Weather situation intelligent analysis method and device based on artificial intelligence multi-mode causal reasoning

Abstract

The invention relates to a weather situation intelligent analysis method based on artificial intelligence multi-modal causal reasoning, and the method comprises the following steps: obtaining multi-modal meteorological data, and carrying out the preprocessing of the multi-modal meteorological data, and obtaining the processed data; processing the processed data through a pre-constructed multi-modal causal reasoning neural network model to obtain a weather situation intelligent interpretation result; performing feedback correction on the weather situation intelligent interpretation result through an automatic feedback correction mechanism based on a large model so as to update the weather situation intelligent interpretation result; and generating a weather situation analysis result based on the updated weather situation intelligent interpretation result, and visually displaying the weather situation analysis result. Through the cooperative work, the weather system can be automatically identified, the causal relationship between the systems can be analyzed, a professional weather situation analysis report can be generated, and multi-mode forecast comparative analysis is supported. According to the invention, the efficiency and accuracy of weather situation analysis can be improved, and reliable technical support is provided for weather forecast service.

Description

Technical Field

[0001] This invention relates to the field of weather forecasting technology, and in particular to a method and apparatus for intelligent analysis of weather patterns based on artificial intelligence multimodal causal reasoning. Background Technology

[0002] Accurate identification and analysis of weather systems are crucial for the accuracy of weather forecasts. Current methods for weather system identification and weather pattern analysis mainly suffer from the following problems:

[0003] Limitations of traditional manual analysis: Currently, the identification of weather systems and the analysis of weather patterns mainly rely on the manual analysis and experience-based judgment of forecasters and chief experts. This method is not only time-consuming and labor-intensive, but also suffers from significant shortcomings in efficiency and accuracy when dealing with complex weather systems and multi-model forecast results.

[0004] There is a disconnect between algorithm research and practical application: most existing weather system identification algorithms remain in the research stage, with limited practical application. Although deep learning has made some progress in weather system identification, its integration with practical applications is still insufficient. Weather systems are influenced by multiple factors, including monsoons, tropical cyclones, and subtropical high pressure systems, exhibiting unique complexity.

[0005] Insufficient understanding of causal relationships in weather systems: Currently, some patents involve weather system identification, such as "A Deep Learning-Based Method, System, Device, and Medium for Atmospheric River Identification, CN119131603B" and "A Deep Learning-Based Radar Image Identification Model and Method for Severe Convective Weather, CN202110714050.6." However, these are limited to identification and lack the ability to understand and reason about the causal relationships between weather systems, resulting in limited analytical capabilities for complex weather situations. There is currently a lack of effective solutions to logically connect weather systems and the weather phenomena they cause, remaining only at the level of weather system identification and lacking a bridge to explain the corresponding weather phenomena.

[0006] Insufficient multimodal fusion: Current technologies do not provide in-depth analysis of the fusion of multimodal data such as meteorological numerical data, weather images and text descriptions, and the application of multimodal fusion in weather forecasting is still in its early stages.

[0007] Lack of intelligent weather situation interpretation: Currently, the meteorological field lacks technology capable of automatically generating professional and accurate weather situation text descriptions. Existing weather situation descriptions mainly rely on manual writing, lacking automation and intelligent methods.

[0008] In summary, existing technologies have significant shortcomings in weather system identification and weather situation analysis, making it difficult to meet the needs of accurate identification and efficient analysis of complex weather systems in target areas, especially in South China. There is an urgent need for a more intelligent and automated weather situation diagnosis and assessment technology. Summary of the Invention

[0009] The purpose of this invention is to at least address one of the shortcomings of the prior art and provide an intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] Specifically, a weather situation intelligent analysis method based on artificial intelligence multimodal causal reasoning is proposed, including the following:

[0012] Acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data;

[0013] The pre-constructed multimodal causal reasoning neural network model is used to process the processed data to obtain intelligent interpretation results of the weather situation;

[0014] The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables.

[0015] The intelligent interpretation results of the weather situation are updated by using an automated feedback correction mechanism based on a large model to correct the errors.

[0016] Based on the updated intelligent interpretation of the weather situation, a weather situation analysis result is generated and the weather situation analysis result is visualized.

[0017] Furthermore, specifically, the acquired multimodal meteorological data includes,

[0018] Numerical weather forecast model output data includes temperature, air pressure, wind field, and precipitation forecast data, as well as other data such as satellite cloud imagery, historical weather pattern text descriptions, and actual observation data from observation stations.

[0019] Furthermore, specifically, the multimodal meteorological data is preprocessed to obtain processed data, including:

[0020] The multimodal meteorological data is standardized by unifying data of different resolutions, formats, and spatiotemporal scales into a standard format, converting numerical data into floating-point tensors, unifying resolution, processing satellite cloud image data into image tensors of uniform size, and vectorizing text data.

[0021] Next, quality control is carried out to remove noise and outliers. The IQR method is used to identify outliers in the data, and statistical boundary values ​​are set as outlier judgment criteria. Low-quality data is marked or removed to obtain processed data.

[0022] Furthermore, specifically, the interaction mechanism combining symbolic rules and neural networks includes,

[0023] Design a bidirectional interface that allows the output of the neural network, i.e., the probability distribution, to be converted into a symbolic representation, while the result of the symbolic operation can be used as a conditional input to the neural network;

[0024] Implement an attention guidance mechanism by using symbolic rules to adjust the attention weights of the neural network;

[0025] Construct differentiable logical operators, converting "AND", "OR", and "NOT" logical operations into differentiable functions. The "AND" operation is represented using the SoftMin function, and the "OR" operation using the SoftMax function. Specifically, the differentiable logical operators are constructed as follows:

[0026] The AND operation in SoftMin is represented as: A∧B≈-log(e^( ... -A +e -B );

[0027] The OR operation in SoftMax is represented as: A∨B≈log(e^(-1 / 2)) -A +e -B );

[0028] The NOT operation is represented as:

[0029] Where A,B∈[0,1], represents the probabilistic representation of logical variables. The above functions guarantee differentiability in neural networks, thereby enabling end-to-end training.

[0030] Cross-modal causal inference rules are constructed. For each pair of modalities Xi and Xj, i,j∈{numerical data, image data, text data}, conditional mutual information I(Xi; Xj|Z) is calculated to identify direct causal relationships, where Z is the set of other common cause variables. A graph-based causal discovery algorithm is adopted, constructing a directed acyclic graph (DAG) to represent the causal relationships between variables. The edge weights are obtained by optimizing the objective function through gradient descent. The optimization objective function is defined as:

[0031]

[0032] in:

[0033] This represents the multimodal feature representation of the input (a concatenated representation of encoded numerical data, image data, and text data);

[0034] The reconstructed features are inferred from the causal graph generation model G;

[0035] ‖G‖1: The L1 norm of the adjacency matrix of the causal graph, introducing sparsity constraints;

[0036] λ: A hyperparameter used to adjust the tradeoff between reconstruction error and graph sparsity;

[0037] Establish a multi-level causal reasoning structure:

[0038] The first layer is the weather system identification layer, which identifies different weather systems as independent entities;

[0039] The second layer is the inter-system interaction layer, which establishes the interaction relationships between weather systems;

[0040] The third layer is the weather phenomenon forecasting layer, which establishes the causal relationship between system characteristics and the final weather phenomenon;

[0041] These three relationships are connected using differentiable programming techniques, and the temporal state changes are encoded using a bidirectional long short-term memory network (Bi-LSTM).

