Extreme weather prediction method and system

Abstract

The invention discloses an extreme weather prediction method and system, and relates to the technical field of artificial intelligence. The method comprises the following steps: collecting multi-modal meteorological observation data; performing space-time alignment and quality correction to obtain an assimilation data set; adjusting a feature extraction weight based on a quality identifier, and projecting the assimilation data set to a low-dimensional hidden space to obtain a hidden space feature vector; based on the sample distance measurement of the hidden space, adopting a meta-learning framework to carry out multi-task prediction on the feature vector of the hidden space; performing multi-target collaborative optimization and antagonism training on the neural network by using the prediction result to obtain a prediction model; and verifying the prediction result by using a digital twin simulation environment to obtain a verification result, and updating model parameters of the prediction model. According to the method, the problem of insufficient prediction precision caused by a complex coupling relationship of meteorological factors and data isomerism is effectively solved, and the accuracy and reliability of extreme weather evolution trend prediction are improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for predicting extreme weather. Background Technology

[0002] Extreme weather evolution trend prediction technology is based on the assimilation and processing of multi-source meteorological observation data, combined with high-resolution simulation of numerical weather prediction models. It describes atmospheric dynamic and thermodynamic processes through physical parameterization schemes, thereby inferring the evolution trend of future extreme weather events. In power meteorological forecasting systems, the model output is statistically calibrated to generate probabilistic forecast products, providing key information for power grid operation, such as predicting the potential impact of typhoons and severe convective weather on transmission facilities and power loads, and helping to optimize dispatch decisions.

[0003] Among related technologies, the technology for predicting the evolution trend of extreme weather has the following technical pain points: Due to the complex nonlinear coupling relationship between meteorological factors, such as the interaction between atmospheric circulation and topography and ocean thermal, as well as the heterogeneity of multi-source data, including the differences in format and resolution between satellite remote sensing, ground observation and ocean buoy data, it is difficult for neural network models based on multi-task learning to efficiently extract key features and accurately predict the evolution trend. For example, in the scenario of power meteorological forecasting system, when predicting the impact of cold waves on wind power output, the model fails to fully integrate the spatiotemporal dependence characteristics of multimodal data such as wind speed, temperature drop and air pressure, and incorrectly estimates the duration and spatial range of cold waves, thereby causing deviations in the assessment of power grid resilience and affecting the reliability of dispatching strategies. Summary of the Invention

[0004] This disclosure provides a method and system for predicting extreme weather, in order to at least solve the above-mentioned technical problems existing in the prior art.

[0005] According to a first aspect of this application, a method for predicting extreme weather is provided, comprising: Collect multimodal meteorological observation data; the meteorological observation data includes temperature, air pressure, humidity, wind speed and precipitation; The meteorological observation data is spatiotemporally aligned and quality corrected to obtain an assimilated dataset; wherein, a quality identifier is generated through quality correction, and the quality identifier is used to record the location of data anomalies and the magnitude of correction. Based on the quality identifier, the feature extraction weights are adjusted, and the assimilated dataset is projected into a low-dimensional latent space to obtain the latent space feature vector. Based on the sample distance metric of latent space, a meta-learning framework is used to perform multi-task prediction on the latent space feature vectors to generate prediction results; the prediction results include the probability of extreme weather occurrence, the scope of impact, and the duration. The prediction results are used to perform multi-objective collaborative optimization and adversarial training on the neural network to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration and computational efficiency. The contribution of the latent space feature vectors in the prediction model is quantified, and the prediction results are verified using a digital twin simulation environment to obtain the verification results. Based on the verification results and real-time meteorological observation data, the model parameters of the prediction model are updated.

[0006] In one possible implementation, the meteorological observation data includes ground station data and satellite remote sensing data; the spatiotemporal alignment of the meteorological observation data includes: According to pre-set configuration rules, corresponding confidence weights are assigned to meteorological observation data of each modality; wherein, the confidence weight of the ground station data is higher than that of the satellite remote sensing data; Based on the confidence weights, a weighted dynamic time warping algorithm is used to align the time series of each modal meteorological observation data with the ground station data time as the reference, and time-aligned data is obtained. The time-aligned data is interpolated to a latitude and longitude grid to give it spatial coordinate identification, thus obtaining standardized data; the standardized data is used as input data for quality correction.

[0007] In one possible implementation, the quality correction of the meteorological observation data includes: The generator receives the standardized data and initial quality identifier, and outputs the correction data. A data quality assessment matrix is ​​generated by comparing the distribution differences between the corrected data and the standard dataset using a discriminator, and the data quality assessment matrix is ​​then fed back to the generator. The generator uses the data quality assessment matrix to iteratively correct anomalous data in the correction data and updates the quality identifier after correcting the anomalous data.

[0008] In one possible implementation, projecting the assimilated dataset into a low-dimensional latent space to obtain latent space feature vectors includes: The spatial features of the assimilation dataset are extracted using a graph attention network, and the multi-scale temporal features of the assimilation dataset are extracted using a temporal convolutional network. The spatial features and the multi-scale temporal features are fused to obtain fused features; The distribution of latent space vectors in the fused features is determined by using contrastive learning constraints, thereby obtaining latent space feature vectors.

[0009] In one possible implementation, the latent space-based sample distance metric employs a meta-learning framework to perform multi-task prediction on the latent space feature vectors to generate prediction results, including: A training set is constructed based on the latent space feature vectors; the training set includes a support set and a query set, and the support set includes typical extreme weather scenarios. Based on the latent space feature vectors, calculate the sample distance metric; Based on the sample distance metric, a meta-learning framework is used to perform a small number of gradient descent steps on the support set to determine the model parameters. The model performance is evaluated on the query set based on the model parameters, and the model parameters are optimized. The optimized model parameters are then used to obtain prediction results on the probability, scope of impact, and duration of extreme weather events.

