Through formula Calculate the adaptive edge weights In the formula, The edge weight sensitivity coefficient, The value ranges from 0.5 to 1.0. For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices, For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices; S5-6. Adaptive edge weights Normalization is performed to obtain the normalized adaptive edge weights. .

[0013] Furthermore, step S6 includes the following steps: S6-1. Candidate Pattern Subset middle A pattern as There are nodes, and the relationships between every two nodes form edges, resulting in a graph structure; S6-2. Through formula Calculate the adaptive adjacency weight In the formula For node feature similarity multipliers, , Let the norm of the vector be... For a subset of candidate patterns The Middle The node feature vectors of each pattern , ; S6-2. Subset of candidate patterns No. Each node's neighbor nodes are multiplied by an adaptive adjacency weight. Summing the results gives the first... The neighborhood aggregation features of each node at the current time ; S6-3. Aggregating Neighborhood Features With node feature vectors Perform a concatenation operation to obtain the combined feature vector. .

[0014] Furthermore, in step S8, the formula is used... Calculate the enhanced features ,in For the GCN network Features of the layer output.

[0015] Furthermore, step S9 includes the following steps: S9-1. Through formula Calculate the fused node features In the formula As weight, , For the real number space, The feature dimension for each grid point; S9-2. Through formula The unified fusion features were calculated. In the formula As weight, ; S9-3. Unifying the integration features Rearrange the grid points according to their latitude and longitude indices to obtain the fused feature tensor. , , This represents the number of grid points along the longitude direction. This represents the number of grid points in the latitudinal direction. .

[0016] Furthermore, in step S10, CNN-based mapping is used to fuse the feature tensors. Mapped to meteorological element prediction field , , This represents the number of grid points in the longitude direction of the meteorological element prediction field. This represents the number of grid points in the dimensional direction of the meteorological element prediction field. This represents the number of channels.

[0017] The beneficial effects of this invention are: Dynamic adaptive model selection: By combining historical model performance, turning point weather indices, and near-real-time observations, an adaptive selection of candidate model subsets is achieved. Under turning point weather conditions, the participating models are automatically adjusted, significantly improving the quality and representativeness of input data and solving the problem of decreased accuracy in traditional fixed-combination models.

[0018] Graph-based nonlinear adaptive fusion: This method utilizes graph convolution to propagate information between patterns, achieving a nonlinear and adaptive fusion mechanism. Through dynamic adjustment of edge weights and optimization of convolution weights, complementary information from different patterns is fully utilized, improving prediction accuracy and the ability to capture local features.

[0019] Multidimensional confidence assessment and closed-loop correction: Taking into account prediction uncertainty, historical performance, and observation consistency, an interpretable confidence index is formed. Automatic mode expansion or edge weight correction is triggered for low-confidence regions to achieve adaptive optimization of prediction results, improving system robustness and reliability.

[0020] Efficient utilization of structured multi-source information: A unified input system is formed by integrating multi-source numerical model outputs, real-time observations, and pivotal weather characteristics, providing comprehensive information support for model selection and fusion. This improves data utilization efficiency and ensures high-precision forecasting even under complex weather conditions.

[0021] An end-to-end closed-loop forecasting system: This system forms a complete closed loop, from input data acquisition and dynamic model selection to adaptive graph structure fusion and multi-dimensional confidence assessment. The system can adaptively update and optimize, achieving high-precision, interpretable, and robust weather forecasts, providing a reliable foundation for decision support and automated applications. Detailed Implementation

[0022] The present invention will be further described below.

[0023] A high-precision forecasting method for dynamic selection of multi-source meteorological forecasts and adaptive fusion of graph structure for transitional weather events includes: S1. Acquisition Mode Spatial grid point location Forecast time Forecast output results The forecast output results will be displayed. Normalization is performed to obtain the standardized pattern output. , , For pattern The total number of grid points, , The number of patterns.