[0042] Furthermore, specifically, automated feedback correction mechanisms based on large models include,

[0043] Build an intelligent assessment and feedback system.

[0044] By integrating multiple large language models to form an evaluation cluster, mutual evaluation and collaborative decision-making among models can be achieved.

[0045] Design a structured evaluation protocol to standardize the output format of large models into a standard JSON structure. The output format includes fields for score, problem diagnosis, and improvement suggestions.

[0046] Develop an efficient model scheduling system that supports batch parallel evaluation and real-time feedback, and manage evaluation tasks through priority queues.

[0047] Establish a multi-dimensional quality assessment system, covering three core dimensions: professional accuracy, logical consistency, and analytical completeness;

[0048] Implement an adaptive analysis and optimization engine.

[0049] A vectorized weather system knowledge base is constructed, which stores the typical characteristics and standard analysis methods of various weather systems using vector embedding.

[0050] Using an intelligent matching algorithm, the optimal analysis path and parameter configuration are automatically selected based on the characteristics of current meteorological data.

[0051] This mechanism implements dynamic parameter tuning, leveraging the intelligent analysis capabilities of large models to automatically adjust analysis weights and thresholds based on intermediate results.

[0052] A multi-scheme comparison system is established, and the optimal result is selected from multiple analysis schemes by setting objective evaluation indicators.

[0053] Construct a closed-loop self-optimizing learning system.

[0054] Design an automated iterative process of "generation-evaluation-optimization-verification" to establish a complete quality closed loop.

[0055] In the "generation" phase, the core analytical model generates preliminary weather situation analysis based on multimodal data.

[0056] During the "evaluation" phase, the evaluation model cluster assesses the results through a triple-testing system:

[0057] Data consistency test: Evaluate the consistency of results across different data sources and calculate the correlation coefficients of key indicators.

[0058] Meteorological rationality verification: Verifying the physical rationality of the analysis results based on meteorological principles and laws.

[0059] System coverage verification: Ensure that the analysis fully covers key weather systems and their interactions.

[0060] During the "optimization" phase, the dedicated optimization model generates specific improvement strategies based on the evaluation results, including model parameter adjustments, analysis logic restructuring, and focus optimization.

[0061] During the "validation" phase, the independent validation model compares the optimized results with historical cases and physical laws to ensure the effectiveness of the improvements.

[0062] Establish a complete performance tracing database to record the detailed process and effects of each optimization, forming a mechanism for the continuous evolution of system capabilities.

[0063] Furthermore, specifically, based on the updated intelligent interpretation results of the weather situation, weather situation analysis results are generated, including:

[0064] Automatically generate weather situation analysis text reports:

[0065] The design incorporates a hierarchical report generation framework, comprising four levels: a "Summary Layer" for a brief overview, a "System Layer" for a detailed description of each weather system's characteristics, an "Impact Layer" for analyzing the weather system's impact on actual weather, and a "Forecast Layer" for forecasting future trends.

[0066] A specific text template library is defined for each level, and templates are organized according to different categories, including system type, impact type, and evolution type.

[0067] A conditional text generation algorithm is used to populate the template, with conditions including system type, intensity, location, and direction of movement. A beam search is then employed to generate the optimal representation.

[0068] Post-processing rules are applied to ensure accurate terminology and fluent sentences, and a professional meteorological terminology mapping table is used for standardization.

[0069] Achieve comparative analysis of multi-model forecasts:

[0070] Automatically acquire and process forecast data from different numerical models, unifying data resolution and time intervals.

[0071] Indices of consistency among calculation models include spatial correlation coefficient, root mean square error (RMSE), and systematic bias.

[0072] Based on historical performance statistics, weight coefficients are assigned to each model to establish a comprehensive weighted evaluation system.

[0073] For key weather systems, confidence intervals are calculated, and outliers in the model are identified and analyzed.

[0074] Generate a consistency analysis report, including a forecast consistency heatmap, a key element comparison table, and a preliminary analysis of the reasons for discrepancies.

[0075] Furthermore, specifically, the weather situation analysis results are visualized, including:

[0076] Building an intelligent visualization system:

[0077] Develop hierarchical visualization components, including weather system overview maps, system detail maps, and causal explanations. Figure 3 Each level

[0078] Automatically identify key areas and add text descriptions to avoid overlapping annotations.

[0079] The visualization focus is dynamically adjusted based on user interests and data importance. The importance score is calculated using a weighted average of information entropy and anomaly degree, with high-scoring areas being highlighted.

[0080] It supports timeline interaction, allowing users to drag along the timeline to view the evolution of weather systems, and the time accuracy can be adjusted between different scales.

[0081] This invention also proposes an intelligent weather situation analysis device based on artificial intelligence multimodal causal reasoning, comprising the following:

[0082] A multimodal data acquisition and preprocessing module is used to acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data.

[0083] A causal reasoning module combining neural networks and symbolic systems is used to process the processed data through a pre-built multimodal causal reasoning neural network model to obtain intelligent interpretation results of weather conditions.

[0084] The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables.

[0085] The large-model-based automated feedback correction module is used to update the intelligent weather interpretation results by feeding back and correcting them through the large-model-based automated feedback correction mechanism.

[0086] The weather situation intelligent analysis product generation module is used to generate weather situation analysis results based on the updated weather situation intelligent interpretation results, and to visualize the weather situation analysis results.

[0087] The beneficial effects of this invention are as follows:

[0088] This invention proposes an intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning.

[0089] By constructing a causal reasoning framework combining neural networks and symbolic systems, intelligent processing and analysis of multimodal meteorological data are achieved. The system comprises four main technical modules: a multimodal data acquisition and preprocessing module, a causal reasoning module combining neural networks and symbolic systems, an automated feedback correction module based on a large model, and a weather situation intelligent analysis product generation module. Through the collaborative work of these four modules, the system can automatically identify weather systems, analyze causal relationships between systems, generate professional weather situation analysis reports, and support multi-model forecast comparative analysis. This method improves the efficiency and accuracy of weather situation analysis, providing reliable technical support for meteorological forecasting operations. Attached Figure Description

[0090] The above and other features of this disclosure will become more apparent from the detailed description of the embodiments illustrated in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings:

[0091] Figure 1 The flowchart shown is a process for the intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning of the present invention.

[0092] Figure 2 The diagram shown is a flowchart illustrating the operation of the multimodal causal reasoning neural network model in this invention.

[0093] Figure 3 The diagram shown is a flowchart for visualizing this invention. Detailed Implementation

[0094] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The same reference numerals used throughout the accompanying drawings indicate the same or similar parts.

