An extreme high-temperature interdecadal prediction method and device and a storage medium

Abstract

The application discloses an extreme high-temperature interdecadal prediction method and device and a storage medium, and belongs to the technical field of interdecadal prediction, which comprises the following steps: obtaining a second extreme high-temperature intensity index extracted based on multi-mode multi-initial field return test data in a target prediction period, inputting the second extreme high-temperature intensity index into a pre-trained regional prediction model based on a Transformer, and obtaining an extreme high-temperature intensity interdecadal prediction sequence of each region in the target geographical region; and reconstructing the extreme high-temperature intensity interdecadal prediction sequence of each region on a spatial grid point by a linear projection method to obtain a spatial distribution prediction result of the extreme high-temperature intensity in the target geographical region. The application significantly improves the interdecadal prediction capability of a dynamic mode on the extreme high-temperature intensity in summer, and effectively makes up for the deficiency of an existing mode in the interdecadal prediction of the extreme high-temperature intensity.

Description

Technical Field

[0001] This invention relates to a method, apparatus, and storage medium for predicting interdecadal extreme high temperatures, belonging to the field of interdecadal prediction technology. Background Technology

[0002] Regions generally exhibit weak adaptability in responding to extreme weather events. In particular, the intensity of extreme summer heat events in mid-to-high latitude Eurasia shows significant interdecadal variability, making effective interdecadal forecasting crucial. Accurate interdecadal forecasts not only provide a scientific basis for formulating adaptive policies, adjusting resource allocation, and optimizing disaster prevention and mitigation strategies, but also offer early warnings for potential future extreme weather events. Currently, interdecadal forecasting primarily relies on initialization-based dynamical models. However, issues such as initialization shocks, initial value errors, boundary value errors, and systematic model errors limit the interdecadal forecasting capabilities of current dynamical models for extreme heat events, falling far short of practical needs. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device and storage medium for interdecadal prediction of extreme high temperature, which significantly improves the ability of dynamic models to predict the interdecadal intensity of summer extreme high temperature and effectively makes up for the shortcomings of existing models in interdecadal prediction of extreme high temperature intensity.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] In a first aspect, the present invention provides a method for predicting interdecadal extreme high temperatures, comprising:

[0006] The second extreme high temperature intensity index, extracted from multi-mode multi-initial field return test data within the target prediction period, is input into a pre-trained Transformer-based regional prediction model to obtain the interdecadal prediction sequence of extreme high temperature intensity for each region within the target geographical area.

[0007] The interdecadal prediction sequence of extreme high temperature intensity in each region is reconstructed onto spatial grid points using a linear projection method to obtain the spatial distribution prediction results of extreme high temperature intensity within the target geographic area.

[0008] The training method for the Transformer-based region prediction model includes:

[0009] Acquire the highest temperature observation data and corresponding multi-mode, multi-initial-field return test data during the historical training period;

[0010] The first extreme high temperature intensity index is extracted based on the highest temperature observation data, and the moving average and regional clustering are performed for a set year to obtain the first regional average index for each region.

[0011] The second extreme high temperature intensity index was extracted based on multi-mode multi-initial-field return test data, and the moving average and regional average of the set years were performed to obtain the second regional average index of each initial field.

[0012] We select the average index of the second region with the highest prediction skill ranking in each initial field as input features, and take the average index of the first region of each region as the prediction target. We then construct and train a Transformer-based deep learning model for each region.

[0013] Furthermore, the first extreme high temperature intensity index is extracted based on the highest temperature observation data, and a moving average and regional clustering are performed for a set number of years to obtain the first regional average index for each region, including:

[0014] The first extreme high temperature intensity index was calculated using the relative threshold method based on the highest temperature observation data.

[0015] The first decadal variability is obtained by taking a moving average of the first extreme high temperature intensity index over a set number of years.

[0016] The first-generation interdecadal variability was divided into regions using a spectral clustering algorithm to obtain the divided regions;

[0017] The first interdecadal variability of each region is averaged regionally to obtain the first regional average index for each region.

[0018] Furthermore, the second extreme high temperature intensity index is extracted based on multi-mode, multi-initial-field return test data, and a moving average and regional average are performed over a set number of years to obtain the second regional average index for each initial field, including:

[0019] Acquire multi-mode, multi-initial-field return test data and calculate the second extreme high-temperature intensity index using the relative threshold method;

[0020] The second extreme high temperature intensity index is averaged over a set number of years to obtain the second decadal variability.