[0010] In one possible implementation, the step of using the prediction results to perform multi-objective collaborative optimization and adversarial training on the neural network to obtain a prediction model includes: A multi-objective optimization problem is established; the multi-objective optimization problem includes an objective function of the inverse of the prediction error, an objective function of the expected calibration error, and an objective function of the inverse of the inference speed. The multi-objective optimization problem is solved using a non-dominated sorting genetic algorithm, and adversarial examples are introduced during the solution process to select the optimal solution, thus obtaining the non-dominated sorting result. The optimal model parameters are selected based on the non-dominated ranking results to obtain the prediction model.

[0011] In one possible implementation, quantifying the contribution of latent space feature vectors in the prediction model includes: Extract the mapping relationship between the latent space feature vectors of the input and the prediction results from the prediction model; The marginal contribution of each input latent space feature vector to the prediction result is calculated using the Shapley sum interpretation method. The importance of features is ranked based on the marginal contribution values. Based on the ranking of feature importance, the contribution of latent space feature vectors is determined.

[0012] In one possible implementation, verifying the prediction results using a digital twin simulation environment to obtain verification results includes: Construct a digital twin environment based on a numerical weather forecast model and set physical parameters and boundary conditions; The prediction results output by the prediction model are used as the initial field perturbation term input to the digital twin environment; Run the digital twin environment to obtain probabilistic forecast results; Calculate the degree of difference between the probabilistic forecast result and the prediction result; The physical plausibility is assessed based on the degree of difference, and the verification results are obtained.

[0013] In one possible implementation, updating the model parameters of the prediction model based on the verification results and real-time meteorological observation data includes: Calculate the importance weights of the model parameters of the prediction model to historical tasks; A regularization term is constructed based on the importance weights, and the regularization term is added to the loss function to constrain the change in the model parameters during the update process; Based on the verification results, the sample selection criteria for incremental learning are determined, and training samples are selected from the real-time meteorological observation data collected in real time in combination with the importance ranking. The prediction model is incrementally trained using the training samples to update the model parameters.

[0014] Real-time monitoring of the prediction error of the prediction model on the validation set; A dynamic threshold is set, and when the prediction error is detected to continuously exceed the dynamic threshold, the model retraining process is initiated. In the retraining process, samples corresponding to high-importance features are selected according to the importance ranking, and a training set is constructed by combining the collected real-time weather data. The prediction model is retrained using the training set, and the model update cycle is adjusted according to the error trend in the validation phase.

[0015] According to a second aspect of this application, an extreme weather prediction system is provided, comprising: The data acquisition module is used to collect multimodal meteorological observation data, including temperature, air pressure, humidity, wind speed, and precipitation. The processing module is used to perform spatiotemporal alignment and quality correction on the meteorological observation data to obtain an assimilated dataset; wherein, a quality identifier is generated through quality correction, and the quality identifier is used to record the location of data anomalies and the magnitude of correction; The extraction module is used to adjust the feature extraction weights based on the quality identifier, project the assimilation dataset into a low-dimensional latent space, and obtain the latent space feature vector. The prediction module is used to perform multi-task prediction on the latent space feature vector based on the sample distance metric of the latent space, and generate prediction results; the prediction results include the probability of extreme weather occurrence, the scope of impact, and the duration. An optimization module is used to perform multi-objective collaborative optimization and adversarial training on the neural network using the prediction results to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration, and computational efficiency. The verification module is used to quantify the contribution of the latent space feature vectors in the prediction model and verify the prediction results using a digital twin simulation environment to obtain the verification results. The update module is used to update the model parameters of the prediction model based on the verification results and real-time meteorological observation data.

[0016] According to a third aspect of this application, an electronic device is provided, comprising: A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in any of the above embodiments.

[0017] According to a fourth aspect of this application, a non-transitory computer-readable storage medium is provided storing computer instructions for causing the computer to perform the methods described in this application.

[0018] According to a fifth aspect of this application, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method described in this application.

[0019] This application utilizes a multimodal meteorological observation data spatiotemporal alignment and quality correction step to effectively handle the heterogeneity of multi-source data in terms of format and resolution, generating standardized assimilation datasets and quality identifiers, thus improving data quality from the source. Latent space projection technology uses graph attention networks and temporal convolutional networks to extract spatial and temporal features, which are then fused and compared to obtain latent space feature vectors representing the complex coupling relationships of meteorological elements, overcoming the challenge of modeling nonlinear interactions. A meta-learning framework rapidly adapts model parameters based on sample distance metrics in the latent space, enhancing the generalization prediction capability for rare extreme weather events. Multi-objective collaborative optimization and adversarial training use a non-dominated sorting genetic algorithm to balance prediction accuracy, model calibration, and computational efficiency, improving model robustness. Interpretability analysis quantifies the contribution of latent space feature vectors, and a digital twin simulation environment verifies the physical rationality of the prediction results, enhancing decision-making transparency. An incremental learning mechanism dynamically updates the model based on verification results and real-time data, and uses elastic weights to consolidate the algorithm and maintain historical knowledge, achieving continuous optimization of the prediction system. The various technical aspects form a closed-loop process, collaboratively addressing the challenges posed by data heterogeneity and complex coupling relationships, and significantly improving the accuracy and reliability of extreme weather evolution trend prediction.

[0020] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily apparent from the following detailed description taken in conjunction with the accompanying drawings. Several embodiments of this application are illustrated in the drawings by way of example and not limitation, in which: In the accompanying drawings, the same or corresponding reference numerals indicate the same or corresponding parts.

