LGU-Net meteorological element correction system and temperature forecast correction method

Through the combination of U-Net model, LSTM structure and GAM attention mechanism, the LGU-Net meteorological element correction system solves the problem of large errors in traditional meteorological forecasts, making full use of multi-element and multi-time phase characteristics to achieve high-precision meteorological data Correction.

CN118962861APending Publication Date: 2024-11-15CHANGGUANG SATELLITE TECH CO LTD
View PDF 0 Cites 7 Cited by

Patent Information

Application Number
CN202411011392.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

The traditional meteorological forecasting model has errors in complex climate systems. The manual correction efficiency is low, the method is complex, and the correlation characteristics between multiple elements and multiple time phases cannot be fully utilized, resulting in large forecast errors.

Method used

The LGU-Net meteorological element correction system, which combines the LSTM structure and GAM attention mechanism, extracts the spatial and timing characteristics of meteorological elements, enhances feature extraction and model performance through the GAM attention mechanism, and optimizes the LSTM structure to reduce computational complexity. Spend.

Benefits of technology

Accurate correction of GFS meteorological mode forecast data was achieved, and the 24-hour forecast error dropped to 0.7503, improving the accuracy of the meteorological factor correction algorithm.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118962861A_ABST
    Figure CN118962861A_ABST
Patent Text Reader

Abstract

The invention discloses an LGU-Net meteorological element correction system and a temperature forecast correction method, relates to the technical field of meteorological forecast correction, and solves the problems that a traditional correction algorithm is large in calculated amount and complex in model. The system comprises a U-Net, an LSTM and a GAM attention mechanism. The U-Net structure is composed of an encoder and a decoder which are connected through a jump link, the LSTM structure is arranged at the jump link part, a Xavier weight initialization method and a gradient cutting method are adopted to stabilize gradient, and a distributed training strategy is adopted and a GRU variant is simulated to optimize the LSTM structure; a GAM attention mechanism is arranged in a decoder part, and double MLP is used for amplifying dependence of a cross-dimension channel and space. The method comprises the following steps: firstly, preprocessing input data; constructing a multi-temporal multi-element temperature correction data set; the application system trains the data set to obtain a correction model; and correcting the temperature data by using the correction model. The method can be applied to correction work of temperature forecast.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of weather forecast correction, and in particular to an LGU-Net weather element correction system and a temperature forecast correction method based on an improved U-Net model. Background Art

[0002] With the continuous development of social economy, the impact of weather on various fields has become increasingly prominent, so the accuracy of weather forecasts has become increasingly critical. In recent decades, numerical weather forecasting has made revolutionary progress and has become an indispensable part of meteorological forecasting services, and continues to develop in a more refined direction. However, traditional model forecasts may have some limitations when facing complex climate systems, such as complex terrain effects, the accuracy of local weather phenomena, etc., coupled with the uncertainty of the initial field of the numerical model simulating the atmosphere and the defects of the model itself, the forecast error of the numerical model still exists. Therefore, it is of great significance to revise the model forecast to obtain more accurate results.

[0003] However, there are also certain problems in data correction: (1) The correction energy of manual forecasters is limited and their levels are uneven, which may cause human errors to be mixed in the model errors; (2) Traditional methods need to combine physical simulation and set various assumptions at the same time, which is cumbersome to operate and the model is complex; (3) The current error correction methods cannot take into account the relationship characteristics between multiple meteorological elements. In the entire weather process, various meteorological elements influence each other, resulting in the loss of a large amount of correlation information; (4) Model forecast data also contains a large number of spatial and temporal features, and the current error correction methods do not fully extract these features.

[0004] In summary, how to make full use of the correlation characteristics between multiple factors and multiple time phases to improve the accuracy of weather forecasts is a key technical issue that urgently needs to be solved. Summary of the invention

[0005] Aiming at the problems of large computational complexity and complex models in traditional algorithms, the present invention uses the U-Net model as the basic framework and develops a LGU-Net meteorological element correction system and a temperature forecast correction method, which realizes the correction of temperature data forecasted by the GFS meteorological model.