[0021] Based on the above-described regional divisions, the second interdecadal variability is averaged regionally to obtain the second regional average index for each initial field.

[0022] Furthermore, the multi-mode multi-initial-field return test data is derived from the return test data of multiple modes in the sixth Coupled Mode Comparison Program decadal forecast program experiment.

[0023] Furthermore, the method for selecting the average index of multiple second-region regions with the highest rankings in each initial in-field prediction skill includes:

[0024] Using the evaluation method of anomaly correlation coefficient and root mean square error skill scoring, the spatial scale prediction skill of the second decadal variability and the first decadal variability of each initial field is evaluated, and the temporal scale prediction skill of the second decadal variability after regional averaging and the regional average observation index is evaluated, and the evaluation results are obtained.

[0025] Based on the evaluation results, the average index of the top-ranked second-region prediction skills in each initial field was extracted.

[0026] Furthermore, the structure of the Transformer-based deep learning model includes:

[0027] A linear projection layer is used to convert the average exponent of multiple second regions of the input into vector features;

[0028] A multi-layer Transformer encoder is used to extract the temporal and cross-regional dependencies of vector features using a global attention mechanism, resulting in extracted vector features.

[0029] A multilayer perceptron is used to perform regression prediction on the extracted vector features and output a predicted sequence.

[0030] Furthermore, the process of reconstructing the interdecadal prediction sequences of extreme high-temperature intensity in each region onto a spatial grid using a linear projection method to obtain the spatial distribution prediction results of extreme high-temperature intensity within the target geographical region includes:

[0031] During the training period, the first detrended regional average exponential regression of the first detrended interdecadal variability of each grid point in the region is used to calculate the variability regression coefficients and intercepts.

[0032] Linear regression was performed on the first interdecadal variability of each grid point to obtain the regression coefficients and intercepts of the linear trend;

[0033] Based on the regression coefficients and intercepts of the variable rate and the regression coefficients and intercepts of the linear trend, the spatial distribution prediction results of each region are calculated using the linear projection formula.

[0034] By merging the spatial distribution prediction results of each region, we obtain the spatial distribution prediction results of extreme high temperature intensity within the target geographical region.

[0035] Furthermore, the calculation formula for the linear projection formula is as follows:

[0036] ;

[0037] in, For the firsti The grid point at the th grid point t The projected value for the year, For the predicted sequence of the th t The value of the year, and The first i The rate of change regression coefficients and intercepts for each grid point; and The first i The linear trend regression coefficients and intercepts for each grid point.

[0038] Secondly, the present invention provides an extreme high-temperature interdecadal prediction device, comprising:

[0039] Memory, used to store computer programs / instructions;

[0040] A processor for executing the computer program / instructions to implement the steps of any of the methods described above.

[0041] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0042] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0043] This invention provides a method, apparatus, and storage medium for interdecadal prediction of extreme high temperatures. By introducing a Transformer-based regional prediction model, it significantly improves the ability of dynamical models to predict the interdecadal intensity of summer extreme high temperatures. Compared with traditional dynamical models, this Transformer-based regional prediction model effectively compensates for the shortcomings of existing models in interdecadal prediction of extreme high temperature intensity. This improvement provides a more reliable predictive basis for regional climate change responses, thereby providing a scientific basis for formulating precise adaptation policies, rationally optimizing resource allocation, and developing more efficient disaster prevention and mitigation strategies. Attached Figure Description

[0044] Figure 1 This is a flowchart of the extreme high temperature interdecadal prediction method provided in the embodiments of the present invention;

[0045] Figure 2 This is a flowchart illustrating the application process of the extreme high temperature interdecadal prediction method provided in this embodiment of the invention. Detailed Implementation

[0046] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0047] Example 1, as Figure 1 As shown in the figure, this embodiment introduces a method for predicting interdecadal extreme high temperatures, including:

[0048] The second extreme high temperature intensity index, extracted from multi-mode multi-initial field return test data within the target prediction period, is input into a pre-trained Transformer-based regional prediction model to obtain the interdecadal prediction sequence of extreme high temperature intensity for each region within the target geographical area.

[0049] The interdecadal prediction sequence of extreme high temperature intensity in each region is reconstructed onto spatial grid points using a linear projection method to obtain the spatial distribution prediction results of extreme high temperature intensity within the target geographic area.