[0022] Figure 1 A flowchart illustrating the extreme weather prediction method in an embodiment of this application is shown; Figure 2 A schematic diagram of the structure of the extreme weather prediction system in an embodiment of this application is shown; Figure 3 A schematic diagram of the composition structure of the electronic device in an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, features, and advantages of this application more apparent and understandable, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] In the following description, the terms "first" and "second" are used merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first" and "second" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0026] The following section, with reference to the accompanying drawings, describes a method and system for predicting extreme weather provided in this application.

[0027] like Figure 1 As shown, this application provides a method for predicting extreme weather, including: S101, Collect multimodal meteorological observation data; the meteorological observation data includes temperature, air pressure, humidity, wind speed and precipitation.

[0028] In this application, multi-modal meteorological observation data, including temperature, air pressure, humidity, wind speed, and precipitation, are first collected using multi-source meteorological observation equipment. The data originate from various sources, such as ground stations, satellite remote sensing platforms, and radiosonde equipment, and have different sampling frequencies and spatial resolutions. The collected meteorological observation data provides the raw data input for subsequent processing, ensuring comprehensive and diverse data coverage.

[0029] S102, perform spatiotemporal alignment and quality correction on the meteorological observation data to obtain an assimilated dataset; wherein, a quality identifier is generated through quality correction, and the quality identifier is used to record the location of data anomalies and the correction magnitude.

[0030] This application performs spatiotemporal alignment and quality correction on collected multimodal meteorological observation data. Spatiotemporal alignment employs a weighted dynamic time warping algorithm, aligning the time series of data from other sources with the ground station data time as the baseline, based on the confidence weights of different data sources. The aligned data is then interpolated to a unified latitude and longitude grid using bilinear interpolation, generating a standardized data cube with timestamps and spatial coordinates. Quality correction is performed through a conditional generative adversarial network (GAN). The generator receives the standardized data cube and an initial quality identifier, outputting the corrected data; the discriminator compares the distribution differences between the corrected data and historical high-quality datasets, generating a data quality assessment matrix. The data quality assessment matrix is ​​fed back to the generator, iteratively correcting data anomalies and updating the quality identifier, thereby obtaining a standardized assimilated dataset. The quality identifier records the location and correction magnitude of data anomalies, providing quality guidance for subsequent feature extraction.

[0031] S103, adjust the feature extraction weights based on the quality identifier, project the assimilation dataset into a low-dimensional latent space, and obtain the latent space feature vector.

[0032] Specifically, the assimilation dataset is projected into a low-dimensional latent space. The projection process uses a graph attention network to extract spatial features from the assimilation dataset, where the calculation of attention weights between nodes considers both terrain elevation differences and distances between nodes. Simultaneously, a temporal convolutional network extracts multi-scale temporal features from the assimilation dataset. After fusing spatial and temporal features, contrastive learning is used to constrain the latent space vector distribution, causing samples with similar weather evolution patterns to cluster in the latent space, thus obtaining latent space feature vectors. The projection process adjusts the feature extraction weights based on quality identifiers to ensure the accuracy and robustness of feature extraction.

[0033] S104, Based on the sample distance metric of the latent space, a meta-learning framework is used to perform multi-task prediction on the latent space feature vector to generate prediction results; the prediction results include the probability of extreme weather occurrence, the scope of impact, and the duration.

[0034] This application employs a meta-learning framework for multi-task prediction of latent space feature vectors. The framework samples a support set and a query set from the training set; the support set includes typical extreme weather scenarios. A sample distance metric is calculated based on the latent space feature vectors, and a small number of gradient descent steps are performed on the support set using this metric to quickly adapt the model parameters. The adapted model parameters are then used to evaluate the model performance on the query set, and the initial parameters are optimized to enable the model to generalize across tasks, generating predictions of the probability of occurrence, the scope of impact, and the duration of extreme weather events. The meta-learning framework utilizes the sample distance metric in the latent space for prediction task adaptation, improving the forecasting capability for rare events.

[0035] S105, using the prediction results to perform multi-objective collaborative optimization and adversarial training on the neural network to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration and computational efficiency.

[0036] In this application, the generated prediction results undergo multi-objective collaborative optimization and adversarial training. The multi-objective collaborative optimization establishes a multi-objective optimization problem, including three objective functions: the reciprocal of the prediction error, the expected calibration error, and the reciprocal of the inference speed. A non-dominated sorting genetic algorithm is used to solve the multi-objective optimization problem, selecting the Pareto optimal solution through non-dominated sorting and crowding calculation. Adversarial example training is introduced during the optimization process; adversarial examples are generated by perturbing the input features and added to the training data to enhance model robustness. The optimal model parameters are selected based on the non-dominated sorting results to obtain the prediction model. Multi-objective collaborative optimization balances prediction accuracy, model calibration, and computational efficiency, ensuring comprehensive model performance.

[0037] S106, quantify the contribution of the latent space feature vector in the prediction model, and verify the prediction results using a digital twin simulation environment to obtain the verification results.

[0038] In this application, interpretability analysis is employed to quantify the contribution of latent space feature vectors in the prediction model. The interpretability analysis method extracts the mapping relationship between input features and prediction results from the prediction model, and uses the Shapley sum interpretation method to calculate the marginal contribution value of each input feature to the prediction result. Feature importance ranking is generated based on the magnitude of the marginal contribution values, and this ranking is output to the incremental learning step. Simultaneously, the physical rationality of the prediction results is verified using a digital twin simulation environment. This digital twin simulation environment is constructed based on a WRF numerical model, using the prediction results as initial field perturbation inputs to run ensemble forecasts and obtain probabilistic forecast results. The difference between the probabilistic forecast results and the prediction results is calculated, and the physical rationality is evaluated based on the difference, obtaining the verification results. Interpretability analysis provides guidance on feature importance, and digital twin verification enhances the credibility of the prediction results.

