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

Abstract

The invention discloses a CS-UNet model-based meteorological element forecast correction method and system and a visualization device, relates to the technical field of meteorological forecast element AI correction, and solves the problems of large calculation amount and complex model structure when a traditional numerical mode is used for processing complex meteorological elements. The method comprises the following steps: S1, constructing a meteorological element data set; s2, preprocessing the meteorological element data set, and respectively making multi-temporal multi-element input data and multi-temporal multi-element true value data based on the preprocessed meteorological element data set; step S3, constructing a CS-Unet model; s4, the CS-UNet model is trained by adopting the multi-temporal and multi-element input data; and S5, the trained CS-UNet model corrects the multi-temporal and multi-element input data, and compares the corrected data with the multi-temporal and multi-element true value data, thereby completing the meteorological element forecast correction method.

Description

Technical Field

[0001] This invention relates to the field of AI correction technology for meteorological forecast elements, specifically to a method, system, and visualization device for correcting meteorological forecasts based on the CS-UNet model. Background Technology

[0002] Against the backdrop of globalization and increasingly intense human activity, the impact of meteorological phenomena on key sectors such as agriculture, transportation, and energy has become increasingly prominent, making the demand for accurate weather forecasts particularly urgent. Although numerical weather prediction technology has made significant leaps and is widely used in operational weather forecasting, it still faces many challenges in handling certain complex localized weather events. For example, numerical models are not sensitive enough to subtle changes in terrain, leading to a significant reduction in forecast accuracy in mountainous areas or specific local regions. Furthermore, uncertainties in model initial conditions, simplifications of physical processes, and limitations in model resolution can all contribute to forecast errors. This problem is particularly pronounced under complex terrain or extreme weather conditions.

[0003] Traditional correction methods, while relying on empirical formulas and statistical techniques to provide some improvement, have limitations in capturing the complex relationships between multi-scale and multi-temporal meteorological elements, and often require frequent manual adjustments, resulting in low efficiency. Therefore, developing correction techniques that can fully utilize multi-element and multi-temporal characteristics and possess adaptive learning capabilities has become a key technological breakthrough for improving forecast accuracy and reducing forecast errors. Summary of the Invention

[0004] This invention solves the problems of large computational load and complex model structure in traditional numerical models when dealing with complex meteorological elements.

[0005] The meteorological element forecast correction method based on the CS-UNet model described in this invention includes the following steps: Step S1: Construct a meteorological element dataset; Step S2: Preprocess the meteorological element dataset, and based on the preprocessed meteorological element dataset, generate multi-temporal and multi-element input data and multi-temporal and multi-element ground truth data respectively. Step S3: Construct the CS-Unet model; Step S4: Train the CS-UNet model using multi-temporal and multi-factor input data; Step S5: Correct the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compare it with the true data of multiple time phases and multiple elements to complete the method for correcting meteorological element forecasts.

[0006] Furthermore, in the embodiments of the present invention, the meteorological element dataset includes topographic data, ERA5 meteorological element data, WRF meteorological element data, regional elevation data, and time feature variables.

[0007] Furthermore, in the embodiments of the present invention, the multi-temporal and multi-element input data consists of WRF meteorological element data, topographic data and time feature variables, and in the time dimension, it is divided into 1-hour intervals and 24-hour periods to form a multi-temporal and multi-element input data with a dimension of 24×12×X×Y. Among them, 12 represents 8 WRF input channels and 4 time feature variables, and X and Y represent the spatial dimensions of the multi-temporal and multi-factor input data, respectively.

[0008] Furthermore, in the embodiments of the present invention, the multi-temporal and multi-element true data uses ERA5 meteorological element data, with 1 hour as the time interval and 24 hours as a time period, to form a 24×5×X×Y true data. Here, 5 represents the five target elements: T2m, D2m, U10, V10, and SLP, while X and Y represent the spatial dimensions of the ground truth data for multiple time phases and multiple elements.

[0009] Furthermore, in the embodiments of the present invention, the CS-UNet model includes a U-Net structure, a CAM attention mechanism, a SAM attention mechanism, and a multilayer perception mechanism; CAM attention mechanism, SAM attention mechanism and multilayer perception mechanism are added to the skip connection part of the U-Net structure.

