A Method for Temperature Correction Forecasting Based on Short-Term Real-Time Information Calibration in Large Meteorological Models

Abstract

This invention relates to a method for correcting temperature forecasts using a large meteorological model based on short-term real-time information calibration. The method includes: S1, data collection; S2, data preprocessing and sample set creation; S3, temperature correction forecast model construction; S4, sample set partitioning and model training and testing; and S5, outputting the temperature forecast result after short-term real-time calibration. This invention, by fully utilizing the prior information of the large meteorological model forecast, introduces residual information between short-term real-time data and the forecast as a constraint, revealing the spatiotemporal correlation structure of the modeling error and its evolution with forecast lead time. Furthermore, it enables the correction process to adaptively adjust the calibration strategy according to the error state, thereby improving the accuracy, stability, and consistency of temperature forecasts and enhancing the operational application level of large meteorological model temperature forecast calibration technology.

Description

Technical Field

[0001] This invention belongs to the field of meteorology and machine learning technology, specifically relating to a method for temperature correction forecasting of a large meteorological model based on short-term real-time information calibration. Background Technology

[0002] Temperature forecasting plays a crucial role in meteorological forecasting, disaster prevention and mitigation, energy dispatching, agricultural production, and urban operation support. Its task is to forecast temperature elements at the grid level for a given area based on observational data and forecast model output. Accurate temperature forecasts are a vital foundation for generating operational forecast products and providing risk alerts, and are of great significance for monitoring disastrous events such as cold waves and heat waves.

[0003] In recent years, large-scale meteorological models have demonstrated significant advantages in rapid inference across multiple variables and timeframes, enabling rapid forecast output at global or regional scales with relatively low computational costs, providing high-frequency and rapid forecast support for operational needs. In practical operation, large-scale meteorological models typically output gridded temperature forecasts for the next several days at fixed start times (e.g., 00:00 and 12:00 daily). However, influenced by factors such as differences in training data distribution, insufficient characterization of boundary layer processes, and systematic biases in the models, temperature forecasts from large-scale meteorological models still exhibit significant systematic biases and regional errors, particularly in areas with complex topography, land-sea interfaces, urban heat islands, and during the evolution of severe weather events. As forecast lead times increase, the accumulation of errors and temporal drift problems are further amplified, limiting the stability, consistency, and operational availability of temperature forecasts.

[0004] Relying solely on temperature forecasts provided by large meteorological models at the time of forecast launch, while ignoring the constraints imposed by short-term real-time information on error evolution, limits the upper limit of correction effectiveness. Temperature forecast errors exhibit significant background dependence; different circulation patterns, flux states, boundary layer stability, and topographical variations lead to distinctly non-stationary error structures in both space and time. Furthermore, errors propagate and accumulate across forecast lead times; independent corrections for each lead time can easily result in unstable correction magnitudes and insufficient temporal consistency, thereby impacting the reliability and usability of operational products.

[0005] Therefore, there is an urgent need to develop a method for calibrating temperature forecasts of large meteorological models based on short-term real-time information. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by providing a method for correcting temperature forecasts in large meteorological models based on short-term real-time information calibration. This method fully utilizes the prior knowledge of large meteorological models and introduces residual information between short-term real-time data and forecasts as a constraint, revealing the spatiotemporal correlation structure of modeling errors and their evolution with forecast lead time. Furthermore, it enables the correction process to adaptively adjust the calibration strategy according to the error state, thereby improving the accuracy, stability, and consistency of temperature forecasts and enhancing the operational application level of large meteorological model temperature forecast calibration technology.