[0050] The training method for the Transformer-based region prediction model includes:

[0051] Acquire the highest temperature observation data and corresponding multi-mode, multi-initial-field return test data during the historical training period;

[0052] The first extreme high temperature intensity index is extracted based on the highest temperature observation data, and the moving average and regional clustering are performed for a set number of years to obtain the first regional average index for each region; where the set number of years can be five years.

[0053] The second extreme high temperature intensity index was extracted based on multi-mode multi-initial-field return test data, and the moving average and regional average of the set years were performed to obtain the second regional average index of each initial field.

[0054] We select the average index of the second region with the highest prediction skill ranking in each initial field as input features, and take the average index of the first region of each region as the prediction target. We then construct and train a Transformer-based deep learning model for each region.

[0055] like Figure 2 As shown in the figure, the application process of the extreme high temperature interdecadal prediction method provided in this embodiment involves the following steps:

[0056] Step 1: Identify the intensity of extreme heat events and clarify the regional characteristics of their interdecadal variability. Daily maximum temperature data from ERA5 (European Centre for Medium-Range Weather Forecasts Reanalysis v5, the fifth-generation global atmospheric reanalysis dataset launched by the Copernicus Climate Change Service of the European Union) from 1966 to 2022 were selected. The summer extreme heat intensity index was calculated at each grid point using the 90% relative threshold method, yielding the first extreme heat intensity index. To extract the interdecadal variability signal, a five-year moving average was applied to the first extreme heat intensity index, obtaining the first interdecadal variability from 1968 to 2020. Spatial clustering analysis of the first interdecadal variability was performed using spectral clustering, dividing the mid-to-high latitudes of Eurasia into 14 regions with significant interdecadal covariance characteristics. Finally, the interdecadal variability of the first extreme heat intensity index at all grid points within each region was averaged regionally, forming the first regional average index for the 14 regions from 1968 to 2020, serving as the observational benchmark for subsequent models.

[0057] Step 2: Evaluate the predictive ability of dynamic climate models for interdecadal variability of extreme heat intensity. Daily maximum temperature data from 86 initial fields across 6 models in the 6th Coupled Model Intercomparison Project (CPIP) interdecadal prediction experiment were selected. Using the 90% relative threshold method consistent with observations, the second extreme heat intensity index for each initial field was calculated. Subsequently, a five-year moving average was applied, and the 6–10 year forecast lead time was extracted to obtain the second interdecadal variability from 1968 to 2026. Combined with the 14 clusters obtained in Step 1, the second interdecadal variability was regionally averaged to form the second regional average index for each initial field. Using 1968–2020 as the comparison period, the anomaly correlation coefficient and the root mean square error technique were used for evaluation.

[0058] Correlation and Techniques Between Second and First Decadal Variability (Spatial Scale Assessment)

[0059] Correlation and techniques between the average index of the second region and the average index of the first region (time scale assessment).

[0060] Finally, the average index of the second region of the top 30 initial fields with the best prediction skills was selected as the input feature for the subsequent deep learning model.

[0061] Step 3: Construct a Transformer-based deep learning model. Building upon the dynamic mode output, deep learning methods are introduced to further improve prediction accuracy. The model adopts a Transformer-based architecture with the following structure: the input layer is a linear projection layer, converting the average exponents of the 30 selected initial fields in the second region into vector representations that can be processed by the Transformer; the hidden layer is a multi-stacked Transformer encoder, utilizing a global attention mechanism to extract temporal and cross-regional dependencies; the output layer is a multilayer perceptron for regression prediction. Models are constructed separately for each of the 14 clustered regions. Each region's model uses the average exponents of the 30 initial fields in the second region as input and the observed average exponents of the first region as the prediction target, and is trained independently. After training, the model can output a predicted sequence with a time scale consistent with the input feature, the average exponents of the second region.

[0062] Step 4: Reconstructing interdecadal predictions at spatial scales. After model training, the second regional average index for each region from 2021 to 2026 is input into the corresponding regional model to predict the period from 2021 to 2026, obtaining the predicted sequences for each region. A spatial reconstruction method based on linear projection is designed to map the regional predicted sequences to grid points within the region, thereby restoring the spatial distribution field. Through the above methods, the extension from regional temporal scale predictions to spatial scale predictions is realized, and the applicability of the prediction results in climate risk assessment and services is enhanced.

[0063] The following description, in conjunction with a preferred embodiment, illustrates the content involved in the above embodiments.