[0010] The present invention discloses a meteorological element forecast visualization device, which is implemented based on any of the above-mentioned meteorological element forecast correction methods based on the CS-UNet model. The device is capable of visualizing ERA5 meteorological element data, WRF meteorological element data, and corrected multi-temporal and multi-element input data.

[0011] The present invention discloses a meteorological element forecast correction system based on the CS-UNet model. The system is implemented based on the aforementioned method for meteorological element forecast correction using the CS-UNet model, and is characterized by comprising the following modules: Module S1, constructs a meteorological element dataset; Module S2 preprocesses the meteorological element dataset and generates multi-temporal, multi-element input data and multi-temporal, multi-element ground truth data based on the preprocessed meteorological element dataset. Module S3, constructs the CS-UNet model; Module S4 uses multi-temporal and multi-factor input data to train the CS-UNet model; Module S5 corrects the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compares it with the true data of multiple time phases and multiple elements, thus completing the method for correcting meteorological element forecasts.

[0012] This invention solves the problems of large computational load and complex model structure in traditional numerical models when dealing with complex meteorological elements. Specific beneficial effects include: This invention presents a meteorological element forecast correction method based on the CS-UNet model, addressing the problem of large errors in traditional meteorological forecast models when processing multi-element data, and solving the challenge that existing correction algorithms cannot effectively utilize the correlation features between multiple elements and multiple time phases. By adopting the U-Net model as the basic architecture, this invention introduces a Channel Attention (CAM) mechanism and a Spatial Attention (SAM) mechanism. The CAM mechanism automatically learns and strengthens the importance of different time points in meteorological element forecasts, while the SAM mechanism improves the model's perception of complex terrain and spatial features by capturing the correlation between spatial locations. Simultaneously, temporal features are introduced to help the model capture the periodic changes of meteorological elements, improving the modeling capability of time-series data; the addition of terrain data further supplements spatial features, enhancing the model's forecasting capability under different terrain conditions. Through these methods, the model can more accurately correct the meteorological elements output by the WRF model, significantly improving forecast accuracy. Attached Figure Description

[0013] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a meteorological element forecast correction method based on the CS-UNet model, as described in Implementation Method 1. Figure 2 This is a structural diagram of the CS-UNet model described in Implementation Method 4; Figure 3 This is a structural diagram of the CAM attention mechanism / SAM attention mechanism described in Implementation Method 4; Figure 4 This is a structural diagram of the multi-layer sensing mechanism described in Implementation Method 4; Figures 5 to 9 This is a comparison chart of the average hourly ERA5, WRF, and corrected results as described in Implementation Method 5. Detailed Implementation

[0014] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0015] Implementation Method 1: The meteorological element forecast correction method based on the CS-UNet model described in this implementation method includes the following steps: Step S1: Construct a meteorological element dataset; Step S2: Preprocess the meteorological element dataset, and based on the preprocessed meteorological element dataset, generate multi-temporal and multi-element input data and multi-temporal and multi-element ground truth data respectively. Step S3: Construct the CS-Unet model; Step S4: Train the CS-UNet model using multi-temporal and multi-factor input data; Step S5: Correct the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compare it with the true data of multiple time phases and multiple elements to complete the method for correcting meteorological element forecasts.

[0016] This embodiment describes a meteorological element forecast correction method based on the CS-UNet model, which mainly includes steps such as dataset construction, correction model training, and correction result verification, and realizes the error correction of meteorological elements in the WRF forecast model. Figure 1 As shown.

[0017] Furthermore, based on the CS-UNet network with a U-Net structure, the meteorological elements predicted by the WRF weather model were corrected.

[0018] To address the issue that most correction algorithms cannot fully utilize the spatial and temporal features of data, channel attention is used to extract the temporal features of the data, spatial attention is used to extract the spatial features of the data, and terrain data is added before spatial attention to improve the accuracy of the correction algorithm.

[0019] Implementation Method 2: This implementation method further defines the meteorological element forecast correction method based on the CS-UNet model described in Implementation Method 1. The meteorological element dataset includes topographic data, ERA5 meteorological element data, WRF meteorological element data, regional elevation data, and time characteristic variables.