[0039] S107, Based on the verification results and real-time meteorological observation data, update the model parameters of the prediction model.

[0040] Finally, based on the validation results and real-time meteorological observation data, the prediction model is dynamically updated through incremental learning. Incremental learning employs an elastic weight consolidation algorithm to calculate the importance weights of the prediction model parameters to historical tasks. A regularization term is constructed based on these importance weights and added to the loss function to constrain changes in important parameters. The sample selection criteria for incremental learning are determined based on the validation results, and training samples are selected from real-time meteorological observation data by combining feature importance ranking. The selected training samples are used to incrementally train the prediction model and update its parameters. Simultaneously, the prediction error of the prediction model on the validation set is monitored in real time, and a dynamic threshold is set. When the prediction error continuously exceeds the dynamic threshold, the model retraining process is initiated. In the retraining process, samples corresponding to high-importance features are selected based on feature importance ranking, and a training set is constructed using recent data. The prediction model is retrained using this training set, and the model update cycle is adjusted based on the error trend observed in the validation phase. Incremental learning uses interpretability analysis results to guide the update strategy, achieving continuous optimization and adaptive evolution of the prediction model.

[0041] This application effectively addresses the heterogeneity of multi-source data in terms of format and resolution through spatiotemporal alignment and quality correction steps of multimodal meteorological observation data, generating standardized assimilation datasets and quality identifiers to improve data quality from the source. Latent space projection technology utilizes graph attention networks and temporal convolutional networks to extract spatial and temporal features, which are then fused and compared to obtain latent space feature vectors representing the complex coupling relationships of meteorological elements, overcoming the challenge of modeling nonlinear interactions. A meta-learning framework rapidly adapts model parameters based on sample distance metrics in latent space, enhancing the generalization prediction capability for rare extreme weather events. Multi-objective collaborative optimization and adversarial training balance prediction accuracy, model calibration, and computational efficiency through a non-dominated sorting genetic algorithm, improving model robustness. Interpretability analysis quantifies the contribution of latent space feature vectors, and a digital twin simulation environment verifies the physical rationality of the prediction results, enhancing decision-making transparency. An incremental learning mechanism dynamically updates the model based on verification results and real-time data, and uses elastic weights to consolidate the algorithm and maintain historical knowledge, achieving continuous optimization of the prediction system. These technical aspects form a closed-loop process, collaboratively addressing the challenges posed by data heterogeneity and complex coupling relationships, significantly improving the accuracy and reliability of extreme weather evolution trend predictions.

[0042] In some embodiments, the meteorological observation data includes ground station data and satellite remote sensing data; the spatiotemporal alignment of the meteorological observation data includes: According to pre-set configuration rules, corresponding confidence weights are assigned to meteorological observation data of each modality; wherein, the confidence weight of the ground station data is higher than that of the satellite remote sensing data; Based on the confidence weights, a weighted dynamic time warping algorithm is used to align the time series of each modal meteorological observation data with the ground station data time as the reference, and time-aligned data is obtained. The time-aligned data is interpolated to a latitude and longitude grid to give it spatial coordinate identification, thus obtaining standardized data; the standardized data is used as input data for quality correction.

[0043] In this application, the reliability weights for spatiotemporal alignment are based on the measurement accuracy and reliability of the data sources. Ground station data receives higher weights due to direct contact measurements, while satellite remote sensing data receives lower weights due to indirect inversion. The weighted dynamic time warping algorithm uses the time series of ground station data as a benchmark and calculates the time offsets of other source data through dynamic programming to achieve time series alignment. The aligned data is then interpolated to a unified latitude and longitude grid using bilinear interpolation, generating a standardized data cube with timestamps and spatial coordinate identifiers. This step resolves the temporal and spatial inconsistencies of multi-source data and provides a standardized input format for subsequent quality correction.

[0044] In some embodiments, the quality correction of the meteorological observation data includes: The generator receives the standardized data and initial quality identifier, and outputs the correction data. A data quality assessment matrix is ​​generated by comparing the distribution differences between the corrected data and the standard dataset using a discriminator, and the data quality assessment matrix is ​​then fed back to the generator. The generator uses the data quality assessment matrix to iteratively correct anomalous data in the correction data and updates the quality identifier after correcting the anomalous data.

[0045] In this application, during the quality correction process, the generator of a conditional generative adversarial network receives a standardized data cube and an initial quality identifier. It learns the data distribution through an encoder-decoder structure and outputs corrected data. The discriminator compares the corrected data with the statistical distribution of historical high-quality datasets, generating a data quality assessment matrix to quantify the differences. The generator iteratively adjusts its parameters based on the data quality assessment matrix, corrects data anomalies, and updates the anomaly locations and correction magnitudes recorded in the quality identifiers. The quality correction process utilizes an adversarial training mechanism to improve data reliability and achieve high-quality feature extraction.

[0046] In some embodiments, projecting the assimilated dataset into a low-dimensional latent space to obtain latent space feature vectors includes: The spatial features of the assimilation dataset are extracted using a graph attention network, and the multi-scale temporal features of the assimilation dataset are extracted using a temporal convolutional network. The spatial features and the multi-scale temporal features are fused to obtain fused features; The distribution of latent space vectors in the fused features is determined by using contrastive learning constraints, thereby obtaining latent space feature vectors.