[0007] The technical problem solved by this invention is achieved through the following technical solution: A method for temperature correction forecasting based on a large meteorological model calibrated with short-term real-time information, comprising the following steps: S1. Using a daily reporting cycle, collect the temperature forecast data of the meteorological big model at 00:00 every day for the next ten days as the temperature forecast data of the meteorological big model, and simultaneously collect the actual temperature data for the corresponding next ten days; collect the actual temperature data at 01:00 every day and the temperature forecast data of the meteorological big model at the corresponding time in the first hour after the reporting starts. S2. Spatial grid unification processing is performed on the temperature forecast data and actual temperature data collected in S1 from the large-scale meteorological model for the next ten days. The bilinear interpolation method is used to convert them into a unified target grid field. Based on the actual temperature data of the first hour after the start of the forecast and the temperature forecast data of the large-scale meteorological model at the same time, the residual field data is calculated and used as the condition input for short-term actual calibration. The temperature forecast data of the large-scale meteorological model for the next ten days is used as input, and the corresponding actual temperature data for the next ten days is used as labels to create a sample set. S3. Using the temperature forecast data of the meteorological big data model for the next ten days as input and the residual field data of the first hour after the start of the forecast as conditional input, a temperature correction forecast model is built. The temperature correction forecast model includes a forecast sequence encoding module, a residual field conditional encoding module and a correction decoding module. The residual field conditional encoding module outputs a conditional vector to perform conditional adaptive modulation on the normalization layer inside the correction forecast model, so as to achieve adaptive calibration under different error backgrounds. S4. Divide the sample set into a training set, a validation set, and a test set. Train the temperature correction forecast model on the training set, perform hyperparameter tuning on the validation set, and test and evaluate the performance of the temperature correction forecast model on the test set. Evaluation metrics include area-weighted mean absolute error and root mean square error. S5. Input the temperature forecast data of the meteorological big model for the next ten days at any given reporting time, and obtain the actual temperature data for the first hour after reporting and the temperature forecast data of the meteorological big model at the same time; use S2 to calculate the residual field data, and input the temperature forecast data of the meteorological big model for the next ten days and the residual field data together into the temperature correction forecast model trained in S4. The model output is the temperature forecast result after short-term actual calibration.

[0008] Furthermore, the meteorological large model temperature forecast data and actual temperature data collected by S1 at 00:00 every day for the next ten days are grid data with equal latitude and longitude. The bilinear interpolation method is used to transform them into an equidistant grid field or an operational grid field under the target grid. If the spatial resolution of the meteorological large model temperature forecast data and the actual temperature data for the next ten days is different, the two are first unified to the same spatial resolution and the same latitude and longitude range, and the grid points are guaranteed to correspond point by point according to the latitude and longitude information. S2 calculates residual field data based on the actual temperature data of the first hour after the start of the report and the temperature forecast data of the meteorological large model at the same time. The residual field data is used to characterize the current error state and serve as the condition input for short-term actual calibration. The residual field data is calculated as follows: the actual temperature data for the first hour after the start of the reporting period is subtracted from the temperature forecast data from the meteorological large model for the first hour after the start of the reporting period, grid by grid, to obtain the residual field data. : ; in: This is the actual temperature data for the first hour; The system uses temperature forecast data from a large meteorological model at the same time as input, residual field data as input, temperature forecast data from the large meteorological model for the next ten days as input, and actual temperature data for the next ten days as labels to create a sample set. The system then matches the data with the starting time and latitude / longitude grid to form a sample set that corresponds one-to-one with "forecast sequence - residual field conditions - actual label".

[0009] Furthermore, the residual field condition encoding module of S3 is used to abstract the residual field data into a global condition vector. The forecast sequence encoding module is used to perform spatiotemporal modeling on the forecast data; the correction decoding module is used to output the correction residual and add it to the temperature forecast data of the meteorological big model for the next ten days to obtain the final calibration result. The residual field conditional encoding module consists of several convolutional layers, used to extract deep features from the residual field data, and to aggregate spatial information into a sample-level representation through global average pooling, and then output a global conditional vector through a fully connected layer. The global condition vector This is used to control the scaling and bias parameters of the normalization layer in subsequent networks, enabling the model to adaptively adjust the correction strength and correction mode according to the current error state, and to use the global condition vector. Mapped to scaling parameters of the normalization layer With bias parameters , and The calculation method is as follows: ; in: Here is the weight matrix of the trained linear layer. For the bias of the linear layer; the and Broadcasting across all timeframes and spaces in the dimension enables samples to achieve consistent feature modulation under error conditions, thus realizing conditional modulation of deep features. The normalized calculation method that can adaptively adjust the correction strength and correction mode according to the current error state is as follows: ; in: Indicates the characteristics of the token The normalization result; This indicates element-wise multiplication.

