Set prediction precipitation information spatiotemporal error correction method and system based on image super-resolution

Abstract

The application provides a set forecast precipitation information space-time error correction method and system based on image super-resolution, and relates to the technical field of precipitation forecast.The application corrects the space-time error of set forecast precipitation information based on the image super-resolution idea and a deep learning algorithm, improves the spatial resolution of set forecast precipitation information, and improves the forecast precision.The method has outstanding advantages in improving the set forecast precipitation space-time structure similarity and the precision of rain or no rain certainty classification forecast compared with the traditional statistical post-processing method, and provides high-precision data support for the application of set forecast precipitation information in flood forecast business.

Description

Technical Field

[0001] This invention relates to the field of precipitation forecasting technology, and in particular to a method and system for correcting spatiotemporal errors in ensemble forecast precipitation information based on image super-resolution. Background Technology

[0002] Precipitation is a key driving data for flood forecasting, and its spatiotemporal accuracy and lead time are important factors affecting the effectiveness of flood forecasts. However, influenced by multi-scale weather systems and topographic relief, precipitation exhibits complex and high-frequency spatiotemporal variability, making it one of the most difficult meteorological variables to forecast accurately. In recent years, the development of atmospheric dynamics and computer science has driven continuous progress in numerical weather prediction technology. Global numerical weather prediction models can provide ensemble forecasts of precipitation information with a long lead time, and coupling them with hydrological models has become an important development direction for improving the accuracy of flood forecasts and extending the effective lead time. However, the computational grid of global numerical weather prediction models is relatively coarse, resulting in low spatial resolution of their ensemble forecast precipitation information, which is mismatched with the spatial scale of flood forecast models. At the same time, the forecast precipitation results are affected by factors such as the initial field, model structure, and perturbation scheme, resulting in problems such as prominent quantitative errors and insufficient ensemble discretization, thus restricting their application in operational flood forecasting.

[0003] Traditional statistical post-processing methods establish forecast-observation statistical relationships for each grid cell to improve the usability of ensemble forecast precipitation information. While these methods can improve forecast accuracy at local locations, they neglect the spatiotemporal dependence and overall nature of precipitation within a given spatial and temporal range, inherently disrupting the spatiotemporal structure of precipitation. Accurate spatiotemporal structure of precipitation is crucial for rationally extrapolating the spatiotemporal evolution of runoff generation and confluence patterns in watersheds. Therefore, developing a novel statistical post-processing method for ensemble forecast precipitation that simultaneously improves spatial resolution, reduces quantitative errors, reconstructs spatiotemporal structure, and enhances probabilistic forecast accuracy is a current research focus. Summary of the Invention

[0004] Purpose of the invention: To propose a spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution, and further to propose a system for implementing the above method. By using a conditional generative deep learning image super-resolution model, the spatial resolution of ensemble forecast precipitation is improved, quantitative error is reduced, spatiotemporal structure is reconstructed, and probabilistic forecast accuracy is increased, so as to solve the above-mentioned problems in the existing technology and provide refined information support for flood forecasting operations.

[0005] In a first aspect, the present invention proposes a method for correcting spatiotemporal errors in ensemble forecast precipitation information based on image super-resolution, comprising the following steps:

[0006] Acquire multi-source data within the defined rectangular boundary of the watershed, including aggregated forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover and DEM data, and perform spatiotemporal scale registration;

[0007] Training and testing samples are divided according to time, effective samples are selected from the training samples and sample augmentation is performed, and data homogenization is carried out on this basis.

[0008] A generator for correcting the spatiotemporal error of ensemble forecast precipitation is constructed. Using a U-shaped network structure combined with the input module G, a false image of ensemble forecast precipitation that approximates the actual spatial distribution of precipitation is generated.

[0009] A discriminator for ensemble forecasting of precipitation spatiotemporal error correction is constructed. The residual network structure is combined with the input module D to determine the probability that the generated precipitation spatial distribution image is a true image.

[0010] The generator and discriminator are connected to form a conditional generative adversarial network. Based on the training samples, the network is trained alternately to obtain a stable ensemble forecast precipitation statistical post-processing model. The test samples are input into the ensemble forecast precipitation statistical post-processing model to generate statistically post-processed ensemble forecast precipitation results.