[0020] This implementation method involves preparing a meteorological element dataset, specifically as follows: The WRF model, jointly developed by multiple institutions including the National Center for Atmospheric Research (NCAR), the Forecasting Systems Laboratory of the National Oceanic and Atmospheric Administration (NOAA), the National Center for Atmospheric Environment Research (FSL, NCEP / NOAA), and the Center for Heavy Rainfall Analysis and Forecasting at the University of Oklahoma, was used as the data to be corrected. In addition to five target elements—temperature at 2-meter altitude (T2m), dew point temperature at 2-meter altitude (D2m), U and V components (U10, V10) of wind speed at 10-meter altitude, and sea level pressure (SLP)—vegetation cover, latent heat flux, and precipitation were selected as auxiliary elements. This dataset has a spatial resolution of 0.09° and a temporal resolution of hourly. For real data, ERA5, developed by the European Centre for Medium-Range Weather Forecasts (ECMWF), was selected. This dataset provides global reanalysis data for various meteorological elements, with a spatial resolution of 0.25° and a temporal resolution of hourly. The five target meteorological elements provided by the ERA5 data are consistent with those in the WRF model. The study area was selected from Northeast China (38.61°–53.51°N, 118.83°–135.13°E).

[0021] Meanwhile, considering the significant topographic relief in the region, the spatial distribution of meteorological elements may be affected by topography. Therefore, we introduced elevation data of the study area as an auxiliary feature to enhance the model's ability to perceive complex topographic features, especially in the spatial correction of weather forecasts.

[0022] Furthermore, to capture the temporal variations of meteorological elements and enable the model to learn the patterns of these variations over time, four temporal feature variables were added at each time point. These temporal features were processed using sine and cosine encoding, representing the periodic variations of hours and dates within the year, respectively. Specifically, the hourly features and date features within the year for a given time point are encoded using the following formulas: ; ; in, The hour of the current time (0-23). This represents the number of days in a year at the current time (1-365). In this way, the model can effectively capture the patterns of meteorological elements changing over time, improving the modeling capabilities of time-series data.

[0023] Implementation Method 3: This implementation method further defines the meteorological element forecast correction method based on the CS-UNet model described in Implementation Method 1. The multi-temporal and multi-element input data consists of WRF meteorological element data, topographic data, and temporal characteristic variables. In the time dimension, the data is divided into 1-hour intervals and 24-hour intervals, forming a multi-temporal and multi-element input data with a dimension of 24×12×X×Y. Among them, 12 represents 8 WRF input channels and 4 time feature variables, and X and Y represent the spatial dimensions of the multi-temporal and multi-factor input data, respectively.

[0024] In this embodiment, the multi-temporal and multi-element true data uses ERA5 meteorological element data, with 1 hour as the time interval and 24 hours as a time period, to form a 24×5×X×Y true data. Here, 5 represents the five target elements: T2m, D2m, U10, V10, and SLP, while X and Y represent the spatial dimensions of the ground truth data for multiple time phases and multiple elements.

[0025] In this embodiment, before generating multi-temporal, multi-element input data and multi-temporal, multi-element ground truth data, the meteorological element dataset needs to be preprocessed, specifically as follows: Because the ERA5 data and the meteorological data output by the WRF model differ in spatial resolution, bilinear interpolation was first used to adjust the meteorological element data and elevation data provided by ERA5 to the same spatial resolution (0.09°) as the WRF model data, ensuring consistency in data comparison and analysis. Simultaneously, noting the potential for missing values ​​in the data, common meteorological data analysis practices were employed to fill these missing values ​​with 0, avoiding interference from missing values ​​in model training. Furthermore, to prevent certain elements from dominating the model training process and causing bias towards specific elements, Z-score normalization was used to standardize the data, avoiding training bias due to different units of measurement. These operations completed the data preparation.

[0026] Next, multi-temporal, multi-feature input data is generated. This part consists of WRF data, terrain data, and temporal feature data, with 1-hour intervals and 24-hour periods as a time dimension. The final result is an input dataset with dimensions of 24×12×X×Y, where 24 represents hours, 12 represents the 8 WRF input channels and 4 temporal features, and X and Y represent the spatial dimensions of the data.