[0006] The technical solution of the present invention is as follows:

[0007] A LGU-Net meteorological element correction system, including a U-Net structure, an LSTM structure, and a GAM attention mechanism;

[0008] The U-Net structure is a symmetrical U-shaped network structure consisting of an encoder and a decoder connected by skip links, which is used to extract the spatial distribution characteristics of meteorological elements; the encoder is composed of convolutional layers and pooling layers, and at each encoder stage, the size of the feature map is gradually reduced and the number of channels is gradually increased to capture meteorological feature information at different levels; the decoder is composed of upsampling layers and convolutional layers, which is used to restore the meteorological feature map extracted in the encoder to the size of the original input image;

[0009] The LSTM structure is arranged in the jump link part of the U-Net structure, and is used to extract the relevant features between the multi-time series meteorological images, and input the multi-time series meteorological features of the encoder stage into the decoder stage; the LSTM structure adopts the Xavier weight initialization method and the gradient clipping method to stabilize the gradient, adopts the distributed training strategy, and imitates the GRU variant to optimize the LSTM structure;

[0010] The GAM attention mechanism is added to the upsampling layer of the decoder, including a channel attention module and a spatial attention module; the spatial attention module uses two convolutions to focus on spatial information; the channel attention module uses 3D arrangement to retain three-dimensional information; and dual MLP is used to amplify the dependency between cross-dimensional channels and space.

[0011] Preferably, the LSTM structure optimized by imitating the GRU variant is as follows:

[0012] Merge the input gate and forget gate of LSTM; optimize the LSTM structure by sharing some parameters or reducing redundant parameters.

[0013] Preferably, when adding the GAM attention mechanism, the optimal hyperparameter configuration of GAM is determined by the Bayesian optimization algorithm to ensure the convergence and performance of the model. In addition, the attention mechanism is also optimized, and efficient matrix operations and parallel computing techniques are used to ensure that the additional computing overhead is reduced as much as possible while improving the model performance. Finally, the strategy of gradually increasing the learning rate is adopted to stabilize the initial training process, and the adaptive learning rate optimizer is used to dynamically adjust the learning rate to ensure the stability and effectiveness of the training process, and avoid the problem of slow convergence due to the introduction of the GAM module.

[0014] The present invention also provides a temperature forecast correction method, which uses the above-mentioned LGU-Net meteorological element correction system, and the correction method specifically includes the following steps:

[0015] S1. Preprocess the input data:

[0016] S2, construct multi-phase and multi-factor input data as the temperature correction data set;

[0017] S3, using the LGU-Net meteorological element correction system to train the temperature correction data set to obtain the LGU-Net temperature correction model of the correction area based on the GFS model prediction data;

[0018] S4. Use the trained LGU-Net temperature correction model to correct the hourly temperature data predicted by the GFS model.

[0019] Preferably, the step S1 specifically includes:

[0020] The GFS forecast data of the correction area is used as the data to be corrected, and preprocessed to obtain multi-factor TIF images with latitude and longitude information at hourly intervals;

[0021] The terrain data of the correction area in the satellite remote sensing data is preprocessed to obtain remote sensing image data.

[0022] Preferably, the multiple factors include temperature, humidity, wind speed, pressure and precipitation factors.

[0023] Preferably, the step S2 specifically includes:

[0024] In the time dimension, 1 hour is used as the time interval and 24 hours as a time period, using the data processed by GFS;

[0025] In the multi-factor dimension, humidity, wind speed, pressure, and precipitation are all input as auxiliary factors of temperature;

[0026] In this way, a 5×24×X×Y multi-phase and multi-factor input data set is constructed as the temperature correction data set, where X and Y represent the length and width of the TIF image, respectively.

[0027] Compared with the prior art, the specific beneficial effects of the present invention are:

[0028] Aiming at the situation that the traditional weather forecast model has large errors and the existing correction algorithm cannot effectively utilize the related features between multiple factors and multiple time phases, the present invention uses the U-Net model as the framework, adds the GAM attention mechanism to extract the related features between multiple factors, adds the LSTM module to capture the time series information in the multi-time phase data, and adds data such as terrain and land utilization rate to supplement the spatial features of the data, thus realizing the error correction of temperature data. The average error of the corrected 24-hour forecast GFS data compared with the real data is only 0.7503, which improves the highest accuracy of the meteorological element correction algorithm in the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is the structural diagram of the LGU-Net model;

[0030] Figure 2 This is the structural diagram of the LSTM model;

[0031] Figure 3 This is the structure diagram of the GAM attention mechanism;

[0032] Figure 4 This is the flow chart of the 2m temperature correction method for the Northeast region of the GFS model;

[0033] Figure 5 Provides a visual display of corrected temperature data and true temperature data. DETAILED DESCRIPTION

[0034] In order to make the technical solution of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the specification of the present invention. It should be noted that the following embodiments are only used to better understand the technical solution of the present invention and should not be understood as a limitation to the present invention.