[0011] In a further embodiment of the first aspect, the process of acquiring multi-source element data within the delineated rectangular boundary of the watershed specifically includes:

[0012] Obtain raw ensemble forecast precipitation information and actual observed precipitation information within the defined rectangular boundary of the watershed, and determine the temporal resolution, spatial resolution, start time, members, and lead time information of the raw ensemble forecast precipitation information, as well as the temporal resolution and spatial resolution of the actual observed precipitation information;

[0013] Acquire 2m air temperature, total cloud cover and DEM data within the defined rectangular boundary of the watershed, and determine the spatial resolution of the three.

[0014] Using the spatial resolution of actual observed precipitation information as the target resolution, bilinear interpolation was used to downscale the original ensemble forecast precipitation information, 2m temperature, total cloud cover and DEM data to register the spatial resolution.

[0015] Using the actual observed precipitation time resolution as the target resolution, linear interpolation is used to downscale the ensemble forecast precipitation information time resolution. Then, the mean of downscaled ensemble forecast precipitation information with the same forecast period at different start times is combined, and time registration is performed according to time, actual observed precipitation information, 2m temperature, and total cloud cover.

[0016] In a further embodiment of the first aspect, effective samples are selected from the training samples and sample augmentation is performed, with the following selection rules:

[0017] In the original ensemble forecast precipitation information, the proportion of grid cells with forecast precipitation greater than 0.1 mm per unit time scale exceeds 50%; in the original ensemble forecast precipitation information for at least 50% of the forecast period, the proportion of grid cells with forecast precipitation greater than 2 mm exceeds 50%; in the original ensemble forecast precipitation information for at least 50% of the forecast period, the proportion of grid cells with forecast precipitation greater than 5 mm exceeds 50%; and in the original ensemble forecast precipitation information for at least 25% of the forecast period, the proportion of grid cells with forecast precipitation greater than 10 mm exceeds 50%.

[0018] The effective training samples are augmented by rotating them counterclockwise by 90°, 180° and 270° respectively, thereby expanding the size of the effective samples.

[0019] In a further embodiment of the first aspect, the process of performing data homogenization specifically includes:

[0020] For both training and test samples, the mean of the forecast precipitation information of the set is subjected to logarithmic transformation and normalization operations that retain zero values, while only normalization operations are performed on the 2m temperature, total cloud cover and DEM data.

[0021] In a further embodiment of the first aspect, the input module G1 of the generator is a Gaussian random variable. , Input module G2 is a condition variable , ;

[0022] in, This represents a normal distribution with a mean of 0 and a variance of 1. For digital elevation; This represents the mean of the downscaled ensemble forecast precipitation. and These are the downscaled 2m air temperature and total cloud cover, respectively.

[0023] In a further embodiment of the first aspect, the process of constructing a generator for ensemble forecast precipitation spatiotemporal error correction specifically includes:

[0024] The encoder of the generator is constructed. It performs feature downsampling on the input tensors of input modules G1 and G2. It consists of several downsampling modules connected together. Each downsampling module comprises two 2D convolutional layers, a max-pooling layer, a batch normalization layer, and a dropout layer, as shown in the following expression:

[0025]

[0026] In the formula, For the first Output features of each downsampling block; For discarding layers; For batch normalization layer; This is a max pooling layer; and The first The weights of the first and second convolutional layers in each downsampling block; For the first The input of each downsampling block; This is a convolution operation;

[0027] The decoder is constructed to build the generator. The decoder upsamples the input features and consists of several interconnected upsampling modules, ensuring that the number of upsampling modules exceeds one downsampling module. These upsampling modules serve as the connection between the encoder and decoder. Each upsampling module consists of two 2D convolutional layers, a transposed convolutional layer, a batch normalization layer, and a dropout layer, as shown in the following expression:

[0028]

[0029] In the formula, For the first Each upsampling module outputs features; and The first The weights of the first and second convolutional layers in each upsampling block; For the first Output features of each downsampling layer; For the first The output features of each upsampling layer; This is a transpose convolution operation; This is a feature splicing operation.

[0030] In a further embodiment of the first aspect, the input module G1 and the input module G2 are connected to the encoder, and then connected to the connection module and the decoder in sequence to construct a U-shaped network. The output features of the downsampling module in the encoder are connected one by one to the two-dimensional convolutional layer of the upsampling module corresponding to the U-shaped network to form a complete generator for spatiotemporal error correction of ensemble forecast precipitation.