[0047] In this application, during the projection to the low-dimensional latent space step, the graph attention network processes the spatial dimension of the assimilation dataset. The calculation of attention weights between nodes incorporates factors such as terrain elevation difference and Euclidean distance to capture complex spatial dependencies. The temporal convolutional network uses dilated convolutional kernels to extract multi-scale temporal features, capturing the long-term and short-term evolution patterns of meteorological data. Spatial and temporal features are fused through fully connected layers, and a contrastive learning loss function is used to constrain the distribution of latent space vectors, causing similar weather patterns to cluster in the latent space. This step achieves efficient representation of meteorological data, providing compact feature inputs for the meta-learning framework.

[0048] In some embodiments, the latent space-based sample distance metric employs a meta-learning framework to perform multi-task prediction on the latent space feature vectors to generate prediction results, including: A training set is constructed based on the latent space feature vectors; the training set includes a support set and a query set, and the support set includes typical extreme weather scenarios. Based on the latent space feature vectors, calculate the sample distance metric; Based on the sample distance metric, a meta-learning framework is used to perform a small number of gradient descent steps on the support set to determine the model parameters. The model performance is evaluated on the query set based on the model parameters, and the model parameters are optimized. The optimized model parameters are then used to obtain prediction results on the probability, scope of impact, and duration of extreme weather events.

[0049] This application provides a meta-learning framework for multi-task prediction. The support set samples typical extreme weather scenarios from the training set, while the query set includes diverse meteorological sequence samples. The sample distance metric is calculated based on the cosine similarity of latent space feature vectors, used to quickly adapt model parameters on the support set through a few steps of gradient descent. The adapted model is then evaluated on the query set, and the initial parameters are adjusted by a meta-optimizer to enhance cross-task generalization. The meta-learning framework leverages inter-task similarity to improve the model's prediction accuracy for rare extreme weather events.

[0050] In some embodiments, the step of using the prediction results to perform multi-objective collaborative optimization and adversarial training on the neural network to obtain a prediction model includes: A multi-objective optimization problem is established; the multi-objective optimization problem includes an objective function of the inverse of the prediction error, an objective function of the expected calibration error, and an objective function of the inverse of the inference speed. The multi-objective optimization problem is solved using a non-dominated sorting genetic algorithm, and adversarial examples are introduced during the solution process to select the optimal solution, thus obtaining the non-dominated sorting result. The optimal model parameters are selected based on the non-dominated ranking results to obtain the prediction model.

[0051] In this application, the multi-objective collaborative optimization and adversarial training problem defines three objective functions: the reciprocal of the prediction error, the reciprocal of the expected calibration error, and the reciprocal of the inference speed, reflecting the model's accuracy, reliability, and efficiency requirements. A non-dominated sorting genetic algorithm selects the Pareto optimal solution set through fast non-dominated sorting and crowding comparison, balancing multiple competing objectives. Adversarial example training generates adversarial examples by perturbing input features using the gradient sign method and adds them to the training data to improve model robustness. This step enhances the model's overall performance in practical applications through multi-objective decision-making and adversarial training.

[0052] In some embodiments, quantifying the contribution of latent space feature vectors in the prediction model includes: Extract the mapping relationship between the latent space feature vectors of the input and the prediction results from the prediction model; The marginal contribution of each input latent space feature vector to the prediction result is calculated using the Shapley sum interpretation method. The importance of features is ranked based on the marginal contribution values. Based on the ranking of feature importance, the contribution of latent space feature vectors is determined.

[0053] This application utilizes interpretability analysis to extract the nonlinear mapping relationship between input features and output results from a predictive model. It employs the Shapley sum interpretation method to calculate the marginal contribution value of each latent space feature vector, which is derived based on the average contribution of the feature across all possible subsets. Feature importance is ranked according to the magnitude of the marginal contribution value, highlighting key driving factors. The interpretability analysis results provide feature selection guidance for incremental learning steps, enhancing the transparency of model decisions.

[0054] In some embodiments, verifying the prediction results using a digital twin simulation environment to obtain verification results includes: Construct a digital twin environment based on a numerical weather forecast model and set physical parameters and boundary conditions; The prediction results output by the prediction model are used as the initial field perturbation term input to the digital twin environment; Run the digital twin environment to obtain probabilistic forecast results; Calculate the degree of difference between the probabilistic forecast result and the prediction result; The physical plausibility is assessed based on the degree of difference, and the verification results are obtained.

[0055] Specifically, for the validation of the digital twin simulation environment, a digital twin environment based on the WRF numerical model is constructed, and physical parameterization schemes and boundary conditions are set to simulate atmospheric dynamic processes. The prediction results output by the prediction model are used as initial field perturbation terms input into the digital twin environment, and probabilistic forecast results are obtained through ensemble integration. The mean absolute error between the probabilistic forecast results and the prediction results is calculated as the degree of discrepancy to evaluate the physical rationality of the prediction results. The validation step combines data-driven prediction with the physical model to improve the credibility of the forecast results.

[0056] In some embodiments, updating the model parameters of the prediction model based on the verification results and real-time meteorological observation data includes: Calculate the importance weights of the model parameters of the prediction model to historical tasks; A regularization term is constructed based on the importance weights, and the regularization term is added to the loss function to constrain the change in the model parameters during the update process; Based on the verification results, the sample selection criteria for incremental learning are determined, and training samples are selected from the real-time meteorological observation data collected in real time in combination with the importance ranking. The prediction model is incrementally trained using the training samples to update the model parameters.

[0057] Real-time monitoring of the prediction error of the prediction model on the validation set; A dynamic threshold is set, and when the prediction error is detected to continuously exceed the dynamic threshold, the model retraining process is initiated. In the retraining process, samples corresponding to high-importance features are selected according to the importance ranking, and a training set is constructed by combining the collected real-time weather data. The prediction model is retrained using the training set, and the model update cycle is adjusted according to the error trend in the validation phase.