[0095] Example 1, referring to Figure 1 , Figure 1 This paper demonstrates the overall architecture of an intelligent weather situation analysis system based on multimodal causal reasoning. The system acquires data from multimodal meteorological data sources, processes it through a multimodal causal reasoning framework, generates intelligent weather situation interpretations, and ultimately outputs weather situation analysis products. The system forms a closed loop through business applications and feedback, enabling continuous optimization. The multimodal data processing and causal reasoning module comprises four main steps: data acquisition and preprocessing, combination of neural networks and symbolic systems, automated feedback correction of large models, and product generation. This invention proposes an intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning, including the following:

[0096] Acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data;

[0097] Reference Figure 2This paper details the processing flow of a large-scale weather situation analysis model based on causal reasoning. Starting with multimodal data acquisition, the system undergoes data normalization and quality control, then moves to a stage combining neural networks and symbolic systems. Multi-level causal reasoning is then performed, ultimately generating weather situation text, comparing multi-model forecasts, and providing visualizations. The system utilizes an automated feedback correction framework to achieve closed-loop optimization, continuously improving analysis quality. The pre-constructed multimodal causal reasoning neural network model processes the pre-processed data to obtain intelligent interpretations of the weather situation.

[0098] The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables.

[0099] The intelligent interpretation results of the weather situation are updated by using an automated feedback correction mechanism based on a large model to correct the errors.

[0100] Reference Figure 3 , Figure 3 The system demonstrates the generation and application process of weather situation analysis products. Starting with multimodal meteorological data, the system processes it through a multimodal causal reasoning framework to generate three main products: text reports, multimodal comparisons, and visualizations. These products, after evaluation and optimization using a large-scale model automated feedback framework, are applied to three main scenarios: weather consultation, forecasting decision-making, and disaster early warning. The system achieves continuous improvement through a feedback correction mechanism; based on the updated intelligent interpretation of the weather situation, it generates weather situation analysis results and visualizes these results.

[0101] In a preferred embodiment of the present invention, the acquired multimodal meteorological data specifically includes...

[0102] Numerical weather forecast model output data includes temperature, air pressure, wind field and precipitation forecasts, as well as other data including satellite cloud imagery, historical weather pattern text descriptions, and observation station data.

[0103] Specifically, the multimodal meteorological data is preprocessed to obtain processed data, including:

[0104] The multimodal meteorological data is standardized by unifying data of different resolutions, formats, and spatiotemporal scales into a standard format, converting numerical data into floating-point tensors, unifying resolution, processing satellite cloud image data into image tensors of uniform size, and vectorizing text data.

[0105] Next, quality control is carried out to remove noise and outliers. The IQR method is used to identify outliers in the data, and statistical boundary values ​​are set as outlier judgment criteria. Low-quality data is marked or removed to obtain processed data.

[0106] As a preferred embodiment of the present invention, specifically, the interaction mechanism combining symbol rules and neural networks includes,

[0107] Design a bidirectional interface that allows the output of the neural network, i.e., the probability distribution, to be converted into a symbolic representation, while the result of the symbolic operation can be used as a conditional input to the neural network;

[0108] Implement an attention guidance mechanism by using symbolic rules to adjust the attention weights of the neural network;

[0109] Construct differentiable logical operators, converting "AND", "OR", and "NOT" logical operations into differentiable functions. The "AND" operation is represented using the SoftMin function, and the "OR" operation using the SoftMax function. Specifically, the differentiable logical operators are constructed as follows:

[0110] The AND operation in SoftMin is represented as: A∧B≈-log(e^( ... -A +e -B );

[0111] The OR operation in SoftMax is represented as: A∨B≈log(e^(-1 / 2)) -A +e -B );

[0112] The NOT operation is represented as:

[0113] Where A,B∈[0,1], represents the probabilistic representation of logical variables. The above functions guarantee differentiability in neural networks, thereby enabling end-to-end training.

[0114] Cross-modal causal inference rules are constructed. For each pair of modalities Xi and Xj, i,j∈{numerical data, image data, text data}, conditional mutual information I(Xi; Xj|Z) is calculated to identify direct causal relationships, where Z is the set of other common cause variables. A graph-based causal discovery algorithm is adopted, constructing a directed acyclic graph (DAG) to represent the causal relationships between variables. The edge weights are obtained by optimizing the objective function through gradient descent. The optimization objective function is defined as:

[0115]

[0116] in:

[0117] This represents the multimodal feature representation of the input (a concatenated representation of encoded numerical data, image data, and text data);

[0118] The reconstructed features are inferred from the causal graph generation model G;

[0119] ‖G‖1: The L1 norm of the adjacency matrix of the causal graph, introducing sparsity constraints;

[0120] λ: A hyperparameter used to adjust the tradeoff between reconstruction error and graph sparsity;

[0121] Establish a multi-level causal reasoning structure:

[0122] The first layer is the weather system identification layer, which identifies different weather systems as independent entities;

[0123] The second layer is the inter-system interaction layer, which establishes the interaction relationships between weather systems;

[0124] The third layer is the weather phenomenon forecasting layer, which establishes the causal relationship between system characteristics and the final weather phenomenon;

[0125] These three relationships are connected using differentiable programming techniques, and the temporal state changes are encoded using a bidirectional long short-term memory network (Bi-LSTM).

[0126] As a preferred embodiment of the present invention, specifically, the automated feedback correction mechanism based on a large model includes,

[0127] Build an intelligent assessment and feedback system.

[0128] By integrating multiple large language models to form an evaluation cluster, mutual evaluation and collaborative decision-making among models can be achieved.

[0129] Design a structured evaluation protocol to standardize the output format of large models into a standard JSON structure. The output format includes fields for score, problem diagnosis, and improvement suggestions.

[0130] Develop an efficient model scheduling system that supports batch parallel evaluation and real-time feedback, and manage evaluation tasks through priority queues.

[0131] Establish a multi-dimensional quality assessment system, covering three core dimensions: professional accuracy, logical consistency, and analytical completeness;

[0132] Implement an adaptive analysis and optimization engine.

[0133] A vectorized weather system knowledge base is constructed, which stores the typical characteristics and standard analysis methods of various weather systems using vector embedding.

[0134] Using an intelligent matching algorithm, the optimal analysis path and parameter configuration are automatically selected based on the characteristics of current meteorological data.

[0135] This mechanism implements dynamic parameter tuning, leveraging the intelligent analysis capabilities of large models to automatically adjust analysis weights and thresholds based on intermediate results.

[0136] A multi-scheme comparison system is established, and the optimal result is selected from multiple analysis schemes by setting objective evaluation indicators.

[0137] Construct a closed-loop self-optimizing learning system.

[0138] Design an automated iterative process of "generation-evaluation-optimization-verification" to establish a complete quality closed loop.

[0139] In the "generation" phase, the core analytical model generates preliminary weather situation analysis based on multimodal data.

[0140] During the "evaluation" phase, the evaluation model cluster assesses the results through a triple-testing system:

[0141] Data consistency test: Evaluate the consistency of results across different data sources and calculate the correlation coefficients of key indicators.

[0142] Meteorological rationality verification: Verifying the physical rationality of the analysis results based on meteorological principles and laws.

[0143] System coverage verification: Ensure that the analysis fully covers key weather systems and their interactions.

[0144] During the "optimization" phase, the dedicated optimization model generates specific improvement strategies based on the evaluation results, including model parameter adjustments, analysis logic restructuring, and focus optimization.

[0145] During the "validation" phase, the independent validation model compares the optimized results with historical cases and physical laws to ensure the effectiveness of the improvements.

[0146] Establish a complete performance tracing database to record the detailed process and effects of each optimization, forming a mechanism for the continuous evolution of system capabilities.