[0010] Furthermore, the prediction sequence encoding module consists of 12 spatiotemporal blocks (ST-Blocks) stacked together. Each ST-Block includes a windowed spatial and temporal attention module and a feedforward network module, and stable training is achieved through residual connections. The windowed spatial and temporal attention module restores the token to a 2D grid, then divides the 2D grid into several 8×8 windows, and adds adaptive spatial and temporal attention within the windows, while modeling the error propagation relationship in the time dimension. The feedforward network module is used for nonlinear feature transformation and representation enhancement, and the calculation formula is as follows: ; ; ; ; ; in: For spatial attention function, the token grid is divided into 8×8 windows in each time period and adaptive spatial attention is performed within the window to learn local spatial correlation. As a time attention function, adaptive time attention is applied to the time sequence of the same spatial token position across 10 days, learning the error propagation law over time; To perform features after the adaptive spatial attention sublayer, Features after the adaptive temporal attention sublayer is completed.

[0011] The advantages and beneficial effects of this invention are as follows: 1. The sample of this invention consists of "short-term actual conditions - 10-day lead-time forecast sequence - 10-day lead-time actual label", which is a vector obtained through conditional encoding. Conditional adaptive modulation is applied to the normalization layer within the network, enabling the model to adaptively adjust the correction strategy according to the current error state. Simultaneously, patch embedding, two-dimensional sin-cos spatial location encoding, temporal embedding, and an ST-Block structure composed of windowed spatial and temporal attention are employed to explicitly learn the spatial correlation structure of the error and its propagation over time. The calibration results are output in the form of residual learning, thereby improving the accuracy, stability, and temporal consistency of temperature forecasts while maintaining forecast rationality.

[0012] 2. This invention can fully utilize the prior information of meteorological large model forecasts, introduce residual information between short-term actual conditions and forecasts after the start of the forecast as a condition constraint, display the spatiotemporal correlation structure of modeling errors and their evolution law with the forecast lead time, and enable the correction process to adaptively adjust the calibration strategy according to the error state, thereby improving the accuracy, stability and consistency of temperature forecasts and enhancing the operational application level of meteorological large model temperature forecast calibration technology. Attached Figure Description

[0013] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the temperature correction prediction model of the present invention; Figure 3 This is a comparison chart of the prediction errors of the present invention and the Pangu meteorological model. Detailed Implementation

[0014] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0015] like Figure 1 As shown, this invention provides a method for temperature correction forecasting based on short-term real-time information calibration of a large meteorological model. Its innovation lies in the following: based on the temperature forecast data output by the large meteorological model at a fixed start time, i.e., 00:00 every day, real-time information from the first hour after the start of the forecast is introduced to construct residual field conditions, and the future temperature forecast data is stably calibrated by a spatiotemporal Transformer correction network modulated by conditional adaptive normalization (AdaLN) to obtain operationally usable temperature correction results.

[0016] The steps of the method are as follows: (1) Data Acquisition and Sample Organization. Temperature forecast data output by the global Pangu meteorological model at 00:00 daily from January 2017 to December 2019 was collected as a forecast data file, and the actual temperature data corresponding to that time was simultaneously collected as a tag data file. Further, in the first hour after the start of the forecast, the actual temperature data and the corresponding temperature forecast data from the meteorological model were collected to construct a short-term actual calibration conditional data file. In this embodiment, samples are stored in a one-to-one correspondence using triplet files: a forecast data file, a tag data file, and a conditional data file. The forecast data file contains gridded temperature forecast data for the next 10 days, the tag data file contains gridded temperature data for the next 10 days, and the conditional data file contains the actual temperature data for the first hour after the start of the forecast and the temperature forecast data from the meteorological model at the same time. These three files are matched with the date and start time to form a sample set.

[0017] (2) Grid unification, regional pruning, and short-term residual field data construction. The output of the meteorological large model and the actual data are subjected to grid unification processing to obtain the target area data, and short-term actual residual field data is constructed as conditional information. Let the actual temperature data for the first hour after the start of reporting be... The temperature forecast data from the large meteorological model at the same time are: Then the residual field data Defined as: ; in: The spatial grid index is used to represent the current error background and serves as the conditional input for subsequent correction networks. To improve training stability, this embodiment calculates the channel mean and standard deviation for the forecast data file and performs channel-by-channel normalization. Let any channel data be... Its normalized form is: ; In this embodiment and These are the mean and standard deviation of the channel, respectively. It is a small, non-zero constant.

[0018] (3) Construct a temperature correction forecast model with temperature forecast data from a large meteorological model for the next ten days as input, residual field data as conditional input, and actual temperature data for the next ten days as output.