[0064] Taking the interdecadal variation prediction of the spatial distribution of extreme high temperature intensity in mid-to-high latitude summers in Eurasia from 2021 to 2026 as a case study, the implementation process is explained in detail.

[0065] Step 1: Identify the intensity of extreme summer heat events in mid-to-high latitude Eurasia and clarify the regional characteristics of their interdecadal variability. Using daily maximum temperatures from ERA5 data (1966-2022) as observational data, calculate the first extreme heat intensity index for summer (June-August) at each grid point in the mid-to-high latitude Eurasia region (60°–130°E, 20°–75°N). The extreme heat intensity index is defined as the 90th percentile of the daily maximum temperature in summer during the baseline period (1981-2010) as the temperature threshold. Each summer, if the daily maximum temperature exceeds this threshold, it is considered an extreme heat day. Calculate the difference between the daily maximum temperature of these extreme heat days and the threshold, and take their average. Perform a five-year moving average of the summer extreme heat index for each grid point to obtain the first interdecadal variability. The five-year moving average result is labeled with the year in the middle of that five-year period. For example, the average value from 1966 to 1970 is labeled as the interdecadal variability in 1968, and so on, ultimately yielding the interdecadal variability from 1968 to 2020. A spectral clustering algorithm is used to perform cluster analysis on the intensity of extreme summer heat in mid-to-high latitude Eurasia. Spectral clustering algorithms typically use Euclidean distance to construct a similarity matrix between samples. However, Euclidean distance may not fully reflect the similarity between samples in some cases, especially in the context of climate data with complex spatiotemporal characteristics. Therefore, the anomaly correlation coefficient (ACC) is introduced as a similarity metric to construct the similarity matrix. Quantifying the similarity between different regions using ACC allows for more accurate identification of regions with similar climate patterns. Furthermore, the number of clusters, k, is set to 20, and the regional average index anomaly correlation coefficient (ACC) and mean square skill score (MSSS) are calculated for each region. Based on the correlation between ACC and MSSS, highly correlated regions are merged to obtain more accurate clustering results. This method clusters the extreme summer heat intensity in mid-to-high latitudes of Eurasia into 14 different regions. Based on the 14 clustered regions, the interdecadal variability of the first extreme heat intensity index is averaged regionally to obtain the first regional average index.

[0066] Step 2: Evaluate the interdecadal forecasting skills of dynamic models for the intensity of extreme summer temperatures in mid-to-high latitudes of Eurasia. Processing the dynamic model dataset: (1) Use the daily maximum temperature data of 86 initial fields from 6 models in the sixth Coupled Model Comparison Project Interdecadal Prediction Experiment Group A. The reporting years for each initial field are from 1961 to 2018, and a 10-year forecast is performed for each reporting year. For example, if the reporting starts in 1961, the forecast is for 1961-1970, and so on; (2) Calculate the intensity index of the second extreme summer temperatures in mid-to-high latitudes of Eurasia for each initial field using the relative threshold method; (3) Perform a five-year moving average for the forecast lead time for each reporting year, and extract the last valid year for each data set. Specifically, after performing a five-year moving average for the forecast lead time (1961-1970) starting in 1961, 1968 is taken as the valid year, and so on, until the forecast data is obtained starting in 2018. Finally, combining the results of all reporting years, the second decadal variability from 1968 to 2026 was obtained. (4) Based on the 14 clusters of extreme high temperature intensity in mid-to-high latitude Eurasia during summer, the decadal variability of the second extreme high temperature intensity index of each initial field was averaged regionally to obtain the second regional average index; (5) The ACC and MSSS between the decadal variability of the second extreme high temperature intensity index from 1968 to 2020 and the decadal variability of the first extreme high temperature intensity index from 1968 to 2020 were calculated to evaluate the model's predictive ability in the spatial field. Next, the ACC and MSSS between the second regional average index from 1968 to 2020 and the first regional average index from 1968 to 2020 were calculated to evaluate the model's interdecadal prediction skill on the time scale. Finally, the second regional average index of the 30 initial fields with the best prediction skill for each region was selected as the input features for the subsequent deep learning model of that region.