[0024] S2. Obtain the position of spatial grid points Forecast time Real-time observation data, real-time observation data Normalization is performed to obtain standardized observations. .

[0025] S3. Calculate the dynamic reliability of the model. .

[0026] S4. Calculate the pattern Spatial grid point location Forecast time Transitional Weather Index .

[0027] S5. Based on mode dynamic reliability and turning point weather index Obtain node feature vectors Normalized adaptive edge weights .

[0028] S6. Based on the normalized adaptive edge weights and node feature vectors Obtain the combined feature vector .

[0029] S7. Combine feature vectors The input is fed into a GCN network (Graph Convolutional Network), and the GCN network's first... Layer output features are , , This represents the number of layers in the GCN network.

[0030] S8. Features Residual enhancement is performed to obtain enhanced features. .

[0031] S9. Utilizing Enhanced Features Obtain the fusion feature tensor .

[0032] S10. Utilizing the fused feature tensor Obtain meteorological element prediction field .

[0033] The high-precision forecasting method of dynamic selection and graph structure adaptive fusion of multi-source meteorological forecasts for transitional weather, as proposed in this invention, aims to solve the following core technical problems: The problem of dynamic model selection: How to combine historical model performance, turning point weather indices, and real-time observations to achieve adaptive screening and subset optimization of candidate models.

[0034] Nonlinear adaptive fusion problem: How to use graph convolution to propagate features between candidate pattern nodes, achieve nonlinear and adaptive fusion between patterns, and improve prediction accuracy under turning weather conditions.

[0035] Adaptive confidence assessment and feedback problem: How to generate multidimensional confidence indices based on the uncertainty of fusion prediction, historical performance and observation consistency, and to trigger mode expansion or edge weight adjustment for low confidence areas to achieve interpretability and robustness of prediction results.

[0036] The problem of utilizing structured multi-source data: How to construct a unified input feature system that integrates multi-source model outputs, real-time observations, transitional weather features, and data quality masks to provide efficient information support for dynamic model selection and graph convolutional fusion.

[0037] Based on traditional multi-model weighted fusion, graph neural network structure modeling and dynamic structure learning mechanisms are introduced to achieve adaptive modeling and time-varying optimization of the correlation between multi-source meteorological models.

[0038] Its core innovation lies in: The model has been upgraded from "weighted average" to "graph structure modeling"—treating each meteorological model as a graph node and constructing a dynamic graph structure through the correlation, consistency and reliability between nodes. Instead of linear independent weighting, the model achieves "dependency propagation" and "complementary enhancement" between models through dynamic learning of the graph adjacency matrix.

[0039] Adaptive graph structure update mechanism: When the system detects that the weather has entered a turning point (such as fronts, sudden changes in precipitation, strong winds, etc.), the graph structure is reparameterized through the turning point index Ct, making the updates of edge weights and node features between models more sensitive to sudden weather features, thus achieving adaptive adjustment at the "structural level". Reliability-guided node feature reconstruction—the reliability of CLDAS real-time observation error inversion is co-encoded with meteorological element inputs to form a time-varying node feature vector, realizing the mapping from the original model field to the dynamic reliable feature space.

[0040] The fusion output is interpretable—the output is not just a single weighted average, but a meteorological element prediction field obtained through graph convolution propagation. The response of each output node can be traced back to the corresponding input model node and edge weight contribution, thus possessing interpretability and stability.

[0041] In one embodiment of the present invention, step S1 involves obtaining the model from the ECMWF reanalysis and forecast dataset and / or the GFS global numerical weather prediction dataset and / or the output data of the local high-resolution numerical weather prediction model. Spatial grid point location Forecast time Forecast output results The ECMWF (European Centre for Medium-Range Weather Forecasts) reanalysis and forecast dataset provides data on temperature, humidity, wind speed, and precipitation. The GFS (Global Forecast System, provided by the National Center for Environmental Prediction, NCEP) global numerical forecast dataset contains key meteorological elements.