[0058] Specifically, in the incremental learning dynamic update process, this application uses an elastic weight consolidation algorithm to calculate the importance weights of model parameters to historical tasks, based on the Fisher information matrix of the parameters in the loss function. A regularization term is constructed based on the importance weights and added to the loss function to constrain the variation of important parameters during incremental training. The sample selection criteria for incremental learning combine validation results and feature importance ranking, selecting high-value training samples from real-time meteorological observation data. The model retraining process is triggered when the validation set prediction error continuously exceeds a dynamic threshold, prioritizing training with samples corresponding to high-importance features and adjusting the model update cycle to maintain prediction performance. The incremental learning mechanism enables the model to adapt to the dynamic changes in meteorological data and maintain long-term effectiveness.

[0059] The extreme weather prediction method provided in this application first addresses the challenge of data heterogeneity by employing spatiotemporal alignment and quality correction steps in the data preprocessing stage. Spatiotemporal alignment assigns confidence weights to observation data from different sources and uses a weighted dynamic time warping algorithm to align multi-source time series based on ground station data. Then, a standardized data cube is generated through interpolation, unifying the inconsistencies in time and spatial scales across the multi-source data. Quality correction utilizes a conditional generative adversarial network (GAN). The generator corrects data anomalies based on the distribution of historical high-quality datasets, while the discriminator generates a data quality evaluation matrix for iterative optimization and updates quality identifiers to record anomaly locations and correction magnitudes. This improves data quality from the source and reduces noise introduced by heterogeneous data.

[0060] To address the complex coupling relationships among meteorological factors, this application introduces a latent space projection technique. Spatial features are extracted using a graph attention network, where the calculation of attention weights between nodes comprehensively considers both terrain elevation differences and inter-node distances, capturing the interaction between terrain and atmospheric circulation. Simultaneously, a temporal convolutional network is used to extract multi-scale temporal features. After fusing spatial and temporal features, contrastive learning is employed to constrain the latent space vector distribution, causing samples with similar evolutionary trends to cluster in the latent space. This projection method transforms high-dimensional nonlinear coupling relationships into low-dimensional, analytically reproducible feature representations, providing a more effective input for the model.

[0061] At the model prediction level, this application employs a meta-learning framework to enhance the modeling capability for complex coupling relationships through a multi-task learning mechanism. The support set includes typical extreme weather scenarios, the query set is used to evaluate generalization performance, and a sample distance metric calculated based on latent space feature vectors guides rapid adaptation of model parameters, enabling the model to learn from a small number of samples and adapt to rare events, thus improving the accuracy of capturing complex weather patterns. Multi-objective collaborative optimization and adversarial training further enhance model robustness. A non-dominated sorting genetic algorithm balances multiple objectives such as prediction accuracy, model calibration, and computational efficiency. Adversarial example training enhances the model's adaptability to data perturbations by perturbing input features, avoiding overfitting.

[0062] This application also combines interpretability analysis methods with digital twin simulation environment verification to ensure the transparency and physical rationality of model decisions. This application quantifies the marginal contribution of each latent space eigenvector to the prediction results through Shapleyga and interpretation methods, thereby generating a feature importance ranking and revealing key driving factors. This application also utilizes a digital twin environment based on a WRF numerical model, using the prediction results as initial field perturbation terms for ensemble forecasting, and verifies the physical consistency of the prediction results through dissimilarity assessment, thus bridging data-driven approaches and physical models and improving forecast reliability.

[0063] Finally, this application achieves dynamic model evolution through an incremental learning mechanism. The elastic weight consolidation algorithm calculates parameter importance weights, and regularization terms constrain changes in important parameters. New data samples are selected for incremental training based on validation results and feature importance ranking. The model retraining process is triggered by a dynamic threshold for prediction error, ensuring the model continuously adapts to changes in meteorological data. This technical solution, through a closed-loop process of data preprocessing, feature extraction, model optimization, and validation updates, collaboratively addresses the challenges posed by data heterogeneity and complex coupling relationships, significantly improving the accuracy of extreme weather evolution trend prediction.

[0064] The method provided in this application can be applied to typhoon forecasting scenarios in power meteorological forecasting systems. During the data acquisition phase, multimodal meteorological observation data, including temperature, air pressure, humidity, wind speed, and precipitation, are collected from meteorological observation stations, satellite remote sensing platforms, and radiosonde equipment. Ground station data, being directly measured, receives a higher confidence weight. In the spatiotemporal alignment process, a weighted dynamic time warping algorithm is used to align the time series of satellite and radiosonde data with the ground station data time as a reference. The aligned data is then used to generate a standardized data cube with a 0.25-degree grid through bilinear interpolation. Quality correction is performed using a conditional generative adversarial network. The generator receives the data cube and an initial quality identifier, outputs the corrected data, and the discriminator compares it with historical high-quality datasets to generate a data quality evaluation matrix. Iterative correction of data anomalies is performed, and the quality identifier is updated. In the latent space projection phase, the graph attention network incorporates terrain elevation and distance factors when calculating the attention weights between nodes. A temporal convolutional network extracts multi-scale temporal features, which are then fused and compared to obtain a latent space feature vector. The meta-learning framework samples support and query sets of typhoon scenarios from the training set, calculates sample distance metrics based on latent space feature vectors, and quickly adapts model parameters to generate predictions of typhoon occurrence probability, impact range, and duration. Multi-objective optimization employs a non-dominated sorting genetic algorithm to balance prediction error, calibration, and efficiency, and introduces adversarial example training to enhance robustness. This application can calculate feature marginal contributions using Shapley and interpretation methods to generate feature importance rankings. The digital twin environment in this application uses prediction results as initial field perturbation terms for ensemble forecasting based on the WRF model to verify physical rationality. This application also supports incremental learning, i.e., selecting new data based on verification results and feature importance rankings, updating model parameters through an elastic weight consolidation algorithm, and triggering a retraining process when the prediction error exceeds a dynamic threshold, thereby optimizing typhoon prediction accuracy.