[0027] Finally, multi-temporal, multi-factor ground truth data is generated. This part uses ERA5 data, and also in the time dimension, it uses 1-hour intervals and 24-hour periods to form a 24×5×X×Y ground truth dataset.

[0028] Implementation Method 4: This implementation method further defines the meteorological element forecast correction method based on the CS-UNet model in Implementation Method 1. The CS-UNet model includes a U-Net structure, a CAM attention mechanism, a SAM attention mechanism, and a multilayer sensing mechanism. CAM attention mechanism, SAM attention mechanism and multilayer perception mechanism are added to the skip connection part of the U-Net structure.

[0029] This implementation method mainly involves constructing the CS-Unet model, specifically: First, the CS-UNet deep learning network is constructed. The structure diagram of the CS-UNet model is shown below. Figure 2 As shown in the figure. This implementation combines multi-element and multi-temporal data features to develop a meteorological element correction algorithm, CS-UNet, which is a joint CAM-SAM-UNet model. The traditional U-Net model consists of an encoder and a decoder. The encoder extracts features by downsampling layer by layer, and the decoder restores resolution by upsampling layer by layer. Features from corresponding layers of the encoder and decoder are fused through skip connections. However, when processing complex multi-temporal meteorological data, the existing U-Net model lacks the ability to distinguish the importance of features from different temporal phases, has limited ability to capture correlations between spatial locations, and is insufficient in its feature representation ability to fully reflect the nonlinear relationships of meteorological elements. Therefore, to address these shortcomings, firstly, in the skip connection part of the U-Net model, channel attention is added to learn the importance of different time points in meteorological element forecasting, spatial attention is added to calculate the correlations between spatial locations to generate attention weights, and the multilayer perceptron is added to enhance the feature representation ability. In addition, elevation data is regarded as a geographical feature and fused with the feature map in the spatial attention mechanism to further enhance the model's spatial feature capture ability. The model structure is mainly divided into four parts: U-Net structure, CAM attention mechanism, SAM attention mechanism, and multilayer perceptron.

[0030] (1) The U-Net model is a deep learning architecture for semantic segmentation. Its structure consists of two parts: an encoder and a decoder, forming a symmetrical U-shaped structure connected by skip connections. In the encoder, the input meteorological data and temporal features are processed through a series of convolutional and max-pooling layers. Each convolutional module contains two 3×3 convolutional layers, gradually reducing the resolution while increasing the number of channels to 256, extracting multi-level features. The decoder performs upsampling through bilinear interpolation to gradually restore the spatial resolution of the feature maps. Through skip connections, the high-resolution feature maps of the corresponding layers in the encoder are combined with the processed feature maps in the decoder. Finally, the output layer generates the prediction result through a 3×3 convolution operation.

[0031] (2) In the channel attention CAM section, such as Figure 3As shown, the time dimension is processed as a channel, with each time point's features treated as an independent channel. First, CAM generates Query, Key, and Value matrices for each time channel through a fully connected layer. The correlation between time channels is obtained by calculating the dot product between the Query and Key, i.e., the importance of a feature at one time point to other time points. Next, normalized attention weights are generated using the Softmax function, and these weights are then used to weight the Value matrix, amplifying important time features and weakening less important ones. Thus, through CAM, the model can dynamically learn the importance of different time points in meteorological element correction forecasts.

[0032] (3) Spatial Attention (SAM) has a similar structure to CAM, but it focuses on the spatial dimension of the input feature map. SAM treats each spatial location (pixel) in the feature map as a Query, Key, and Value, and generates attention weights by calculating the correlation between spatial locations. Similar to channel attention, SAM normalizes the weights using Softmax, amplifies important spatial location features, and enhances the model's ability to capture complex spatial information through these weighted features. Simultaneously, elevation data is expanded to the same resolution as other feature maps through interpolation, aligning with the features output by CAM. Figure 1 Enter SAM.