[0042] In one embodiment of the present invention, in step S2, spatial grid locations are obtained from the CLDAS dataset (real-time product dataset of the China Meteorological Administration Land Surface Data Assimilation System). Forecast time Real-time observation data .

[0043] In one embodiment of the present invention, step S3 is performed using the formula The dynamic reliability of the model is calculated. In the formula, This is the sensitivity coefficient. The time decay coefficient, The value ranges from 0.8 to 0.95. The value ranges from 0.5 to 1.0. The length of the sliding window. The value ranges from 24 to 72 hours. The number of time steps in the sliding window. , For the standardized model Spatial grid point location Forecast time The output, For the standardized grid point positions Forecast time Observations.

[0044] In one embodiment of the present invention, through formula The turning point weather index was calculated. In the formula For pattern Forecast time Spatial grid point location The normalized value of the maximum temperature gradient reflects the activity of the cold and hot boundary or frontal zone. For pattern Forecast time Spatial grid point location The value after normalization of the maximum humidity gradient reveals the signal of abrupt changes in dryness and wetness. For pattern Forecast time Spatial grid point location The value after normalizing the maximum wind shear gradient is used to capture wind field disturbances and convection triggering conditions. For pattern Forecast time Spatial grid point location The value after normalizing the maximum value of the precipitation intensity gradient is used to identify areas of heavy precipitation or convective activity. , , , All are weights. Higher weight values ​​indicate more dramatic weather changes at grid points or in specific regions. The system can dynamically adjust model weights and edge weights accordingly to improve prediction accuracy under extreme weather conditions. In one embodiment of the present invention, step S5 includes the following steps: S5-1. Through formula Calculate the first Dynamic mode reliability score for each mode , In the formula, , All are weights. , This approach ensures a comprehensive consideration of both the model's recent reliability and its adaptability to changing weather conditions. Each model's performance at the current moment is quantitatively scored, taking into account both recent statistical performance and adaptability to changing weather conditions, providing a quantitative basis for model selection.

[0045] S5-2. Sort all dynamic modes of each model by reliability score from highest to lowest, and then select the top [models]. The patterns corresponding to the reliability scores of each dynamic pattern constitute a subset of candidate patterns. ,in, Selecting a high-quality subset of patterns from multiple sources ensures the accuracy and diversity of subsequent fusion.

[0046] S5-3. Through formula Calculate the subset of candidate patterns The Middle Node feature vectors of each pattern , ,in For a subset of candidate patterns The Middle The pattern is in the forecast time. The standardized output field (corresponding to meteorological element values ​​at different spatial grid points). Node feature vectors It comprehensively characterizes the spatial forecast distribution and current performance level of the model, and serves as the basic input for subsequent graph structure modeling.

[0047] S5-4. Through formula Calculate the subset of candidate patterns The Middle The pattern and the first Initial values ​​of edge weights for each pattern In the formula, The Pearson correlation coefficient is used. For a subset of candidate patterns The Middle The pattern is in the forecast time. Standardized output field By the forecast time Standardized output field The set, Adding 1 and dividing by 2 is used to linearly map the correlation coefficient to the [0,1] interval, which is convenient for use as input for graph edge weights.

[0048] S5-5. To reflect the impact of transitional weather on model correlation, a transitional weather index is introduced for dynamic correction. Specifically, this is achieved through the formula... Calculate the adaptive edge weights In the formula, The edge weight sensitivity coefficient, The value ranges from 0.5 to 1.0. For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices, For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices.

[0049] S5-6. Adaptive edge weights Normalization is performed to obtain the normalized adaptive edge weights. This ensures that the sum of the weights of the neighboring edges of each node is 1, which facilitates subsequent graph convolution processing.

[0050] Traditional methods use only fixed edge weights or simple correlations. This method, however, achieves dynamic propagation of information between models through adaptive edge weight adjustment driven by a dynamic weather transition index, thus improving fusion performance under extreme weather conditions. Node features not only include model outputs but also incorporate dynamic reliability, providing high-quality input for subsequent graph convolution.