[0065] As a specific embodiment, this application addresses the prediction of the impact of cold waves on wind power output. During the data acquisition phase, data from ground stations, satellites, and ocean buoys are integrated. Ground station data is given higher weight during spatiotemporal alignment, and a weighted dynamic time warping algorithm is used to align heterogeneous data sources, generating a standardized data cube. A quality correction network corrects data anomalies through a generator, and a discriminator evaluates data quality based on historical cold wave datasets, updating quality identifiers to record correction information. In latent space projection, a graph attention network captures the interaction between topography and atmospheric circulation, a temporal convolutional network extracts the temporal patterns of cold wave evolution, and contrastive learning constrains the latent space distribution to cluster similar cold wave samples. The meta-learning framework supports sets including cold wave development scenarios, rapidly adapts the model based on latent space distance metrics, and generates predictions of cold wave duration, intensity, and impact range. During multi-objective optimization, a non-dominated sorting genetic algorithm solves for Pareto optimal solutions, and adversarial training improves the model's generalization ability by perturbing key features such as wind speed and air pressure. Interpretability analysis quantifies the contribution of each latent space feature vector to wind power output prediction. A digital twin environment simulates the cold wave propagation path to verify the consistency between the prediction results and the physical model. An incremental learning mechanism updates the model based on real-time meteorological data, an elastic weight consolidation algorithm retains historical cold wave pattern knowledge, and the retraining process prioritizes high-importance feature samples to dynamically adjust the model for accurate prediction of the impact of cold waves on wind power output.

[0066] like Figure 2 As shown, this application provides an extreme weather prediction system, comprising: The acquisition module 201 is used to acquire multimodal meteorological observation data, including temperature, air pressure, humidity, wind speed, and precipitation. Processing module 202 is used to perform spatiotemporal alignment and quality correction on the meteorological observation data to obtain an assimilated dataset; wherein, a quality identifier is generated through quality correction, and the quality identifier is used to record the location of data anomalies and the correction magnitude; Extraction module 203 is used to adjust the feature extraction weights based on the quality identifier, project the assimilation dataset into a low-dimensional latent space, and obtain a latent space feature vector; The prediction module 204 is used to perform multi-task prediction on the latent space feature vector based on the sample distance metric of the latent space, and generate prediction results; the prediction results include the probability of extreme weather occurrence, the scope of impact, and the duration. The optimization module 205 is used to perform multi-objective collaborative optimization and adversarial training on the neural network using the prediction results to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration and computational efficiency. The verification module 206 is used to quantify the contribution of the latent space feature vectors in the prediction model and verify the prediction results using a digital twin simulation environment to obtain the verification results. The update module 207 is used to update the model parameters of the prediction model based on the verification results and real-time meteorological observation data.

[0067] According to embodiments of this application, this application also provides an electronic device and a readable storage medium.

[0068] The electronic device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the at least one processor, which, when executed, enable the at least one processor to perform the extreme weather prediction method described in this application. The computer instructions are used to cause the computer to perform the extreme weather prediction method described in this application.

[0069] This application also provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the extreme weather prediction method of this application.

[0070] Figure 3 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of this application is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the application described and / or claimed herein.

[0071] like Figure 3 As shown, device 800 includes a computing unit 801, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 802 or a computer program loaded from storage unit 808 into random access memory (RAM) 803. RAM 803 may also store various programs and data required for the operation of device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via bus 804. Input / output (I / O) interface 805 is also connected to bus 804.

[0072] Multiple components in device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of monitors, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0073] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as extreme weather prediction methods. For example, in some embodiments, the extreme weather prediction method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program may be loaded and / or installed on device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the extreme weather prediction method described above may be performed. Alternatively, in other embodiments, the computing unit 801 may be configured to perform extreme weather prediction methods by any other suitable means (e.g., by means of firmware).

[0074] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0075] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0076] In the context of this application, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0077] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0078] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0079] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. A client-server mapping is created by computer programs running on the respective computers and having client-server relationships with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0080] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method of predicting extreme weather, characterized by, The method comprises the following steps: Collecting multi-modal meteorological observation data; the meteorological observation data includes temperature, air pressure, humidity, wind speed and precipitation; Performing space-time alignment and quality correction on the meteorological observation data to obtain an assimilated data set; wherein a quality identifier is generated through quality correction, and the quality identifier is used to record the position and correction amplitude of abnormal data; Adjusting the feature extraction weight based on the quality identifier, projecting the assimilated data set to a low-dimensional latent space to obtain a latent space feature vector; the latent space feature vector includes meteorological features; Based on the sample distance measurement of the latent space, a meta-learning framework is used to perform multi-task prediction on the latent space feature vector to generate a prediction result; the prediction result includes the probability, impact range and duration of extreme weather occurrence; Using the prediction result to perform multi-objective collaborative optimization and adversarial training on the neural network to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration and computational efficiency; Quantifying the contribution of the latent space feature vector in the prediction model and verifying the prediction result using digital twin simulation environment to obtain a verification result; Based on the verification result and real-time meteorological observation data, updating the model parameters of the prediction model.