[0031] In a further embodiment of the first aspect, the process of constructing a discriminant for ensemble forecast precipitation spatiotemporal error correction specifically includes:

[0032] The input module D1 of the discriminator is the output feature of the generator. Input module D2 is a condition , , This indicates actual observed precipitation;

[0033] A symmetrical residual block stack structure consisting of two paths is constructed. Each residual block is composed of a two-dimensional convolutional layer and two sets of leaky activation function layers connected to two-dimensional convolutional layers, as shown in the following expression:

[0034]

[0035]

[0036] In the formula, and The first and second pathways are respectively Output characteristics of each residual block; and These are the input features of the first and second pathways, respectively. and The first and second pathways are respectively The weights of the first convolutional layer in each residual block; , and , The first and second pathways are respectively the first and second pathways. The weights of the second and third convolutional layers in each residual block; This is to leak the activation layer function;

[0037] Construct the probability output module in the discriminator. The probability output module is constructed by concatenating... and The probability that the output of the leaky activation function layer, the 2D convolutional layer, the global average pooling layer, the fully connected layer, and the activation function layer is true is expressed as follows:

[0038]

[0039] In the formula, Output probability; It is a fully connected layer; This is a global average pooling layer; These are the weights of the two-dimensional convolutional layer in the probability output module;

[0040] Input modules D1 and D2 are sequentially connected to a symmetrical residual block stacking structure and a probability output module to form a complete discriminator for ensemble forecast precipitation spatiotemporal error correction.

[0041] In a further embodiment of the first aspect, a conditional generative adversarial network is constructed by connecting a generator and a discriminator, wherein the generator receives random variables. With condition variables It outputs ensemble forecast precipitation false images, and the discriminator receives the ensemble forecast precipitation false images and condition variables. Output probability;

[0042] Define the loss function of the discriminator respectively. Loss function of generator :

[0043]

[0044] in, ;

[0045]

[0046] in,

[0047]

[0048] In the formula, and These are the discriminator and the generator, respectively. and These are the parameters for the discriminator and the generator, respectively. The weights are used for gradient penalty. For actual precipitation observation; and The mean probability of generating true or false images and The penalty factor; This represents the function for calculating the mean square error; Fairness skill score calculation function;

[0049] A phased training strategy is adopted. First, the generator is pre-trained based on the training samples using MSE as the loss function, and then... The generator is fixed for the loss function, the discriminator is trained, and finally... The loss function is fixed for the discriminator, and the generator is trained to obtain a stable ensemble forecast precipitation statistical post-processing model.

[0050] The test samples are placed into the trained stable ensemble forecast precipitation statistical post-processing model, and the output is the statistically post-processed ensemble forecast precipitation data.

[0051] A second aspect of the present invention proposes a spatiotemporal error correction system for ensemble forecast precipitation information based on image super-resolution. This system automatically executes the spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution disclosed in the first aspect and its further embodiments. The system includes:

[0052] The data pre-extraction module is used to acquire multi-source element data within the defined rectangular boundary of the watershed, including aggregated forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover and DEM data, and to perform spatiotemporal scale registration.

[0053] The sample preprocessing module is used to divide training samples and test samples according to time, filter effective samples in the training samples and perform sample augmentation, and carry out data homogenization processing on this basis.

[0054] Generator; The generator uses a U-shaped network structure combined with the input module G to generate ensemble forecast precipitation false images that approximate the actual spatial distribution of precipitation;

[0055] Discriminator; the discriminator uses a residual network structure combined with input module D to determine the probability that the generated precipitation spatial distribution image is a real image;

[0056] The post-processing module is formed by connecting the generator and the discriminator to create an adversarial network, and is obtained by alternating training based on the training samples.

[0057] The error correction result output module is used to input the test samples into the post-processing module to generate statistically post-processed ensemble forecast precipitation results.

[0058] Beneficial Effects: This invention corrects the spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution and deep learning algorithms. This not only improves the spatial resolution of the ensemble forecast precipitation information but also enhances forecast accuracy. Compared to traditional statistical post-processing methods, this approach has significant advantages in improving the spatiotemporal structural similarity of ensemble forecast precipitation and the accuracy of deterministic classification forecasts of rain or no rain, providing high-precision data support for the application of ensemble forecast precipitation information in flood forecasting operations. Attached Figure Description

[0059] Figure 1 This is an overall flowchart of a method for correcting spatiotemporal errors in ensemble forecast precipitation information based on image super-resolution.

[0060] Figure 2 This is a flowchart detailing the spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution.