[0019] First, the residual field conditional encoding module: taking the normalized conditional field as input, it extracts features through several convolutional layers, obtains a sample-level global representation through global average pooling, and then obtains the conditional vector through fully connected mapping. The global condition vector This is used to control the scaling and bias parameters of the normalization layer in subsequent networks, enabling the model to adaptively adjust the correction strength and correction mode according to the current error state, and to control the global conditional vector. Mapped to scaling parameters of the normalization layer With bias parameters , and The calculation method is as follows: ; in: Here is the weight matrix of the trained linear layer. For the bias of the linear layer; the and Broadcasting across all timeframes and spatial dimensions enables consistent feature modulation of samples under error conditions, achieving conditional modulation of deep features; the calculation method for adaptively adjusting correction intensity and correction mode normalization based on the current error state is as follows: ; in: Indicates the characteristics of the token The normalization result; Indicates element-wise multiplication; Secondly, the forecast sequence coding module performs spatiotemporal tokenization and coding on the temperature forecast data of the meteorological big data model for the next ten days. It divides the temperature grid field of each timeframe into several 16×16 patches spatially, and maps each patch to a token vector, forming (…). Token data in the form of ); where, For batch size, To predict the number of steps in advance, For the number of patches, The token feature dimension is defined as follows: A two-dimensional sin-cos spatial location code and a learnable temporal embedding vector are superimposed on the token data, specifically represented as follows: ; in: for( token data in the form of ) Represented as a two-dimensional spatial location encoding vector, It is a learnable temporal embedding vector.

[0020] The prediction sequence encoding module consists of 12 stacked spatiotemporal blocks (ST-Blocks). Each ST-Block includes a windowed spatial and temporal attention module and a feedforward network module, and stable training is achieved through residual connections. The windowed spatial and temporal attention module restores the token to a 2D grid, then divides the 2D grid into several 8×8 windows, and adds adaptive spatial and temporal attention within the windows, while modeling the error propagation relationship in the time dimension. The feedforward network module is used for nonlinear feature transformation and representation enhancement, and the calculation formula is as follows: ; ; ; ； ; in: For spatial attention function, the token grid is divided into 8×8 windows in each time period and adaptive spatial attention is performed within the window to learn local spatial correlation. As a time attention function, adaptive time attention is applied to the time sequence of the same spatial token position across 10 days, learning the error propagation law over time; To perform features after the adaptive spatial attention sublayer, Features after the adaptive temporal attention sublayer is completed.

[0021] Finally, the correction decoding module restores the token data output by ST-Block into grid residual field data. The transposed convolution is used for depatch to restore the original spatial resolution, where the transposed convolution kernel size and stride are both patch_size=16. The padding region is then cropped to restore the original resolution. Finally, residual learning is used to add the residuals back into the large model prediction to obtain the calibration output, which is in the following format: ; in: The temperature forecast data output by the large meteorological model. For the corrected residual data predicted by the model, This is temperature forecast data after short-term real-time calibration.

[0022] The temperature correction prediction model structure of this invention is as follows: Figure 2As shown in the diagram, part A represents the input, which consists of temperature forecast data from a large-scale meteorological model for the next ten days and residual field data. Part B illustrates how the residual data undergoes operations such as convolution and pooling to ultimately obtain the global conditional vector. Part C of the figure shows the spatiotemporal tokenization and encoding of the temperature forecast data from the large-scale meteorological model for the next ten days. Part D of the figure shows the forecast sequence module, and the final correction result is obtained by adding the output residuals to the large-scale model forecast. The input parameter settings are shown in Table 1.

[0023] Table 1 Encoder and Decoder Parameter Settings

[0024] (4) Divide the sample set into a training set, a validation set and a test set, train the temperature correction prediction model on the training set, perform hyperparameter tuning on the validation set, and test the model on the test set.

[0025] The training set used 10-day temperature forecasts from the Pangu meteorological model (January 2017 to December 2019) at midnight as input, and Pangu model temperature forecasts for the first hour after the forecast and the corresponding actual temperature data at that time as conditional inputs. The actual temperature data for the next 10 days was used as labels, resulting in 1095 samples. A 10% random subset of the 2020 data was used as the validation set, and the remaining 2020 data was used as the test set. The final training set consisted of 1095 samples, the validation set of 36 samples, and the test set of 330 samples.