[0035] Example 1.

[0036] This embodiment combines the data characteristics of multiple factors and multiple time phases to build a LGU-Net meteorological element correction system. The system structure diagram is shown in Figure 1 As shown in the figure, the overall structure is mainly divided into three parts: U-Net structure, LSTM structure and GAM attention mechanism.

[0037] First, in the cross-link part of the U-Net model, an LSTM structure is added to extract the features of multi-time series images, and the multi-time series features of the encoding stage are input into the decoder stage, making full use of the correlation features between multi-time series images. Then, in order to make full use of the correlation between multiple elements and reduce the information loss in the decoding part, a GAM attention mechanism is added to the decoding part to fully integrate the features between multiple bands and improve the ability of the decoder part to recover important information.

[0038] (1) The U-Net model is a deep learning architecture for semantic segmentation. Its structure is divided into two parts: encoder and decoder, forming a symmetrical U-shaped structure connected by skip connections. The encoder part consists of a series of convolutional layers and pooling layers. In each encoder stage, the size of the feature map is gradually reduced and the number of channels is gradually increased to capture feature information at different levels. The decoder part consists of a series of upsampling layers and convolutional layers, which are used to restore the feature map extracted in the encoder to the size of the original input image. Skip connections help retain multi-scale feature information and improve the accuracy of segmentation results.

[0039] (2) Use LSTM network to extract relevant features between multiple time phases. Its structure is as follows: Figure 2As shown in the figure. The LSTM network can dynamically update the hidden state and memory cells to capture the timing information in the sequence data. The forget gate and input gate are responsible for regulating the retention of historical information and the update of new information, so that the network can effectively learn the key features in the data. After training, feature vectors can be extracted from the hidden layer or output layer of the LSTM network. These vectors contain important information in the sequence data.

[0040] When inserting the LSTM structure, the feature map of the convolution output needs to be serialized into the input format of the time step, and then the output after LSTM processing needs to be reshaped to adapt to the subsequent convolution operation. In addition, in order to deal with the problem of gradient disappearance or explosion that may occur when LSTM processes long sequence data, the Xavier weight initialization method and gradient clipping method are used to stabilize the gradient. Finally, in order to reduce the complexity and training time of the model increased by the LSTM structure, a distributed training strategy is adopted, and the structure of LSTM is optimized by imitating the GRU variant, and the input gate and forget gate of LSTM are merged, which is similar to the update gate of GRU; by sharing some parameters or reducing redundant parameters, the LSTM structure is further optimized to reduce the consumption of computing resources while maintaining performance.

[0041] (3) The attention mechanism can focus on basic features and suppress unnecessary features, thereby improving the model's ability to extract features. GAM includes a channel attention module and a spatial attention module, which can amplify cross-dimensional interactions and simultaneously obtain features of multiple dimensions to avoid information loss, thereby improving detection accuracy. The GAM structure diagram is shown in the figure. Figure 3 As shown in Figure 2, the channel attention module uses 3D arrangement to retain three-dimensional information, and uses dual MLP to amplify the dependencies between cross-dimensional channels and space, while the spatial attention module uses two convolutions to focus on spatial information.

[0042] When adding the GAM attention mechanism, the Bayesian optimization algorithm is used to determine the optimal hyperparameter configuration of GAM to ensure the convergence and performance of the model. In addition, the attention mechanism is also optimized, and efficient matrix operations and parallel computing techniques are used to ensure that the additional computing overhead is reduced as much as possible while improving the performance of the model. Finally, the strategy of gradually increasing the learning rate is adopted to stabilize the initial training process, and the adaptive learning rate optimizer is used to dynamically adjust the learning rate to ensure the stability and effectiveness of the training process and avoid the problem of slow convergence due to the introduction of the GAM module.

[0043] Compared with other attention mechanisms, the GAM attention mechanism combines channel and spatial attention, while considering the interaction between channels and space, retaining more feature information and improving the performance of the model.

[0044] Example 2.