[0051] In one embodiment of the present invention, step S6 includes the following steps: S6-1. Candidate Pattern Subset middle A pattern as There are nodes, and the relationships between every two nodes form edges, resulting in a graph structure.

[0052] S6-2. Through formula Calculate the adaptive adjacency weight In the formula For node feature similarity multipliers, , Let the norm of the vector be... For a subset of candidate patterns The Middle The node feature vectors of each pattern , .

[0053] S6-2. Subset of candidate patterns No. Each node's neighbor nodes are multiplied by an adaptive adjacency weight. Summing the results gives the first... The neighborhood aggregation features of each node at the current time .

[0054] S6-3. Aggregating Neighborhood Features With node feature vectors Perform a concatenation operation to obtain the combined feature vector. .

[0055] In one embodiment of the present invention, step S8 is performed using the formula Calculate the enhanced features ,in For the GCN network The features output by the layer are used for feature updates, thereby preserving local information in the lower layers, so that deep convolutions do not cause excessive smoothing of features, while maintaining the continuity of spatiotemporal information of nodes.

[0056] In one embodiment of the present invention, step S9 includes the following steps: S9-1. Through formula Calculate the fused node features In the formula As weight, , For the real number space, The feature dimension for each grid point. Node features. It contains both local neighbor information and deep global information, providing input for subsequent adaptive adjacency learning and feature aggregation layers.

[0057] S9-2. Through formula The unified fusion features were calculated. In the formula As weight, .

[0058] S9-3. Unifying the integration features Rearrange the grid points according to their latitude and longitude indices to obtain the fused feature tensor. , , This represents the number of grid points along the longitude direction. This represents the number of grid points in the latitudinal direction. .

[0059] In one embodiment of the present invention, step S10 employs CNN-based mapping to fuse the feature tensors. Mapped to meteorological element prediction field , , This represents the number of grid points in the longitude direction of the meteorological element prediction field. This represents the number of grid points in the dimensional direction of the meteorological element prediction field. This represents the number of channels. Convolutional mapping utilizes CNNs to capture the spatial dependencies between grid points, thereby outputting a continuous physical quantity field. (Meteorological element prediction field) Indicates the forecast time The multi-element meteorological forecast results covering the entire region can be directly used as the final output for applications such as meteorological forecast correction, power load forecasting, or energy output assessment.

[0060] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A high-precision forecasting method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transitional weather events, characterized in that: include: S1. Acquisition Mode Spatial grid point location Forecast time Forecast output results The forecast output results will be displayed. Normalization is performed to obtain the standardized pattern output. , , For pattern The total number of grid points, , Number of patterns; S2. Obtain the position of spatial grid points Forecast time Real-time observation data, real-time observation data Normalization is performed to obtain standardized observations. ; S3. Calculate the dynamic reliability of the model. ; S4. Calculate the pattern Spatial grid point location Forecast time Transitional Weather Index ; S5. Based on mode dynamic reliability and turning point weather index Obtain node feature vectors Normalized adaptive edge weights ; S6. Based on the normalized adaptive edge weights and node feature vectors Obtain the combined feature vector ; S7. Combine feature vectors Input into the GCN network, the GCN network's first Layer output features are , , The number of layers in the GCN network; S8. Features Residual enhancement is performed to obtain enhanced features. ; S9. Utilizing Enhanced Features Obtain the fusion feature tensor ; S10. Utilizing the fused feature tensor Obtain meteorological element prediction field .

2. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that: In step S1, the model is obtained from the ECMWF reanalysis and forecast dataset and / or the GFS global numerical weather prediction dataset and / or the output data of the local high-resolution numerical weather prediction model. Spatial grid point location Forecast time Forecast output results .

3. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that: In step S2, spatial grid locations are obtained from the CLDAS dataset. Forecast time Real-time observation data .

4. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that: In step S3, the formula is used. The dynamic reliability of the model is calculated. In the formula, This is the sensitivity coefficient. The value ranges from 0.5 to 1.