[0026] The temperature correction model is trained using the inversely normalized mean squared error (MSE) of the true temperature domain as the loss function. The MSE is defined as follows: ; in: The predicted temperature after inverse normalization. This is a true temperature label. Because meteorological data is usually presented in grid format, and the actual ground area represented by the latitude grid is different, an area-weighted approach is used to evaluate the model. The formula is: ; ; in: Indicates the first Forecast lead time, grid points The error between the predicted value and the actual value, For grid points The predicted value on, For grid points Upload the actual value, For forecast lead time index, For spatial grid indexing, Area weights are used for "area correction" in global grid statistics. For the first The latitude value corresponding to the row.

[0027] This embodiment uses area-weighted mean absolute error (MAE) and root mean square error (RMSE) as evaluation metrics to represent the difference between the predicted and actual temperatures, wherein: ; ; in: , For grid size, To traverse all grid points of the entire input, It is a finite value indicator function, which is used to skip missing and illegal grid points, and avoid including invalid points in the indicator, which would cause the result to crash.

[0028] This embodiment uses temperature data to train the corresponding temperature correction forecast model. The training process is implemented based on the PyTorch architecture, with a learning rate of 0.0001, a batch size of 1, and 50 training epochs for each model. The model with the lowest MSE on the validation set corresponding to the lowest number of epochs is selected as the final model for testing. The area-weighted mean absolute error and root mean square error of the tests are shown in Table 2. Error comparisons are shown below. Figure 3 express.

[0029] Table 2. Error magnitude under different forecast days

[0030] From Table 2 and Figure 3 It can be seen that: (1) From the perspective of the daily error change with time, the error of the Pangu meteorological model forecast accumulates significantly with the number of days: its MAE increases from 0.36 on day 1 to 1.59 on day 10, and RMSE increases from 0.56 to 2.65; the temperature correction forecast model of the present invention can effectively suppress the error growth in the middle and late stages, and the RMSE of the method of the present invention is lower than that of the Pangu meteorological model from day 2 onwards and the gap widens in the middle and late stages.

[0031] (2) The correction gain is most stable in the medium term: Compared with the Pangu meteorological model, the MAE of the method of this invention decreases by an average of about 10.10% in the 3-10 days, and the RMSE decreases by an average of about 10.11% in the 3-10 days. The improvement increases with time: the MAE decreases by about 12.07%-16.35% in the 7-10 days, and the RMSE decreases by about 11.73%-16.23% in the 7-10 days, indicating that the method of this invention has a stronger ability to suppress the accumulation of errors in the medium term.

[0032] (3) Taking the long-term results as an example on day 10, the correction model of the present invention reduced the MAE from 1.59 to 1.33 (a relative decrease of 16.35%) and the RMSE from 2.65 to 2.22 (a relative decrease of 16.23%). Combined with residual learning, it can output an effective correction for the accumulation of medium-term errors while maintaining the consistency of the background at the scale of the Pangu meteorological model, thereby improving the operational availability and stability of temperature forecasts.

[0033] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for temperature correction forecasting based on a large meteorological model calibrated with short-term real-time information, characterized in that: The steps of the method are as follows: S1. Using a daily reporting cycle, collect the temperature forecast data of the meteorological big model at 00:00 every day for the next ten days as the temperature forecast data of the meteorological big model, and simultaneously collect the actual temperature data for the corresponding next ten days; collect the actual temperature data at 01:00 every day and the temperature forecast data of the meteorological big model at the corresponding time in the first hour after the reporting starts. S2. Spatial grid unification processing is performed on the temperature forecast data and actual temperature data collected in S1 from the large-scale meteorological model for the next ten days. The bilinear interpolation method is used to convert them into a unified target grid field. Based on the actual temperature data of the first hour after the start of the forecast and the temperature forecast data of the large-scale meteorological model at the same time, the residual field data is calculated and used as the condition input for short-term actual calibration. The temperature forecast data of the large-scale meteorological model for the next ten days is used as input, and the corresponding actual temperature data for the next ten days is used as labels to create a sample set. S3. Using the temperature forecast data of the meteorological big data model for the next ten days as input and the residual field data of the first hour after the start of the forecast as conditional input, a temperature correction forecast model is built. The temperature correction forecast model includes a forecast sequence encoding module, a residual field conditional encoding module and a correction decoding module. The residual field conditional encoding module outputs a conditional vector to perform conditional adaptive modulation on the normalization layer inside the correction forecast model, so as to achieve adaptive calibration under different error backgrounds. S4. Divide the sample set into a training set, a validation set, and a test set. Train the temperature correction forecast model on the training set, perform hyperparameter tuning on the validation set, and test and evaluate the performance of the temperature correction forecast model on the test set. Evaluation metrics include area-weighted mean absolute error and root mean square error. S5. Input the temperature forecast data of the meteorological big model for the next ten days at any given reporting time, and obtain the actual temperature data for the first hour after reporting and the temperature forecast data of the meteorological big model at the same time; use S2 to calculate the residual field data, and input the temperature forecast data of the meteorological big model for the next ten days and the residual field data together into the temperature correction forecast model trained in S4. The model output is the temperature forecast result after short-term actual calibration.