[0045] This embodiment uses the LGU-Net meteorological element correction system constructed in Example 1 to correct the 2m temperature in the northeast region of the GFS model, and evaluates the accuracy of the correction method. The specific process diagram is shown in FIG. Figure 4 As shown:

[0046] Step 1: Data preprocessing:

[0047] The GFS global numerical weather forecast model developed by the National Center for Environmental Prediction of the United States is used as the data to be revised, and the fifth-generation global reanalysis data product ERA5 provided by the European Center for Medium-Term Weather Forecasts ECMWF is used as the real data in the data set. The parameters of the data to be revised and the real data are the same, with a spatial resolution of 0.25°. 38.61°N~53.51°N, 118.83°E~135.13°E are selected as the test area, and the 2m temperature is used as the research object. The GFS forecast data and ERA5 data are processed into TIF images with latitude and longitude information at hourly intervals. The pixel points of the image represent the temperature of the point in K degrees. The TIF image uses the WGS84 coordinate system.

[0048] In addition, this embodiment also uses Jilin-1 satellite remote sensing data, mainly including terrain data taken by Jilin-1, such as data on various land use types, ground height data, etc. After performing atmospheric correction, geometric correction, image fusion, and image registration preprocessing on the Jilin-1 satellite remote sensing image, remote sensing image data is obtained.

[0049] Step 2: Create a temperature correction data set:

[0050] First, the GFS forecast data was processed to obtain TIF images of the temperature, humidity, wind speed, air pressure and other elements of the experimental area. The ERA5 data was processed to obtain a TIF image of the temperature of the experimental area. The terrain data and land use data of Jilin No. 1 were clipped according to the latitude and longitude range of the experimental area. The above operations completed the data preparation.

[0051] Then, we create multi-temporal and multi-factor input data. This part uses the data processed by GFS. In the time dimension, we use 1 hour as the time interval and 24 hours as a time period. In the multi-factor dimension, we input humidity, wind speed, pressure, and precipitation as auxiliary factors of temperature. Finally, we construct a 5×24×X×Y input data.

[0052] Finally, we create multi-temporal truth data. This part uses ERA5 data. In the time dimension, we use 1 hour as the time interval and 24 hours as a time period. We construct a 24×X×Y truth data, where the time and size of the data correspond to the input data one by one.

[0053] Step 3: LGU-Net network training:

[0054] The LGU-Net meteorological element correction system constructed in Example 1 is used to train the temperature correction data set prepared in step 2 to obtain a temperature correction model for Northeast China based on GFS model prediction data.

[0055] Step 4: Use the trained LGU-Net temperature correction model to correct the hourly temperature data predicted by the GFS model, and compare it with the 24-hour temperature data of the ERA5 reanalysis data as the true value. The root mean square error RMSE and determination coefficient R2 are used as evaluation parameters. The average hourly correction accuracy evaluation results are shown in Table 1.

[0056] Table 1

[0057]

[0058]

[0059] It can be seen from the data in Table 1 that the correction method can achieve 24-hour temperature data correction, and the average RMSE between the corrected data and the true data is 0.7503, and the average R2 is 0.9198. The correction accuracy of the first 20 hours is higher, and the error RMSE is less than 1.

[0060] Step 5. Visualize the ERA5 temperature data and the corrected temperature data in Kelvin. Figure 5 This is the visualization result of the corrected temperature data and the real temperature data.

[0061] The present invention proposes a temperature forecast correction method for Northeast China based on the LGU-Net model, which makes full use of the correlation between various meteorological elements and the continuity characteristics between multi-phase data, improves the accuracy of 2m temperature data prediction of the GFS meteorological forecast model, and provides technical support for refined and accurate meteorological forecasts.

[0062] In order to prove the excellent performance of the correction method described in the present invention, Northeast China was selected as the experimental area. The region has a large dimensional span, with many terrains such as oceans, mountains, plains, and obvious differences in climate. On this basis, the accuracy of the temperature forecast correction method for Northeast China of the LGU-Net model was verified. The back propagation algorithm used for model training is Adam, which requires less memory and has higher computational efficiency. In order to further improve the computational efficiency and the robustness of the model, the network is batch trained and the batch size is set to 32, the initial learning rate (r) of the network is set to 0.001, and the learning rate decay factor is set to 0.5; if the validation set score index does not decrease for two consecutive times, the learning rate is decayed to 0.5×r; the total number of training iterations is 100 times, and the validation set score is calculated once for each iteration; finally, the model with the highest validation set score is selected and migrated to the test set for correction and score calculation. The overall model training process uses a GPU (graphics processing unit) server, the GPU chip is NVIDIA GeForce RTX 3060, and the memory capacity is 12GB.

[0063] The above experimental results show that the LGU-Net meteorological element correction system proposed in the present invention can fully extract the relevant features between multiple elements and multiple time phases, and realize the correction of GFS numerical meteorological forecast model data. The data error after correction is reduced, which can be used as the preferred method for correcting the error of meteorological numerical forecast.