[0033] (4) Multilayer perceptron In the multilayer perceptron section, such as Figure 4 As shown, the input features are first flattened into a one-dimensional vector and then processed through multiple fully connected layers. After the output of each layer, the MLP performs dropout regularization to reduce the risk of overfitting. To further enhance the fluidity of the features, residual connections are added to the MLP structure. By adding the input features to the output of each layer, residual connections ensure that information propagates effectively within the network, mitigating the gradient vanishing problem. Finally, the features processed by the MLP are combined with the decoder output to generate the final prediction result.

[0034] After the model is built, it needs to be trained, specifically as follows: In this embodiment, the training objective of the model is to minimize the error between the predicted meteorological element data and the actual data, thereby accurately correcting the WRF model's predicted data. Therefore, a difference term, `diff`, is first defined to represent the error between the WRF model's predicted data and the actual ERA5 data. Specifically, the difference term is calculated using the following formula: ; in, This is ERA5 data after spatial interpolation. The difference term reflects the deviation between the predicted and actual data, serving as the basis for the loss function.

[0035] During model training, this implementation uses Mean Absolute Error (MAE) as the loss function to optimize the model. Specifically, the formula for calculating MAE is: ; in, It is the sample size. It is the first The true deviation of a sample No. Prediction bias for each sample.

[0036] Then, the CS-UNet model was used to train the correction dataset created in step two. After adjusting the experimental parameters, a meteorological element correction model based on WRF model prediction data was obtained.

[0037] Finally, the model is evaluated, specifically as follows: The hourly meteorological elements predicted by the WRF model were corrected using a trained CS-UNet correction model, and compared with the ERA5 reanalysis data as the true values. Root mean square error (RMSE), bias, and correlation coefficient (PCC) were used as evaluation parameters, and their calculation formulas are as follows: ; ; ; The average hourly accuracy evaluation results before and after correction are shown in Table 1.

[0038] Table 1. Accuracy evaluation results before and after correction

[0039] By comparing the bias, RMSE, and PCC of each meteorological element before and after correction, it can be seen that the CS-UNet model shows significant improvement in correcting the WRF model's predictions. The corrected bias and RMSE are significantly reduced, indicating that the model effectively reduces prediction bias and improves accuracy. Particularly in temperature and humidity, the corrected results are significantly closer to the actual values, and the PCC value is also improved, reflecting a strengthened linear correlation between the model's predictions and the actual values.

[0040] Therefore, the meteorological element forecast correction method based on the CS-UNet model described in any of the embodiments one to four makes full use of the correlation between various meteorological elements and the continuity between multi-temporal data, improves the accuracy of meteorological element data prediction in the WRF meteorological forecast model, and provides technical support for refined and accurate meteorological forecasts.

[0041] To demonstrate the superior performance of the method used in this implementation, Northeast China was selected as the experimental area. This region has a wide latitudinal range, with diverse terrains including mountains and plains, resulting in significant climate variations. Based on this, the accuracy of the meteorological element forecast correction method of the CS-UNet model was verified. The backpropagation algorithm used for model training was Adam, which requires less memory and has higher computational efficiency. To further improve computational efficiency and model robustness, batch training was performed on the network with a batch size of 1 and an initial learning rate (r) of 0.0001. The total number of training iterations was 100, with a validation set score calculated after each iteration. Finally, the model with the highest validation set score was transferred to the test set for correction and score calculation. The entire model training process utilized a GPU (Graphics Processing Unit) server with an NVIDIA GeForce RTX3090 GPU and 24GB of memory.

[0042] The experiments described above demonstrate that the proposed CS-UNet meteorological element correction algorithm can effectively extract the correlation features between multiple elements and multiple time phases, thus achieving the correction of WRF numerical weather prediction model data. The corrected data exhibits reduced errors and can be considered the preferred method for correcting errors in meteorological numerical forecasts.

[0043] Implementation Method 5: A meteorological element forecast visualization device described in this implementation method is implemented based on a meteorological element forecast correction method based on the CS-UNet model described in any one of Implementation Methods 1-4. The device is capable of visualizing ERA5 meteorological element data, WRF meteorological element data, and corrected multi-temporal and multi-element input data.