2. The method for temperature correction forecasting based on short-term real-time information calibration of a large meteorological model according to claim 1, characterized in that: The meteorological large model temperature forecast data and actual temperature data collected by S1 at 00:00 every day for the next ten days are equal latitude and longitude grid data. They are converted into an equidistant grid field or an operational grid field under the target grid using bilinear interpolation. If the spatial resolution of the meteorological large model temperature forecast data and the actual temperature data for the next ten days is different, they are first unified to the same spatial resolution and the same latitude and longitude range, and the grid points are guaranteed to correspond point by point according to the latitude and longitude information. S2 calculates residual field data based on the actual temperature data of the first hour after the start of the report and the temperature forecast data of the meteorological large model at the same time. The residual field data is used to characterize the current error state and serve as the condition input for short-term actual calibration. The residual field data is calculated as follows: the actual temperature data for the first hour after the start of the reporting period is subtracted from the temperature forecast data from the meteorological large model for the first hour after the start of the reporting period, grid by grid, to obtain the residual field data. : ； in: This is the actual temperature data for the first hour; The system uses temperature forecast data from a large meteorological model at the same time as input, residual field data as input, temperature forecast data from the large meteorological model for the next ten days as input, and actual temperature data for the next ten days as labels to create a sample set. The system then matches the data according to the start time and latitude / longitude grid to form a sample set that corresponds one-to-one with "forecast sequence - residual field conditions - actual label".

3. The method for temperature correction forecasting based on short-term real-time information calibration of a large meteorological model according to claim 1, characterized in that: The residual field condition encoding module of S3 is used to abstract the residual field data into a global condition vector. The forecast sequence encoding module is used to perform spatiotemporal modeling on the forecast data. The correction decoding module is used to output the correction residual and add it to the temperature forecast data of the meteorological big model for the next ten days to obtain the final calibration result; The residual field conditional encoding module consists of several convolutional layers, used to extract deep features from the residual field data, and to aggregate spatial information into a sample-level representation through global average pooling, and then output a global conditional vector through a fully connected layer. The global condition vector This is used to control the scaling and bias parameters of the normalization layer in subsequent networks, enabling the model to adaptively adjust the correction strength and correction mode according to the current error state, and to use the global condition vector. Mapped to scaling parameters of the normalization layer With bias parameters , and The calculation method is as follows: ； in: Here is the weight matrix of the trained linear layer. For the bias of the linear layer; the and Broadcasting across all timeframes and spaces in the dimension enables samples to achieve consistent feature modulation under error conditions, thus realizing conditional modulation of deep features. The normalized calculation method that can adaptively adjust the correction strength and correction mode according to the current error state is as follows: ； in: Indicates the characteristics of the token The normalization result; This indicates element-wise multiplication.

4. The method for temperature correction forecasting based on short-term real-time information calibration of a large meteorological model according to claim 3, characterized in that: The prediction sequence encoding module consists of 12 stacked spatiotemporal blocks (ST-Blocks). Each ST-Block includes a windowed spatial and temporal attention module and a feedforward network module, and stable training is achieved through residual connections. The windowed spatial and temporal attention module restores the token to a 2D grid, then divides the 2D grid into several 8×8 windows, and adds adaptive spatial and temporal attention within the windows, while modeling the error propagation relationship in the time dimension. The feedforward network module is used for nonlinear feature transformation and representation enhancement, and the calculation formula is as follows: ； ； ； ； ； in: For spatial attention function, the token grid is divided into 8×8 windows in each time period and adaptive spatial attention is performed within the window to learn local spatial correlation. As a time attention function, adaptive time attention is applied to the time sequence of the same spatial token position across 10 days, learning the error propagation law over time; To perform features after the adaptive spatial attention sublayer, Features after the adaptive temporal attention sublayer is completed.

Images