[0061] Figure 3 This is a schematic diagram of the spatiotemporal error correction system for ensemble forecast precipitation information based on image super-resolution.

[0062] Figure 4 The ensemble forecasts of precipitation are output for different post-processing methods for a certain period.

[0063] Figure 5 Output the spatiotemporal distribution of ETS for ensemble forecast precipitation for different statistical post-processing methods.

[0064] Figure 6 Output the KGE' spatiotemporal distribution of ensemble forecast precipitation for different statistical post-processing methods.

[0065] Figure 7 Output the SSIM spatiotemporal distribution of ensemble forecast precipitation for different statistical post-processing methods. Detailed Implementation

[0066] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0067] Example 1:

[0068] This embodiment discloses a spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution. It has significant advantages in improving the spatiotemporal structure similarity of ensemble forecast precipitation and the accuracy of deterministic classification forecasts of rain or no rain, providing high-precision data support for the application of ensemble forecast precipitation information in flood forecasting operations. The method flowchart is as follows: Figure 1 As shown, the specific steps are as follows:

[0069] Step 1: Obtain multi-source element data within the defined rectangular boundary of the watershed, including aggregated forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover and DEM data, and perform spatiotemporal scale registration;

[0070] Step 2: Divide the training samples and test samples according to time, select the effective samples in the training samples and perform sample augmentation, and then carry out data homogenization processing on this basis.

[0071] Step 3: Construct a generator for spatiotemporal error correction of ensemble forecast precipitation. Using a U-shaped network structure combined with the input module, generate ensemble forecast precipitation false images that approximate the actual spatial distribution of precipitation.

[0072] Step 4: Construct a discriminator for ensemble forecast precipitation spatiotemporal error correction, and use a residual network structure combined with the input module to determine the probability that the generated precipitation spatial distribution image is a true image;

[0073] Step 5: Connect the generator and discriminator to form a conditional generative adversarial network. Train alternately based on training samples to obtain a stable ensemble forecast precipitation statistical post-processing model. Input the test samples to generate the statistically post-processed ensemble forecast precipitation.

[0074] Example 2:

[0075] Building upon Example 1, this example discloses a refined process for a spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution, combined with... Figure 1 and Figure 2 It should be noted that the detailed process disclosed in this embodiment is only a feasible solution for correcting the spatiotemporal error of ensemble forecast precipitation information based on image super-resolution. In actual application, it can be reduced or supplemented according to objective factors such as environment and equipment.

[0076] Step 1: Obtain multi-source element data within the defined rectangular boundary of the watershed, including aggregated forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover and DEM data, and perform spatiotemporal scale registration;

[0077] Step 11: Obtain the original ensemble forecast precipitation information and actual observed precipitation information within the defined rectangular boundary of the watershed, and determine the spatiotemporal resolution, start time, members, and lead time information of the original ensemble forecast precipitation information, as well as the spatiotemporal resolution of the actual observed precipitation information;

[0078] Step 12: Obtain 2m air temperature, total cloud cover, and DEM data within the defined rectangular boundary of the watershed, and determine the spatial resolution of the three.

[0079] Step 13: Using the spatial resolution of the actual observed precipitation as the target resolution, use bilinear interpolation to downscale the original ensemble forecast precipitation information, 2m temperature, total cloud cover and DEM data, and register the spatial resolution.

[0080] Step 14: Using the actual observed precipitation time resolution as the target resolution, linear interpolation is used to downscale the ensemble forecast precipitation information time resolution. Then, the downscaled ensemble forecast precipitation information with the same forecast period at different start times is combined and its mean is calculated. Time registration is performed according to time, actual observed precipitation information, 2m temperature, and total cloud cover.

[0081] Step 2: Divide the training samples and test samples according to time, select the effective samples in the training samples and perform sample augmentation, and then carry out data homogenization processing on this basis.

[0082] Step 21: According to the predetermined time range, the first 60% of the spatiotemporally registered data is used as training samples, and the last 40% is used as test samples.

[0083] Step 22: Expand the effective training samples by rotating them counterclockwise by 90°, 180° and 270° respectively.

[0084] Step 23: For the training set, perform logarithmic transformation and normalization on the precipitation data while retaining zero values. Perform normalization only on the 2m air temperature, total cloud cover and DEM. Use the same parameters to transform the data for the test samples.

[0085] Step 3: Construct a generator for spatiotemporal error correction of ensemble forecast precipitation. Using a U-shaped network structure combined with the input module G, generate a false image of ensemble forecast precipitation that approximates the actual spatial distribution of precipitation.