[0147] In a preferred embodiment of the present invention, specifically, generating weather situation analysis results based on the updated intelligent interpretation results of the weather situation includes,

[0148] Automatically generate weather situation analysis text reports:

[0149] The design incorporates a hierarchical report generation framework, comprising four levels: a "Summary Layer" for a brief overview, a "System Layer" for a detailed description of each weather system's characteristics, an "Impact Layer" for analyzing the weather system's impact on actual weather, and a "Forecast Layer" for forecasting future trends.

[0150] A specific text template library is defined for each level, and templates are organized according to different categories, including system type, impact type, and evolution type.

[0151] A conditional text generation algorithm is used to populate the template, with conditions including system type, intensity, location, and direction of movement. A beam search is then employed to generate the optimal representation.

[0152] Post-processing rules are applied to ensure accurate terminology and fluent sentences, and a professional meteorological terminology mapping table is used for standardization.

[0153] Achieve comparative analysis of multi-model forecasts:

[0154] Automatically acquire and process forecast data from different numerical models, unifying data resolution and time intervals.

[0155] Indices of consistency among calculation models include spatial correlation coefficient, root mean square error (RMSE), and systematic bias.

[0156] Based on historical performance statistics, weight coefficients are assigned to each model to establish a comprehensive weighted evaluation system.

[0157] For key weather systems, confidence intervals are calculated, and outliers in the model are identified and analyzed.

[0158] Generate a consistency analysis report, including a forecast consistency heatmap, a key element comparison table, and a preliminary analysis of the reasons for discrepancies.

[0159] In a preferred embodiment of the present invention, specifically, the weather situation analysis results are visualized, including...

[0160] Building an intelligent visualization system:

[0161] Develop hierarchical visualization components, including weather system overview maps, system detail maps, and causal explanations. Figure 3 Each level

[0162] Automatically identify key areas and add text descriptions to avoid overlapping annotations.

[0163] The visualization focus is dynamically adjusted based on user interests and data importance. The importance score is calculated using a weighted average of information entropy and anomaly degree, with high-scoring areas being highlighted.

[0164] It supports timeline interaction, allowing users to drag along the timeline to view the evolution of weather systems, and the time accuracy can be adjusted between different scales.

[0165] The above preferred embodiments are integrated to form the overall solution as follows.

[0166] Step A: Collect and preprocess multimodal meteorological data

[0167] A1. Obtain meteorological data from multiple sources, including but not limited to: numerical weather prediction model output data (such as temperature, air pressure, wind field, precipitation forecasts, etc.), satellite cloud image data, historical weather situation text descriptions, and actual measured data from observation stations.

[0168] A2. Standardize the collected data, unifying data of different resolutions, formats, and spatiotemporal scales into a standard format. Convert numerical data into floating-point tensors to unify resolution; process satellite cloud image data into image tensors of uniform size; and vectorize text data.

[0169] A3. Perform quality control on the processed data to remove noise and outliers. Use the IQR (Inter Quartile Range) method to identify outliers and set statistical boundary values ​​as outlier judgment criteria. Mark or remove low-quality data.

[0170] Step B: Construct a multimodal causal inference neural network model

[0171] B1. Construct a unified framework combining symbolic systems and neural networks. This framework consists of three layers:

[0172] The bottom layer is the neuronal perception layer, which uses multiple feature extraction networks to process different modalities of data: multilayer perceptron (MLP) is used for numerical data; convolutional neural network (CNN) is used for image data, and ResNet architecture is used to extract features; and a pre-trained BERT encoder is used to extract features for text data.

[0173] The middle layer is the representation layer, which projects features from different modalities onto a unified representation space and uses a self-attention mechanism to capture cross-modal correlations.

[0174] The top layer is the logical reasoning layer, which combines symbolic rules and neural networks. Each weather system is represented by a set of symbolic variables, including attributes such as location, intensity, and system type. Causal links are constructed through the relationship graph between variables.

[0175] B2. Implement the interaction mechanism between neural networks and symbolic logic. The specific steps are:

[0176] Design a bidirectional interface that allows the output (probability distribution) of a neural network to be converted into a symbolic representation, while the result of the symbolic operation can be used as a conditional input to the neural network.

[0177] An attention-guided mechanism is implemented by adjusting the attention weights of the neural network using symbolic rules. For example, when a strong convective system is identified, the attention weights for radar echoes and satellite cloud imagery data are increased.

[0178] Construct differentiable logical operators, converting logical operations such as "AND", "OR", and "NOT" into differentiable functions. The "AND" operation is represented by the SoftMin function, and the "OR" operation is represented by the SoftMax function.

[0179] B3. Build cross-modal causal reasoning capabilities. For each pair of modalities Xi and Xj (i,j∈{numerical data, image data, text data}), calculate the conditional mutual information I(Xi; Xj|Z) to identify direct causal relationships, where Z is the set of other common cause variables. Employ a graph-based causal discovery algorithm, constructing a directed acyclic graph (DAG) to represent the causal relationships between variables. Edge weights are optimized using gradient descent to optimize the objective function. Where λ is the sparsity regularization coefficient, and ||G||1 is the L1 norm of the graph structure.

[0180] B4. Establish a multi-level causal reasoning structure:

[0181] The first layer is the weather system identification layer, which identifies different weather systems (such as cold fronts, warm fronts, low-pressure centers, high-pressure ridges, shear lines, etc.) as independent entities.

[0182] The second layer is the inter-system interaction layer, which establishes the interaction relationships between weather systems (such as the coordination between fronts and low-pressure systems, the relationship between cyclone development and upper-level troughs, etc.).

[0183] The third layer is the weather phenomenon forecasting layer, which establishes the causal relationship between system characteristics and the final weather phenomena (such as precipitation, strong winds, temperature changes, etc.).

[0184] These three relationships are connected using differentiable programming techniques, and temporal state changes are encoded using a bidirectional long short-term memory network (Bi-LSTM).

[0185] Step C: Automated Feedback Correction Framework Based on Large Model

[0186] C1. Construct an intelligent evaluation and feedback system:

[0187] Multiple large language models (such as GPT-4, Claude, Llama, etc.) are connected to form an evaluation cluster to achieve mutual evaluation and collaborative decision-making among models.

[0188] Design a structured evaluation protocol to standardize the output format of large models into a standard JSON structure, including fields for scoring, problem diagnosis, and improvement suggestions.

[0189] Develop an efficient model scheduling system that supports batch parallel evaluation and real-time feedback, and manage evaluation tasks through priority queues.

[0190] Establish a multi-dimensional quality assessment system that covers three core dimensions: professional accuracy, logical consistency, and analytical completeness.

[0191] C2. Implement an adaptive analysis and optimization engine:

[0192] Construct a vectorized weather system knowledge base to store the typical characteristics and standard analysis methods of various weather systems using vector embedding.

[0193] The system automatically selects the optimal analysis path and parameter configuration based on the characteristics of current meteorological data through intelligent matching algorithms.

[0194] Implement a dynamic parameter tuning mechanism, leveraging the intelligent analysis capabilities of large models to automatically adjust analysis weights and thresholds based on intermediate results.

[0195] Establish a multi-scheme comparison system and select the optimal result from multiple analysis schemes by setting objective evaluation indicators.

[0196] C3. Construct a closed-loop self-optimizing learning system:

[0197] Design an automated iterative process of "generate-evaluate-optimize-verify" to establish a complete quality closed loop.