[0067] Step 3: Construct a Transformer-based deep learning model to improve the interdecadal prediction skills of patterns for extreme summer heat intensity in mid-to-high latitude Eurasia. This model consists of an input layer, hidden layers, and an output layer. In the input layer, a linear projection layer transforms the input features into a format that the Transformer encoder can process. The shape of the input data is set to (batch_size, 1, 30), where batch_size is a hyperparameter representing the number of samples in one training iteration; 1 indicates that only one time step is predicted at a time, i.e., the value for the current year; and 30 represents the number of input features, containing 30 selected features. The hidden layer uses a multi-stacked Transformer encoder for feature extraction. The feature dimension input to the encoder is (batch_size, 1, d_model), where d_model represents the hidden state dimension of the Transformer encoder, a hyperparameter that determines the dimension of the features within the encoder. In this layer, the encoder extracts complex patterns of input features through a multi-head self-attention mechanism and a feedforward neural network. Ultimately, the encoder's output dimension is (batch_size, 1, d_model), the same as the input dimension, representing the feature table at each time step. The output layer consists of a multilayer perceptron, used to map the features extracted by the encoder to the final regression output. The input dimension to the multilayer perceptron layer is (batch_size, 1, d_model). In MLP, feature transformation and numerical prediction are performed through a series of fully connected layers, activation functions, layer normalization, and Dropout, outputting the regression value y. The input features are then added to the regression value y through residual connections to preserve the original input information and improve gradient flow during training, ultimately generating the model's predicted value Y. For each cluster region, the average index of the second region of the 30 best initial fields is selected as the input feature for separate training. The training period was from 1968 to 2005, and the testing period was from 2006 to 2020.

[0068] Step 4: After model training is complete, the second regional average index for each region from 2021 to 2026 is input into the corresponding regional model to predict the period from 2021 to 2026, obtaining the predicted sequence for each region. The model's predicted sequence is then linearly projected into the spatial field to obtain the interdecadal prediction of the spatial field. The specific steps are as follows:

[0069] (1) Calculate the regression coefficient and intercept of the variability for each grid point during the training period: During the training period, the first detrended interdecadal variability of each grid point in the region is regressed using the first regional average exponent after detrended, and the regression coefficient and intercept of the variability are calculated.

[0070] (2) Calculate the linear trend regression coefficient and intercept for each grid point during the training period: Perform linear regression on the first interdecadal variability of each grid point to obtain the linear trend regression coefficient and intercept;

[0071] (3) Projection of the spatial field: The final predicted value of the spatial field is generated by the following linear projection formula:

[0072] ;

[0073] in, For the first i The grid point at the th grid point t The projected value for the year, For the predicted sequence of the th t The value of the year, and The first i The rate of change regression coefficients and intercepts for each grid point; and The first i Linear trend regression coefficients and intercepts for each grid point;

[0074] (4) Merging prediction results: The spatial field prediction results of 14 regions are merged to finally obtain the prediction of the extreme high temperature intensity spatial field of the entire Eurasian mid-high latitude region from 2021 to 2026.

[0075] This embodiment significantly improves the ability of dynamical models to predict the interdecadal intensity of extreme summer temperatures by introducing a Transformer-based regional prediction model. Compared with traditional dynamical models, this Transformer-based regional prediction model can effectively compensate for the shortcomings of existing models in predicting the interdecadal intensity of extreme temperatures. This improvement provides a more reliable predictive basis for regional climate change responses, thereby providing a scientific basis for formulating precise adaptation policies, rationally optimizing resource allocation, and developing more efficient disaster prevention and mitigation strategies.

[0076] Example 2: This example provides an extreme high-temperature interdecadal prediction device, comprising:

[0077] Memory, used to store computer programs / instructions;

[0078] A processor for executing the computer program / instructions to implement the steps of any of the methods described in Embodiment 1.

[0079] Example 3: This example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described in Example 1.

[0080] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0081] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0083] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit its protection scope. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading this disclosure, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the protection scope of the pending claims.

Claims

1. A method for predicting extreme high temperatures across decades, characterized in that, include: The second extreme high temperature intensity index, extracted from multi-mode multi-initial field return test data within the target prediction period, is input into a pre-trained Transformer-based regional prediction model to obtain the interdecadal prediction sequence of extreme high temperature intensity in each region within the target geographical area. The interdecadal prediction sequence of extreme high temperature intensity in each region is reconstructed onto spatial grid points using a linear projection method to obtain the spatial distribution prediction results of extreme high temperature intensity within the target geographic area. The training method for the Transformer-based region prediction model includes: Acquire the highest temperature observation data and corresponding multi-mode, multi-initial-field return test data during the historical training period; The first extreme high temperature intensity index is extracted based on the highest temperature observation data, and the moving average and regional clustering are performed for a set year to obtain the first regional average index for each region. The second extreme high temperature intensity index was extracted based on multi-mode multi-initial-field return test data, and the moving average and regional average of the set years were performed to obtain the second regional average index of each initial field. We select the average index of the second region with the highest prediction skill ranking in each initial field as input features, and take the average index of the first region of each region as the prediction target. We then construct and train a Transformer-based deep learning model for each region.