[0086] Step 31: The input module G1 of the generator is a Gaussian random variable. Input module G2 is a condition variable ;

[0087] Step 32: Construct the encoder of the generator. The encoder is used to downsample the input tensors of input module G1 and input module G2. It is composed of several downsampling modules connected together. Each downsampling module consists of two two-dimensional convolutional layers, a max pooling layer, a batch normalization layer and a dropout layer.

[0088] Step 33: Construct the decoder of the generator. The decoder is used to upsample the input features. It consists of several upsampling modules connected together, ensuring that the number of upsampling modules exceeds one downsampling module. It serves as the connecting module between the encoder and the decoder. Each upsampling module consists of two two-dimensional convolutional layers, a transposed convolutional layer, a batch normalization layer, and a dropout layer.

[0089] Step 34: Connect input module G1 and input module G2 to the encoder, and connect them sequentially to the connection module and decoder to construct a U-shaped network. Connect the output features of the downsampling module in the encoder one by one to the two-dimensional convolutional layer of the upsampling module corresponding to the U-shaped network to form a complete generator for spatiotemporal error correction of ensemble forecast precipitation.

[0090] Step 4: Construct a discriminator for ensemble forecast precipitation spatiotemporal error correction, and use a residual network structure in conjunction with input module D to determine the probability that the generated precipitation spatial distribution image is a true image;

[0091] Step 41: The input module D1 of the discriminator is the generator output feature. Input module D2 is a condition ;

[0092] Step 42: Construct a symmetrical residual block stack structure consisting of two paths. Each residual block is composed of a two-dimensional convolutional layer and two sets of leaky activation function layers connected to two-dimensional convolutional layers.

[0093] Step 43: Construct the probability output module in the discriminator. The probability output module is constructed by concatenating... and Then, the probability of generating a true image is output by connecting a leaky activation function layer, a two-dimensional convolutional layer, a global average pooling layer, a fully connected layer, and an activation function layer.

[0094] Step 44: Connect input module D1 and input module D2 in sequence with the symmetrical residual block stacking structure and the probability output module to form a complete discriminator for ensemble forecast precipitation spatiotemporal error correction.

[0095] Step 5: Connect the generator and discriminator to form a conditional generative adversarial network. Train alternately based on training samples to obtain a stable ensemble forecast precipitation statistical post-processing model. Input the test samples to generate the statistically post-processed ensemble forecast precipitation.

[0096] Step 51: Connect the generator and discriminator to construct a conditional generative adversarial network, where the generator receives random variables. With condition variables It outputs ensemble forecast precipitation false images, and the discriminator receives the ensemble forecast precipitation false images and condition variables. Output probability;

[0097] Step 52: Design the loss function, and define the loss functions for the discriminator and generator respectively. and ;

[0098] Step 53: Adopt a phased training strategy. First, pre-train the generator based on the training samples using MSE as the loss function, and then... The generator is fixed for the loss function, the discriminator is trained, and finally... The loss function is fixed for the discriminator, and the generator is trained to obtain a stable ensemble forecast precipitation statistical post-processing model.

[0099] Step 54: Place the test samples into the trained stable ensemble forecast precipitation statistical post-processing model and output the statistically post-processed ensemble forecast precipitation data.

[0100] Example 3:

[0101] See Figure 3 This embodiment discloses a spatiotemporal error correction system 600 for ensemble forecast precipitation information based on image super-resolution. The system 600 comprises a data pre-extraction module 601, a sample preprocessing module 602, a generator 603, a discriminator 604, a post-processing module 605, and an error correction result output module 606. The data pre-extraction module 601 acquires multi-source element data within the defined rectangular boundary of the watershed, including ensemble forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover, and DEM data, and performs spatiotemporal scale registration. The sample preprocessing module 602 is used to divide training samples and test samples according to time, select effective samples from the training samples and perform sample augmentation, and then perform data homogenization processing. Generator 603 uses a U-shaped network structure combined with the input module to generate a false image of ensemble forecast precipitation that approximates the actual spatial distribution of precipitation; discriminator 604 uses a residual network structure combined with the input module to determine the probability that the generated spatial distribution image of precipitation is a real image; a conditional generative adversarial network is formed by connecting generator 603 and discriminator 604, and post-processing module 605 is obtained by alternating training based on training samples; error correction result output module 606 is used to input test samples into post-processing module 605 to generate statistically post-processed ensemble forecast precipitation results.