[0044] This implementation method is used to visualize the results, specifically as follows: Visualize 24-hour ERA5 meteorological element data, WRF meteorological element data, and corrected data, such as... Figures 5 to 9 As shown in the figure, the CS-UNet correction model effectively corrects the prediction error of the WRF model, making the corrected predictions closer to the ERA5 data, and significantly reducing the difference between the prediction results and the true values.

[0045] Implementation Method Six: This implementation method describes a meteorological element forecast correction system based on the CS-UNet model. The system is implemented based on the meteorological element forecast correction method based on the CS-UNet model described in Implementation Method One, and includes the following modules: Module S1, constructs a meteorological element dataset; Module S2 preprocesses the meteorological element dataset and generates multi-temporal, multi-element input data and multi-temporal, multi-element ground truth data based on the preprocessed meteorological element dataset. Module S3, constructs the CS-UNet model; Module S4 uses multi-temporal and multi-factor input data to train the CS-UNet model; Module S5 corrects the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compares it with the true data of multiple time phases and multiple elements, thus completing the method for correcting meteorological element forecasts.

[0046] The above provides a detailed description of the meteorological element forecast correction method, system, and visualization device based on the CS-UNet model proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for correcting meteorological element forecasts based on the CS-UNet model, characterized in that, Includes the following steps: Step S1: Construct a meteorological element dataset; Step S2: Preprocess the meteorological element dataset, and based on the preprocessed meteorological element dataset, generate multi-temporal and multi-element input data and multi-temporal and multi-element ground truth data respectively. Step S3: Construct the CS-Unet model; Step S4: Train the CS-UNet model using multi-temporal and multi-factor input data; Step S5: Correct the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compare it with the true data of multiple time phases and multiple elements to complete the method for correcting meteorological element forecasts.

2. The meteorological element forecast correction method based on the CS-UNet model according to claim 1, characterized in that, The meteorological element dataset includes topographic data, ERA5 meteorological element data, WRF meteorological element data, regional elevation data, and time characteristic variables.

3. The meteorological element forecast correction method based on the CS-UNet model according to claim 1, characterized in that, The multi-temporal, multi-element input data consists of WRF meteorological element data, topographic data, and temporal characteristic variables. In the time dimension, it is divided into 1-hour intervals and 24-hour periods, forming a multi-temporal, multi-element input data with a dimension of 24×12×X×Y. Among them, 12 represents 8 WRF input channels and 4 time feature variables, and X and Y represent the spatial dimensions of the multi-temporal and multi-factor input data, respectively.

4. The meteorological element forecast correction method based on the CS-UNet model according to claim 1, characterized in that, The multi-temporal, multi-element true data uses ERA5 meteorological element data, with 1-hour intervals and 24-hour periods to form a 24×5×X×Y true data set. Here, 5 represents the five target elements: T2m, D2m, U10, V10, and SLP, while X and Y represent the spatial dimensions of the ground truth data for multiple time phases and multiple elements.

5. The meteorological element forecast correction method based on the CS-UNet model according to claim 1, characterized in that, The CS-UNet model includes a U-Net structure, a CAM attention mechanism, a SAM attention mechanism, and a multilayer perceptron mechanism. CAM attention mechanism, SAM attention mechanism and multilayer perception mechanism are added to the skip connection part of the U-Net structure.

6. A meteorological element forecast visualization device, wherein the device is implemented based on a meteorological element forecast correction method based on the CS-UNet model as described in any one of claims 1-5, and is capable of visualizing ERA5 meteorological element data, WRF meteorological element data, and corrected multi-temporal and multi-element input data.

7. A meteorological element forecast correction system based on the CS-UNet model, wherein the system is implemented based on the meteorological element forecast correction method based on the CS-UNet model as described in claim 1, characterized in that, Includes the following modules: Module S1, constructs a meteorological element dataset; Module S2 preprocesses the meteorological element dataset and generates multi-temporal, multi-element input data and multi-temporal, multi-element ground truth data based on the preprocessed meteorological element dataset. Module S3, constructs the CS-UNet model; Module S4 uses multi-temporal and multi-factor input data to train the CS-UNet model; Module S5 corrects the input data of the trained CS-UNet model for multiple time phases and multiple elements, and compares it with the true data of multiple time phases and multiple elements, thus completing the method for correcting meteorological element forecasts.

Images