[0198] In the "generation" phase, the core analytical model generates preliminary weather situation analysis based on multimodal data.

[0199] During the "evaluation" phase, the evaluation model cluster assesses the results through a triple-testing system:

[0200] Data consistency test: Evaluate the consistency of results across different data sources and calculate the correlation coefficients of key indicators.

[0201] Meteorological rationality verification: verifying the physical rationality of the analysis results based on meteorological principles and laws.

[0202] System coverage verification: ensuring comprehensive analysis covering key weather systems and their interactions.

[0203] During the "optimization" phase, the dedicated optimization model generates specific improvement strategies based on the evaluation results, including model parameter adjustment, analysis logic restructuring, and focus optimization.

[0204] During the "validation" phase, the independent validation model compares the optimized results with historical cases and physical laws to ensure the effectiveness of the improvements.

[0205] Establish a complete performance tracing database to record the detailed process and effects of each optimization, forming a mechanism for the continuous evolution of system capabilities.

[0206] Step D: Generate weather pattern analysis product

[0207] D1. Automatically generate weather situation analysis text reports:

[0208] The design incorporates a hierarchical report generation framework, comprising four levels: "Summary Layer" (brief overview), "System Layer" (detailed description of the characteristics of each weather system), "Impact Layer" (analysis of the impact of weather systems on actual weather), and "Forecast Layer" (future evolution trends).

[0209] Define a specific text template library for each level and organize the templates according to different categories (system type, impact type, evolution type, etc.).

[0210] The template is filled using a conditional text generation algorithm, with conditions including system type, intensity, location, and direction of movement. The best expression is generated using a cluster search.

[0211] Post-processing rules are applied to ensure accurate terminology and fluent sentences, and a professional meteorological terminology mapping table is used for standardization.

[0212] D2. Implement multi-model forecast comparative analysis:

[0213] Automatically acquire and process forecast data from different numerical models, and standardize data resolution and time intervals.

[0214] Indicators of consistency among calculation models include spatial correlation coefficient, root mean square error (RMSE), and systematic bias.

[0215] Based on historical performance statistics, weight coefficients are assigned to each model to establish a comprehensive weight evaluation system.

[0216] For key weather systems (such as frontal location, low-pressure center intensity, etc.), calculate confidence intervals and identify and analyze model outliers.

[0217] Generate a consistency analysis report, including a forecast consistency heatmap, a key element comparison table, and a preliminary analysis of the reasons for discrepancies.

[0218] D3. Construct an intelligent visualization system:

[0219] Develop hierarchical visualization components, including weather system overview maps, system detail maps, and causal explanations. Figure 3 Each level.

[0220] Implement an intelligent annotation system that automatically identifies key areas (such as frontal convergence points and convection development areas) and adds text descriptions to avoid overlapping annotations.

[0221] The visualization focus is dynamically adjusted based on user interests and data importance. The importance score is calculated using a weighted average of information entropy and anomaly, with high-scoring areas being highlighted.

[0222] It supports timeline interaction, allowing users to drag along the timeline to view the evolution of weather systems, with time accuracy adjustable across different scales.

[0223] The advantages of this invention compared to existing technologies are as follows:

[0224] A multimodal causal reasoning framework integrating symbolic systems and neural networks supports cross-modal, multi-task weather system analysis.

[0225] A differentiable causal reasoning structure was designed to represent the complex weather system relationships as a trainable causal graph structure.

[0226] Build an automated quality assessment and correction mechanism based on a large model, and improve system reliability through a closed loop of "generation-assessment-optimization-verification".

[0227] Develop natural language processing technologies for meteorology professionals to achieve accurate description and interpretation of complex weather systems.

[0228] Example 2: This invention also proposes an intelligent weather situation analysis device based on artificial intelligence multimodal causal reasoning, comprising the following:

[0229] A multimodal data acquisition and preprocessing module is used to acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data.

[0230] A causal reasoning module combining neural networks and symbolic systems is used to process the processed data through a pre-built multimodal causal reasoning neural network model to obtain intelligent interpretation results of weather conditions.

[0231] The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables.

[0232] The large-model-based automated feedback correction module is used to update the intelligent weather interpretation results by feeding back and correcting them through the large-model-based automated feedback correction mechanism.

[0233] The weather situation intelligent analysis product generation module is used to generate weather situation analysis results based on the updated weather situation intelligent interpretation results, and to visualize the weather situation analysis results.

[0234] Specific application examples of the present invention are as follows:

[0235] Intelligent analysis case of typhoon weather patterns (taking Typhoon Mangkhut in 2018 as an example)

[0236] Multimodal data acquisition and preprocessing (implementation step A):

[0237] Data acquisition (implementation step A1):

[0238] Numerical weather forecast data: ECMWF model data: horizontal resolution 0.1°×0.1°, time interval 3 hours; GFS model data: horizontal resolution 0.25°×0.25°, time interval 6 hours; CMA-TRAMS model data: horizontal resolution 0.05°×0.05°, time interval 1 hour.

[0239] Observational data: FY-4A satellite data: visible light channel (0.5km), infrared channel (2km), water vapor channel (4km); Doppler radar data: radial velocity, reflectivity factor, spectral width; automatic weather station data: wind speed, wind direction, air pressure, temperature, humidity; time resolution 1 minute.

[0240] Data normalization (implementation step A2): Spatial interpolation: Use bilinear interpolation to unify data of different resolutions to a 0.1°×0.1° grid; Time alignment: Interpolate data of different time intervals to a 1-hour time step.

[0241] Data format conversion: Numerical data is converted to float32 type tensors, image data is converted to RGB tensors of size [224,224,3], and text data is converted to 768-dimensional vectors through the BERT model.

[0242] Quality Control (Step A3): Outlier Detection: Wind Speed: Use the IQR method, setting upper and lower limits as Q1 - 1.5IQR and Q3 + 1.5IQR; Air Pressure: Use the 3σ principle to remove data exceeding the mean ± 3 standard deviations.

[0243] Data imputation: The KNN algorithm is used to impute missing observational data, and spatiotemporal interpolation methods are used to fill in missing regions of satellite data. A causal inference framework is constructed (implementation step B).

[0244] Neural Network Architecture: Numerical Data Processing: 5-layer MLP with [256, 128, 64, 32, 16 neurons per layer]; Image Data Processing: ResNet50 for feature extraction, outputting a 1024-dimensional vector.

[0245] Text data processing: Using the BERT model, the output is a 768-dimensional vector.

[0246] Symbol system design: Typhoon location: longitude, latitude, direction of movement, speed of movement; Typhoon intensity: central pressure, maximum wind speed, radius of the 7-level wind circle.

[0247] Environmental conditions: Sea surface temperature (SST): location of the 26.5℃ isotherm, intensity of the warm pool, SST gradient; Vertical wind shear: 200-850 hPa wind shear, shear direction, shear intensity; Water vapor transport: intensity of the low-level jet stream, water vapor flux, water vapor convergence zone; Upper-level divergence: 200 hPa divergence field, location of the upper-level jet stream, divergence intensity; Steering airflow: 500 hPa geopotential height field, direction of the steering airflow, intensity of the steering airflow.