2. The method of claim 1, wherein, The multi-modal meteorological observation data includes ground station data and satellite remote sensing data; the space-time alignment of the meteorological observation data comprises: According to the pre-set configuration rule, configure the confidence weight of each modal meteorological observation data; wherein the confidence weight of the ground station data is higher than that of the satellite remote sensing data; Based on the confidence weight, use the weighted dynamic time warping algorithm to align the time series of each modal meteorological observation data based on the ground station data time to obtain time-aligned data; Interpolate the time-aligned data to the latitude and longitude grid so that the time-aligned data has spatial coordinate identification to obtain standardized data; the standardized data is used as input data for quality correction.

3. The method of claim 2, wherein, The quality correction of the meteorological observation data comprises: The generator receives the standardized data and the initial quality identifier and outputs corrected data; The discriminator compares the distribution difference between the corrected data and the standard data set to generate a data quality evaluation matrix, and feeds the data quality evaluation matrix back to the generator The generator iteratively corrects abnormal data in the corrected data using the data quality evaluation matrix, and updates the quality identifier after correcting the abnormal data.

4. The method of claim 1, wherein, The projection of the assimilated data set to a low-dimensional latent space to obtain a latent space feature vector comprises: Using a graph attention network to extract the spatial features of the assimilated data set and using a time series convolution network to extract the multi-scale time features of the assimilated data set; Fuse the spatial features and the multi-scale time features to obtain fused features; Determine the distribution of the latent space vector in the fused features using contrastive learning constraints to obtain the latent space feature vector.

5. The method of claim 1, wherein, The sample distance measurement based on the latent space, using a meta-learning framework to perform multi-task prediction on the latent space feature vector to generate a prediction result, comprises: constructing a training set based on the latent space feature vectors; the training set includes a support set and a query set, and the support set includes typical extreme weather scenarios; calculating a sample distance metric based on the latent space feature vectors; determining model parameters by performing a few-shot gradient descent on the support set using a meta-learning framework based on the sample distance metric; evaluating model performance on the query set based on the model parameters, optimizing the model parameters, and obtaining a prediction result with the probability, impact range, and duration of extreme weather occurrence.

6. The method of claim 1, wherein, The multi-objective collaborative optimization and adversarial training of the neural network using the prediction result to obtain a prediction model, comprising: establishing a multi-objective optimization problem; the multi-objective optimization problem includes a prediction error reciprocal objective function, an expected calibration error objective function, and an inference speed reciprocal objective function; solving the multi-objective optimization problem using a non-dominated sorting genetic algorithm, and introducing an adversarial sample training to select the optimal solution in the solving process to obtain a non-dominated sorting result; selecting the optimal model parameters according to the non-dominated sorting result to obtain the prediction model.

7. The method of claim 1, wherein, The contribution degree of the latent space feature vector in the prediction model is quantified, comprising: extracting the mapping relationship between the input latent space feature vector and the prediction result from the prediction model; calculating the marginal contribution value of each input latent space feature vector to the prediction result using the Shapley sum explanation method; generating a feature importance ranking according to the marginal contribution value; determining the contribution degree of the latent space feature vector based on the feature importance ranking.

8. The method of claim 1, wherein, The prediction result is verified using a digital twin simulation environment to obtain a verification result, comprising: constructing a digital twin environment based on a weather forecast numerical model, and setting physical parameters and boundary conditions; inputting the prediction result output by the prediction model as an initial field disturbance term of the digital twin environment; running the digital twin environment to obtain a probabilistic prediction result; calculating the difference between the probabilistic prediction result and the prediction result; evaluating the physical rationality according to the difference to obtain the verification result.

9. The method of claim 7, wherein, The model parameters of the prediction model are updated based on the verification result and real-time meteorological observation data, comprising: calculating the importance weight of the model parameters of the prediction model for historical tasks; adding a regularization term to the loss function to constrain the change amplitude of the model parameters in the updating process according to the importance weight; determining a sample selection standard for incremental learning according to the verification result, and screening training samples from real-time meteorological observation data collected in real time according to the importance ranking; updating the model parameters by incrementally training the prediction model using the training samples; monitoring the prediction error of the prediction model on the verification set in real time; setting a dynamic threshold, and starting a model retraining process when the prediction error is continuously monitored to exceed the dynamic threshold; In the retraining process, samples corresponding to high importance features are selected according to the importance ranking, and a training set is constructed combining the real-time weather data collected; retrain the prediction model using the training set, and adjust the model update period according to the error trend of the verification link.

10. A system for predicting extreme weather, characterized in that, ​ The collection module is used for collecting multi-modal meteorological observation data; the meteorological observation data includes temperature, air pressure, humidity, wind speed and precipitation; The processing module is used for spatio-temporal alignment and quality correction of the meteorological observation data to obtain an assimilation data set; wherein a quality identifier is generated through quality correction, and the quality identifier is used for recording an abnormal data position and a correction amplitude; The extraction module is used for adjusting feature extraction weights based on the quality identifier, projecting the assimilation data set to a low-dimensional hidden space, and obtaining a hidden space feature vector; The prediction module is used for multi-task prediction of the hidden space feature vector based on sample distance measurement of the hidden space to generate a prediction result; the prediction result includes probability, influence range and duration of extreme weather occurrence; The optimization module is used for multi-objective collaborative optimization and adversarial training of a neural network based on the prediction result to obtain a prediction model; the multi-objective collaborative optimization includes balancing prediction accuracy, model calibration degree and calculation efficiency; The verification module is used for quantifying contribution degrees of the hidden space feature vector in the prediction model, and verifying the prediction result by using a digital twin simulation environment to obtain a verification result; The update module is used for updating model parameters of the prediction model based on the verification result and real-time meteorological observation data.

Images