0. The time decay coefficient, The value ranges from 0.8 to 0.

95. The length of the sliding window. The value ranges from 24 to 72 hours. The number of time steps in the sliding window. , For the standardized model Spatial grid point location Forecast time The output, For the standardized grid point positions Forecast time Observations.

5. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that: Through formula The turning point weather index was calculated. In the formula For pattern Forecast time Spatial grid point location The value after normalizing the maximum temperature gradient. For pattern Forecast time Spatial grid point location The value after normalizing the maximum humidity gradient. For pattern Forecast time Spatial grid point location The value after normalization of the maximum wind shear gradient. For pattern Forecast time Spatial grid point location The value after normalizing the maximum value of the precipitation intensity gradient; , , , All are weights. .

6. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 4, characterized in that, Step S5 includes the following steps: S5-1. Through formula Calculate the first Dynamic mode reliability score for each mode , In the formula, , All are weights. , ; S5-2. Sort all dynamic modes of each model by reliability score from highest to lowest, and then select the top [models]. The patterns corresponding to the reliability scores of each dynamic pattern constitute a subset of candidate patterns. ,in, ; S5-3. Through formula Calculate the subset of candidate patterns The Middle Node feature vectors of each pattern , ,in For a subset of candidate patterns The Middle The pattern is in the forecast time. Standardized output venue ; S5-4. Through formula Calculate the subset of candidate patterns The Middle The pattern and the first Initial values ​​of edge weights for each pattern In the formula, The Pearson correlation coefficient is used. For a subset of candidate patterns The Middle The pattern is in the forecast time. Standardized output field By the forecast time Standardized output field The set, ; S5-5. Through formula Calculate the adaptive edge weights In the formula, The edge weight sensitivity coefficient, The value ranges from 0.5 to 1.

0. For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices, For a subset of candidate patterns The Middle The pattern is in the forecast time. All The average of several pivotal weather indices; S5-6. Adaptive edge weights Normalization is performed to obtain the normalized adaptive edge weights. .

7. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events as described in claim 6, is characterized in that... Step S6 includes the following steps: S6-1. Candidate Pattern Subset middle A pattern as There are nodes, and the relationships between every two nodes form edges, resulting in a graph structure; S6-2. Through formula Calculate the adaptive adjacency weight In the formula For node feature similarity multipliers, , Let the norm of the vector be... For a subset of candidate patterns The Middle The node feature vectors of each pattern , ; S6-2. Subset of candidate patterns No. Each node's neighbor nodes are multiplied by an adaptive adjacency weight. Summing the results gives the first... The neighborhood aggregation features of each node at the current time ; S6-3. Aggregating Neighborhood Features With node feature vectors Perform a concatenation operation to obtain the combined feature vector. .

8. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that: In step S8, the formula is used. Calculate the enhanced features ,in For the GCN network Features of the layer output.

9. The high-precision prediction method for dynamic selection and adaptive fusion of multi-source meteorological forecasts for transformative weather events according to claim 1, characterized in that, Step S9 includes the following steps: S9-1. Through formula Calculate the fused node features In the formula As weight, , For the real number space, The feature dimension for each grid point; S9-2. Through formula The unified fusion features were calculated. In the formula As weight, ; S9-3. Unifying the integration features Rearrange the grid points according to their latitude and longitude indices to obtain the fused feature tensor. , , This represents the number of grid points along the longitude direction. This represents the number of grid points in the latitudinal direction. .

10. The high-precision prediction method for dynamic selection of multi-source meteorological forecasts and adaptive fusion of graph structure for transitional weather as described in claim 1, characterized in that: In step S10, CNN-based mapping is used to fuse the feature tensors. Mapped to meteorological element prediction field , , This represents the number of grid points in the longitude direction of the meteorological element prediction field. This represents the number of grid points in the dimensional direction of the meteorological element prediction field. This represents the number of channels.

Images