[0248] Causal reasoning mechanisms: Using conditional mutual information to calculate the strength of causal relationships between variables: Sea surface temperature and typhoon intensity: calculating the correlation between the location of the 26.5℃ isotherm and changes in typhoon intensity; Vertical wind shear and typhoon structure: analyzing the relationship between wind shear intensity and typhoon eyewall structure; Water vapor transport and precipitation: calculating the causal relationship between water vapor flux and typhoon precipitation intensity; Upper-level divergence and typhoon development: analyzing the relationship between the 200hPa divergence field and changes in typhoon intensity; Steering airflow and path: calculating the correlation between the 500hPa geopotential height field and the typhoon's movement path. A directed acyclic graph (DAG) is constructed to represent the causal relationship network: Sea surface temperature (SST) → Typhoon intensity: Warm SST promotes typhoon development; Vertical wind shear → Typhoon structure: Weak wind shear is conducive to typhoon maintenance; Moisture transport → Precipitation: Strong moisture transport leads to heavy precipitation; Upper-level divergence → Typhoon intensity: Strong divergence promotes typhoon development; Guiding airflow → Path: Guiding airflow determines the direction of typhoon movement. Gradient descent is used to optimize edge weights: The weights of each factor are adjusted based on historical typhoon cases, considering the changes in the importance of each factor at different development stages, and the intensity threshold of the causal relationship is optimized.

[0249] Automated feedback correction based on a large model (implementation step C):

[0250] Intelligent evaluation system (implementation step C1):

[0251] The evaluation models are as follows: Model 1 is responsible for typhoon track forecasting and evaluation; Model 2 is responsible for typhoon intensity forecasting and evaluation; and Model 3 is responsible for precipitation forecasting and evaluation.

[0252] Evaluation index design: Track forecast: average distance error, direction error; Intensity forecast: central pressure error, maximum wind speed error; Precipitation forecast: TS score, false alarm rate, missed alarm rate.

[0253] Adaptive optimization (implementation step C2):

[0254] Knowledge base construction: Historical typhoon case library: containing 1000 typhoon cases; Typhoon path similarity calculation: using the DTW algorithm; Typhoon intensity change pattern recognition: using K-means clustering.

[0255] Parameter optimization: Track forecast weights: dynamically adjusted based on historical performance; Intensity forecast thresholds: adaptively adjusted based on environmental field characteristics; Precipitation forecast parameters: automatically optimized based on topographic features.

[0256] Closed-loop learning (implementation step C3):

[0257] Generation phase: Preliminary forecasts are generated based on multimodal data, and ensemble learning methods are used to fuse the results of multiple models.

[0258] Evaluation Phase: Data Consistency Test: Calculate the correlation coefficients between different data sources; Physical Reasonableness Test: Verify whether the forecast results conform to physical laws; Historical Similarity Test: Compare with similar historical cases.

[0259] Optimization phase:

[0260] Adjusting model parameters: Using gradient descent; updating the knowledge base: Adding new success cases; optimizing inference rules: Adjusting causal strength based on evaluation results.

[0261] Weather situation analysis product generation (implementation step D):

[0262] Automatically generate typhoon analysis reports (implementation step D1):

[0263] Summary layer: Current typhoon status: location, intensity, direction of movement; 72-hour trend: path, intensity changes, and affected area.

[0264] System layer: Typhoon structural characteristics: eye region, spiral rainbands, outer circulation; Environmental field analysis: subtropical high, westerly trough, vertical wind shear.

[0265] Impact layers: Gale impact: Range of winds at levels 7 and 10, maximum wind speed; Precipitation impact: Precipitation area, intensity, and duration; Storm surge impact: Increase in water level, affected shoreline.

[0266] Forecast layer: Track forecast: probability distribution of track over the next 72 hours; Intensity forecast: intensity trend over the next 72 hours; Precipitation forecast: precipitation area forecast over the next 72 hours.

[0267] Multi-model forecast comparative analysis (implementation step D2):

[0268] Data Acquisition and Processing: Acquire forecast data from ECMWF, GFS, and CMA-TRAMS models, and standardize data resolution and time intervals.

[0269] Consistency analysis: This involves calculating differences in track forecasts among different models, analyzing the reasons for discrepancies in intensity forecasts, and assessing the reliability of precipitation forecasts.

[0270] Comprehensive analysis: Based on historical performance, weights are assigned to each model, a path probability distribution map is generated, and forecast uncertainty analysis is provided.

[0271] Intelligent visualization presentation (implementation step D3): Typhoon path visualization: comparison of historical and forecast paths, path probability distribution map, path forecast error analysis.

[0272] Typhoon structure visualization: satellite cloud image animation, radar echo evolution, 3D wind field structure.

[0273] Impact analysis visualization: Map showing the area affected by strong winds, forecast map of precipitation areas, and forecast map of storm surge.

[0274] Example 2: Case Study of Intelligent Analysis of Rainstorm Weather Situation (Taking the Rainstorm Process in South China Before the Flood Season in 2023 as an Example)

[0275] Multimodal data acquisition and preprocessing (implementation step A):

[0276] Data acquisition (implementation step A1):

[0277] Numerical weather forecast data: ECMWF model data: horizontal resolution 0.1°×0.1°, time interval 3 hours; GFS model data: horizontal resolution 0.25°×0.25°, time interval 6 hours; CMA-GFS model data: horizontal resolution 0.1°×0.1°, time interval 1 hour.

[0278] Observational data: Radar network data: reflectivity factor, radial velocity, liquid water content; Automatic weather station data: precipitation, temperature, humidity, air pressure; Radiosonde data: temperature, humidity, vertical wind field profile.

[0279] Data normalization (implementation step A2): Spatial interpolation: Use Kriging interpolation to unify data of different resolutions to a 0.1°×0.1° grid; Time alignment: Interpolate data of different time intervals to a 1-hour time step.

[0280] Data format conversion: Numerical data is converted to a float32 type tensor, radar data is converted to a tensor of size [256, 256, 3], and sounding data is converted to a vertical profile vector.

[0281] Quality control (implementation step A3):

[0282] Outlier detection: Precipitation: IQR method was used, with upper and lower limits set as Q1 - 1.5IQR and Q3 + 1.5IQR. Temperature: The 3σ rule was used to remove data exceeding the mean ± 3 standard deviations.

[0283] Data imputation: Spatiotemporal interpolation methods are used to fill in missing areas of radar data, and linear interpolation is used to fill in missing layers of sounding data.

[0284] Construction of the causal reasoning framework (implementation step B):

[0285] Neural Network Architecture: Numerical Data Processing: 5-layer MLP, with [256, 128, 64, 32, 16 neurons per layer]; Radar Data Processing: U-Net is used to extract features, outputting a 512-dimensional vector; Radiosonde Data Processing: LSTM is used to process the vertical profile, outputting a 256-dimensional vector.

[0286] Symbol system design: Low-level jet stream: location, intensity, water vapor transport; Shear line: location, intensity, direction of movement.

[0287] Topography: altitude, slope, aspect

[0288] Water vapor conditions: Water vapor transport: intensity of the low-level jet stream, water vapor flux, water vapor convergence zone; Water vapor convergence: intensity of lower-level convergence, location of the water vapor convergence center; Vertical water vapor distribution: total precipitable water volume, vertical water vapor profile.

[0289] Unstable conditions: Convective Available Potential Energy (CAPE): CAPE value, CAPE vertical distribution; Lifting Index (LI): LI value, LI spatial distribution; Vertical wind shear: 0-6km wind shear, shear direction.