[0064] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The above implementation modes are only used to illustrate the present invention, but not to limit the present invention. Ordinary technicians in the relevant technical field can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A LGU-Net meteorological element correction system, characterized in that: Including U-Net structure, LSTM structure and GAM attention mechanism; The U-Net structure is a symmetrical U-shaped network structure consisting of an encoder and a decoder connected by skip links, which is used to extract the spatial distribution characteristics of meteorological elements; the encoder is composed of convolutional layers and pooling layers, and at each encoder stage, the size of the feature map is gradually reduced and the number of channels is gradually increased to capture meteorological feature information at different levels; the decoder is composed of upsampling layers and convolutional layers, which is used to restore the meteorological feature map extracted in the encoder to the size of the original input image; The LSTM structure is arranged in the jump link part of the U-Net structure, and is used to extract the relevant features between the multi-time series meteorological images, and input the multi-time series meteorological features of the encoder stage into the decoder stage; the LSTM structure adopts the Xavier weight initialization method and the gradient clipping method to stabilize the gradient, adopts the distributed training strategy, and imitates the GRU variant to optimize the LSTM structure; The GAM attention mechanism is added to the upsampling layer of the decoder, including a channel attention module and a spatial attention module; the spatial attention module uses two convolutions to focus on spatial information; the channel attention module uses 3D arrangement to retain three-dimensional information; and dual MLP is used to amplify the dependency between cross-dimensional channels and space.

2. The LGU-Net meteorological element correction system according to claim 1, characterized in that: The specific structure of the LSTM optimized by imitating the GRU variant is as follows: Merge the input gate and forget gate of LSTM; optimize the LSTM structure by sharing some parameters or reducing redundant parameters.

3. The LGU-Net meteorological element correction system according to claim 1, characterized in that: When adding the GAM attention mechanism, the Bayesian optimization algorithm is used to determine the optimal hyperparameter configuration of GAM to ensure the convergence and performance of the model. In addition, the attention mechanism is also optimized, and efficient matrix operations and parallel computing techniques are used to ensure that the additional computing overhead is reduced as much as possible while improving the performance of the model. Finally, the strategy of gradually increasing the learning rate is adopted to stabilize the initial training process, and the adaptive learning rate optimizer is used to dynamically adjust the learning rate to ensure the stability and effectiveness of the training process and avoid the problem of slow convergence due to the introduction of the GAM module.

4. A temperature forecast correction method, characterized in that: The LGU-Net meteorological element correction system according to any one of claims 1 to 3 is applied, wherein the correction method specifically comprises the following steps: S1. Preprocess the input data: S2, construct multi-phase and multi-factor input data as the temperature correction data set; S3, using the LGU-Net meteorological element correction system to train the temperature correction data set to obtain the LGU-Net temperature correction model of the correction area based on the GFS model prediction data; S4. Use the trained LGU-Net temperature correction model to correct the hourly temperature data predicted by the GFS model.

5. The temperature forecast correction method according to claim 4, characterized in that: The step S1 specifically includes: The GFS forecast data of the correction area is used as the data to be corrected, and preprocessed to obtain multi-factor TIF images with latitude and longitude information at hourly intervals; The terrain data of the correction area in the satellite remote sensing data is preprocessed to obtain remote sensing image data.

6. The temperature forecast correction method according to claim 5, characterized in that: The multiple factors include temperature, humidity, wind speed, pressure and precipitation factors.

7. The temperature forecast correction method according to claim 4, characterized in that: The step S2 specifically includes: In the time dimension, 1 hour is used as the time interval and 24 hours as a time period, using the data processed by GFS; In the multi-factor dimension, humidity, wind speed, pressure, and precipitation are all input as auxiliary factors of temperature; In this way, a 5×24×X×Y multi-phase and multi-factor input data set is constructed as the temperature correction data set, where X and Y represent the length and width of the TIF image, respectively.

Citation Information

Cited By

  • Att-TS-UKAN rainfall deviation correction method based on conditional dynamic region sampling

    CN119557578A

  • Rainfall forecast correction method based on U-CS-Net model

    CN119830734A

  • Precipitation forecast correction method based on u-cs-net model

    CN119830734B

  • Meteorological element forecast correction method and system based on CS-UNet model and visualization device

    CN121502353A

  • Sea surface high wind speed reanalysis product correction and resolution improvement method based on bimodal U-Net

    CN121937296A