2. The method for predicting extreme high temperatures across decades according to claim 1, characterized in that, The method involves extracting the first extreme high temperature intensity index based on the highest temperature observation data, performing a moving average for a set number of years, and dividing the data into regional clusters to obtain the first regional average index for each region, including: The first extreme high temperature intensity index was calculated using the relative threshold method based on the highest temperature observation data. The first decadal variability is obtained by taking a moving average of the first extreme high temperature intensity index over a set number of years. The first-generation interdecadal variability was divided into regions using a spectral clustering algorithm to obtain the divided regions; The first interdecadal variability of each region is averaged regionally to obtain the first regional average index for each region.

3. The method for predicting extreme high temperatures in different decades according to claim 2, characterized in that, The second extreme high temperature intensity index is extracted based on multi-mode, multi-initial-field return test data, and a moving average and regional average are calculated for a set number of years to obtain the second regional average index for each initial field, including: Acquire multi-mode, multi-initial-field return test data and calculate the second extreme high-temperature intensity index using the relative threshold method; The second extreme high temperature intensity index is averaged over a set number of years to obtain the second decadal variability. Based on the above-described regional divisions, the second interdecadal variability is averaged regionally to obtain the second regional average index for each initial field.

4. The method for predicting extreme high temperatures in different decades according to claim 3, characterized in that, The multi-mode, multi-initial-field return test data are derived from the return test data of multiple modes in the sixth Coupled Mode Comparison Program decadal forecasting program experiment.

5. The method for predicting extreme high temperatures in different decades according to claim 3, characterized in that, The method for selecting the average index of multiple second-region regions with the highest ranking of initial in-field prediction skills includes: Using the evaluation method of anomaly correlation coefficient and root mean square error skill scoring, the spatial scale prediction skill of the second decadal variability and the first decadal variability of each initial field is evaluated, and the temporal scale prediction skill of the second decadal variability after regional averaging and the regional average observation index is evaluated, and the evaluation results are obtained. Based on the evaluation results, the average index of the top-ranked second-region prediction skills in each initial field was extracted.

6. The method for predicting extreme high temperatures in different decades according to claim 1, characterized in that, The structure of the Transformer-based deep learning model includes: A linear projection layer is used to convert the average exponent of multiple second regions of the input into vector features; A multi-layer Transformer encoder is used to extract the temporal and cross-regional dependencies of vector features using a global attention mechanism, resulting in extracted vector features. A multilayer perceptron is used to perform regression prediction on the extracted vector features and output a predicted sequence.

7. The method for predicting extreme high temperatures in different decades according to claim 1, characterized in that, The process of reconstructing the interdecadal prediction sequences of extreme high-temperature intensity in each region onto a spatial grid using a linear projection method to obtain the spatial distribution prediction results of extreme high-temperature intensity within the target geographical region includes: During the training period, the first region mean index is processed to remove the linear trend. The detrended first region mean index is used to regress the detrended first interdecadal variability of each grid point in the region, and the variability regression coefficient and intercept are calculated. Linear regression was performed on the first interdecadal variability of each grid point to obtain the regression coefficients and intercepts of the linear trend; Based on the regression coefficients and intercepts of the variable rate and the regression coefficients and intercepts of the linear trend, the spatial distribution prediction results of each region are calculated using the linear projection formula. By merging the spatial distribution prediction results of each region, we obtain the spatial distribution prediction results of extreme high temperature intensity within the target geographical region.

8. The method for predicting extreme high temperatures in different decades according to claim 7, characterized in that, The calculation formula for the linear projection formula is as follows: ； in, For the first i The grid point at the th grid point t The projected value for the year, For the predicted sequence of the th t The value of the year, and The first i The rate of change regression coefficients and intercepts for each grid point; and The first i Linear trend regression coefficients and intercepts for each grid point.

9. An extreme high-temperature interdecadal prediction device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the method according to any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1-8.

Images