[0290] Thermal conditions: Temperature field: lower-level temperature, temperature advection, temperature gradient; Humidity field: specific humidity, relative humidity, dew point temperature; Thermal instability: pseudo-equivalent potential temperature, potential instability.

[0291] Dynamic conditions: Lower-level convergence: 850 hPa divergence, convergence center location; Upper-level divergence: 200 hPa divergence, divergence center location; Vertical motion: vertical velocity, upward motion region.

[0292] Causal Reasoning Mechanisms: Using Granger causality tests to identify causal relationships in time series: Water vapor transport and precipitation: analyzing the causal relationship between changes in water vapor flux and precipitation intensity; Unstable conditions and convection: examining the relationship between changes in CAPE and convection development; Thermal conditions and precipitation: analyzing the relationship between changes in the temperature field and precipitation intensity; Dynamic conditions and precipitation: examining the relationship between vertical motion and precipitation intensity.

[0293] Constructing a Bayesian network to represent causal relationships: Water vapor conditions → Precipitation: Sufficient water vapor is a necessary condition for precipitation; Instability conditions → Convection: Strong instability promotes convection development; Thermal conditions → Precipitation: Thermal instability promotes precipitation; Dynamic conditions → Precipitation: Strong upward motion leads to heavy precipitation; Topography → Precipitation: Orographic lifting enhances precipitation.

[0294] Optimize network parameters using the EM algorithm: Adjust the weights of each factor based on historical heavy rainfall cases, consider the differences in the importance of each factor for different precipitation types, and optimize the intensity threshold of causal relationships.

[0295] Automated feedback correction based on a large model (implementation step C):

[0296] Intelligent evaluation system (implementation step C1):

[0297] The assessment models are as follows: Model 1 is responsible for assessing the forecast area of ​​heavy rainfall; Model 2 is responsible for assessing the forecast of precipitation intensity; and Model 3 is responsible for assessing the timeliness of early warnings.

[0298] Evaluation index design: Precipitation area forecast: hit rate, false alarm rate, critical success index; Intensity forecast: accuracy of tiered precipitation, deviation score; Warning timeliness: advance warning time, warning accuracy.

[0299] Adaptive optimization (implementation step C2):

[0300] Knowledge base construction: Historical rainstorm case library: containing 500 rainstorm events; Rainstorm system configuration similarity calculation: using cosine similarity; Precipitation area pattern recognition: using DBSCAN clustering.

[0301] Parameter optimization: Landing area forecast threshold: dynamically adjusted based on terrain features; intensity forecast parameters: adaptively adjusted based on moisture conditions; warning lead time parameters: automatically optimized based on system movement speed.

[0302] Closed-loop learning (implementation step C3):

[0303] Generation phase: Preliminary forecasts are generated based on multi-source data, and uncertainties are fused using Bayesian methods.

[0304] Evaluation Phase: Data Consistency Verification: Verifying the coordination between different data sources; Physical Reasonableness Verification: Checking whether the forecast results conform to physical laws; System Coverage Verification: Ensuring that critical systems are considered.

[0305] Optimization phase: Adjusting model parameters: using stochastic gradient descent; updating the knowledge base: adding new success cases; optimizing inference rules: adjusting causal strength based on evaluation results.

[0306] Weather situation analysis product generation (implementation step D):

[0307] Automatically generate a rainstorm analysis report (implementation step D1):

[0308] Abstract Layer: Overview of the Rainstorm Process: Period of Impact, Main Areas of Impact, Key Influencing Systems: Low-Level Jet, Shear Line, Topography

[0309] System layer: Factors affecting system configuration: system location, intensity, and evolution; Water vapor transport analysis: water vapor sources, transport channels, and convergence zones.

[0310] Impact layer: Precipitation characteristics: intensity, location, duration; potential impacts: flash floods, geological disasters, urban flooding.

[0311] Forecast Layer: Precipitation Trend: Evolution of precipitation over the next 24 hours; Extreme Precipitation: Potential periods and areas of extreme precipitation.

[0312] Multi-model forecast comparative analysis (implementation step D2):

[0313] Data Acquisition and Processing: Acquire forecast data from ECMWF, GFS, and CMA-GFS models, and standardize data resolution and time intervals.

[0314] Consistency analysis: Discrepancies in precipitation forecasts among different calculation models are analyzed to determine the causes of these discrepancies and assess forecast reliability.

[0315] Comprehensive analysis: Based on historical performance, weights are assigned to each model to generate a precipitation probability forecast map, providing forecast uncertainty analysis and intelligent visualization (implementation step D3): Precipitation data visualization: precipitation distribution map, precipitation intensity evolution map, cumulative precipitation map; Influencing system visualization: low-level jet distribution map, shear line location map, water vapor transport channel map; Forecast product visualization: precipitation area forecast map, precipitation intensity forecast map, extreme precipitation probability map.

[0316] Although the description of the invention has been quite detailed and particularly of several described embodiments, it is not intended to limit it to any of these details or embodiments or any particular embodiment, but should be considered as providing a broad possible interpretation of the claims by referring to the appended claims and taking into account the prior art, thereby effectively covering the intended scope of the invention. Furthermore, the invention has been described above with respect to embodiments foreseeable by the inventors in order to provide a useful description, and non-substantial modifications to the invention that have not yet been foreseen may still represent equivalent modifications.

[0317] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. Any embodiment that achieves the technical effects of the present invention using the same means should fall within the protection scope of the present invention. Within the protection scope of the present invention, various modifications and variations can be made to the technical solutions and / or implementation methods.

Claims

1. A weather situation intelligent analysis method based on artificial intelligence multimodal causal reasoning, characterized in that, Including the following: Acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data; The pre-constructed multimodal causal reasoning neural network model is used to process the processed data to obtain intelligent interpretation results of the weather situation; The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables. The intelligent interpretation results of the weather situation are updated by using an automated feedback correction mechanism based on a large model to correct the errors. Based on the updated intelligent interpretation of the weather situation, a weather situation analysis result is generated and the weather situation analysis result is visualized.

2. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 1, characterized in that, Specifically, the acquired multimodal meteorological data includes, Numerical weather forecast model output data includes temperature, air pressure, wind field, precipitation forecasts, and other data, such as satellite cloud imagery, historical weather pattern descriptions, and measured data from observation stations.

3. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 2, characterized in that, Specifically, the multimodal meteorological data is preprocessed to obtain processed data, including: The multimodal meteorological data is standardized by unifying data of different resolutions, formats, and spatiotemporal scales into a standard format, converting numerical data into floating-point tensors, unifying resolution, processing satellite cloud image data into image tensors of uniform size, and vectorizing text data. Next, quality control is carried out to remove noise and outliers. The IQR method is used to identify outliers in the data, and statistical boundary values ​​are set as outlier judgment criteria. Low-quality data is marked or removed to obtain processed data.

4. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 1, characterized in that, Specifically, the interaction mechanism combining symbolic rules and neural networks includes, Design a bidirectional interface that allows the output of the neural network, i.e., the probability distribution, to be converted into a symbolic representation, while the result of the symbolic operation can be used as a conditional input to the neural network; Implement an attention guidance mechanism by using symbolic rules to adjust the attention weights of the neural network; Construct differentiable logical operators, converting "AND", "OR", and "NOT" logical operations into differentiable functions. The "AND" operation is represented using the SoftMin function, and the "OR" operation using the SoftMax function. Specifically, the differentiable logical operators are constructed as follows: The AND operation in SoftMin is represented as: A∧B≈-log(e^( ... -A +e -B ); The OR operation in SoftMax is represented as: A∨B≈log(e^(-1 / 2)) -A +e -B ); The NOT operation is represented as: Where A,B∈[0,1], represents the probabilistic representation of logical variables. The above functions guarantee differentiability in neural networks, thereby enabling end-to-end training. Cross-modal causal inference rules are constructed. For each pair of modalities Xi and Xj, i,j∈{numerical data, image data, text data}, conditional mutual information I(Xi; Xj|Z) is calculated to identify direct causal relationships, where Z is the set of other common cause variables. A graph-based causal discovery algorithm is adopted, constructing a directed acyclic graph (DAG) to represent the causal relationships between variables. The edge weights are obtained by optimizing the objective function through gradient descent. The optimization objective function is defined as: in: This represents the multimodal feature representation of the input; The reconstructed features are inferred from the causal graph generation model G; ‖G‖1: The L1 norm of the adjacency matrix of the causal graph, introducing sparsity constraints; λ: A hyperparameter used to adjust the tradeoff between reconstruction error and graph sparsity; Establish a multi-level causal reasoning structure: The first layer is the weather system identification layer, which identifies different weather systems as independent entities; The second layer is the inter-system interaction layer, which establishes the interaction relationships between weather systems; The third layer is the weather phenomenon forecasting layer, which establishes the causal relationship between system characteristics and the final weather phenomenon; These three relationships are connected using differentiable programming techniques, and the temporal state changes are encoded using a bidirectional long short-term memory network (Bi-LSTM).

5. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 3, characterized in that, Specifically, automated feedback correction mechanisms based on large models include, Build an intelligent assessment and feedback system. By integrating multiple large language models to form an evaluation cluster, mutual evaluation and collaborative decision-making among models can be achieved. Design a structured evaluation protocol to standardize the output format of large models into a standard JSON structure. The output format includes fields for score, problem diagnosis, and improvement suggestions. Develop an efficient model scheduling system that supports batch parallel evaluation and real-time feedback, and manage evaluation tasks through priority queues. Establish a multi-dimensional quality assessment system, covering three core dimensions: professional accuracy, logical consistency, and analytical completeness; Implement an adaptive analysis and optimization engine. A vectorized weather system knowledge base is constructed, which stores the typical characteristics and standard analysis methods of various weather systems using vector embedding. Using an intelligent matching algorithm, the optimal analysis path and parameter configuration are automatically selected based on the characteristics of current meteorological data. This mechanism implements dynamic parameter tuning, leveraging the intelligent analysis capabilities of large models to automatically adjust analysis weights and thresholds based on intermediate results. A multi-scheme comparison system is established, and the optimal result is selected from multiple analysis schemes by setting objective evaluation indicators. Construct a closed-loop self-optimizing learning system. Design an automated iterative process of "generation-evaluation-optimization-verification" to establish a complete quality closed loop. In the "generation" phase, the core analytical model generates preliminary weather situation analysis based on multimodal data. During the "evaluation" phase, the evaluation model cluster assesses the results through a triple-testing system: Data consistency test: Evaluate the consistency of results across different data sources and calculate the correlation coefficients of key indicators. Meteorological rationality verification: Verifying the physical rationality of the analysis results based on meteorological principles and laws. System coverage verification: Ensure that the analysis fully covers key weather systems and their interactions. During the "optimization" phase, the dedicated optimization model generates specific improvement strategies based on the evaluation results, including model parameter adjustments, analysis logic restructuring, and focus optimization. During the "validation" phase, the independent validation model compares the optimized results with historical cases and physical laws to ensure the effectiveness of the improvements. Establish a complete performance tracing database to record the detailed process and effects of each optimization, forming a mechanism for the continuous evolution of system capabilities.

6. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 1, characterized in that, Specifically, based on the updated intelligent interpretation of the weather situation, weather situation analysis results are generated, including: Automatically generate weather situation analysis text reports: The design incorporates a hierarchical report generation framework, comprising four levels: a "Summary Layer" for a brief overview, a "System Layer" for a detailed description of each weather system's characteristics, an "Impact Layer" for analyzing the weather system's impact on actual weather, and a "Forecast Layer" for forecasting future trends. A specific text template library is defined for each level, and templates are organized according to different categories, including system type, impact type, and evolution type. A conditional text generation algorithm is used to populate the template, with conditions including system type, intensity, location, and direction of movement. A beam search is then employed to generate the optimal representation. Post-processing rules are applied to ensure accurate terminology and fluent sentences, and a professional meteorological terminology mapping table is used for standardization. Achieve comparative analysis of multi-model forecasts: Automatically acquire and process forecast data from different numerical models, unifying data resolution and time intervals. Indices of consistency among calculation models include spatial correlation coefficient, root mean square error (RMSE), and systematic bias. Based on historical performance statistics, weight coefficients are assigned to each model to establish a comprehensive weighted evaluation system. For key weather systems, confidence intervals are calculated, and outliers in the model are identified and analyzed. Generate a consistency analysis report, including a forecast consistency heatmap, a key element comparison table, and a preliminary analysis of the reasons for discrepancies.

7. The intelligent weather situation analysis method based on artificial intelligence multimodal causal reasoning according to claim 1, characterized in that, Specifically, the results of the weather situation analysis are visualized, including: Building an intelligent visualization system: Develop a hierarchical visualization component, including three levels: a weather system overview map, a system detail map, and a cause-and-effect explanation diagram. Automatically identify key areas and add text descriptions to avoid overlapping annotations. The visualization focus is dynamically adjusted based on user interests and data importance. The importance score is calculated using a weighted average of information entropy and anomaly degree, with high-scoring areas being highlighted. It supports timeline interaction, allowing users to drag along the timeline to view the evolution of weather systems, and the time accuracy can be adjusted between different scales.

8. A weather situation intelligent analysis device based on artificial intelligence multimodal causal reasoning, characterized in that, Including the following: A multimodal data acquisition and preprocessing module is used to acquire multimodal meteorological data and preprocess the multimodal meteorological data to obtain processed data. A causal reasoning module combining neural networks and symbolic systems is used to process the processed data through a pre-built multimodal causal reasoning neural network model to obtain intelligent interpretation results of weather conditions. The specific architecture of the multimodal causal reasoning neural network model is as follows: the bottom layer is the neuron perception layer, which is used to process different modal data in the input data using multiple feature extraction networks; the middle layer is the representation layer, which is used to project different modal features onto a unified representation space and use a self-attention mechanism to capture cross-modal correlations; and the top layer is the logical reasoning layer, which combines symbolic rules and neural networks to represent each weather system with a set of symbolic variables and construct causal links through the relationship graph between symbolic variables. The large-model-based automated feedback correction module is used to update the intelligent weather interpretation results by feeding back and correcting them through the large-model-based automated feedback correction mechanism. The weather situation intelligent analysis product generation module is used to generate weather situation analysis results based on the updated weather situation intelligent interpretation results, and to visualize the weather situation analysis results.

Images