[0102] Example 4:

[0103] The Han River basin lies between 106.25° and 114.33° east longitude and 30.17° and 34.33° north latitude, with a drainage area of ​​approximately 159,000 km². 2 The Han River's main stream is 1577 km long, with the Danjiangkou Reservoir as the boundary between its upper and middle reaches, and Zhongxiang as the boundary between its middle and lower reaches. The basin's terrain slopes from west to east, with the Qinling-Bashan Mountains in the west, the Nanxiang Basin in the middle (elevation 100-300 m), and the Jianghan Plain in the east. The basin belongs to the East Asian subtropical monsoon climate zone, characterized by a warm and humid climate, with an average annual temperature of 12℃-16℃ and an average annual precipitation of 894 mm. Precipitation is unevenly distributed throughout the year, with the maximum continuous four months of rainfall accounting for approximately 55-65% of the annual precipitation. The summer flood season is from late June to late July, and the autumn flood season is from late August to mid-October. Overall, the basin has relatively abundant precipitation resources. This embodiment constructs a spatiotemporal error correction scenario for ensemble forecast precipitation information based on image super-resolution in the Han River basin. Specifically:

[0104] S1): Acquisition of multi-source data and sample segmentation. Global medium-range ensemble forecast precipitation data from the ECMWF Weather Forecast Center in the TIGGE database for the period from May to November 2018 to 2022 were acquired. The ensemble consisted of 50 members, with a spatial resolution of 0.50°, a temporal resolution of 6 hours, and forecast times of 0:00 and 12:00 (UTC) daily. Actual observed precipitation information was also acquired, with a spatial resolution of 0.10° and a temporal resolution of 3 hours. 2m air temperature and total cloud cover data were also acquired, with a spatial resolution of 0.25° and a temporal resolution of 3 hours. DEM data were also acquired, with a spatial resolution of 0.1°.

[0105] With a target resolution of 0.1° and 3h, bilinear interpolation is used spatially and linear interpolation is used temporally to unify the resolution of ensemble forecast precipitation, 2m temperature, total cloud cover and DEM data. Then, downscaled ensemble forecast precipitation information with the same forecast period at different reporting times is combined and its mean is calculated. Time registration is performed according to time and actual observed precipitation information, 2m temperature and total cloud cover.

[0106] The first 60% of the registered data is used as training samples and the last 40% is used as test samples. The effective training samples are further screened, and the effective training samples are augmented by rotating them counterclockwise by 90°, 180° and 270° respectively to expand the scale of the effective samples.

[0107] For the training set, a logarithmic transformation and normalization operation were performed on the precipitation data to retain zero values, while only the 2m air temperature, total cloud cover, and DEM were normalized. The same parameters were used to transform the data for the test samples.

[0108] S2): Construct a generator for correcting spatiotemporal errors in ensemble forecast precipitation. The generator consists of a U-shaped encoder and a decoder. The encoder consists of five downsampling blocks, each consisting of two 2D convolutional layers (3×3 kernel, zero-padded), a max pooling layer (2×2), a batch normalization layer, and a dropout layer (30% dropout). The decoder consists of six upsampling blocks, each consisting of two 2D convolutional layers (3×3 kernel, zero-padded), a transposed convolutional layer (2×2 kernel, 2×2 stride), a batch normalization layer, and a dropout layer (30% dropout). One of the upsampling blocks is used as the connection module. Furthermore, the dropout layer of the downsampling block in the encoder is concatenated with the first 2D convolutional layer of the downsampling block.

[0109] The inputs of the generator are respectively including The condition variable c1 and the Gaussian random variable z are included, where the Gaussian random variable z is connected to the beginning of the encoder, and the condition variable c1 is connected to the first two-dimensional convolutional layer in the first downsampling block. The formula is as follows:

[0110]

[0111]

[0112] In the formula: For the first Output features of each downsampling block; For the first Each upsampling module outputs features; For discarding layers; For batch normalization layer; This is a max pooling layer; and The first The weights of the first and second convolutional layers in each downsampling block; and The first The weights of the first and second convolutional layers in each upsampling block; For the first The input of each downsampling block; For the first Output features of each downsampling layer; For the first The output features of each upsampling layer; This is a convolution operation; This is a transpose convolution operation; This is a feature splicing operation.

[0113] S3): Construct a discriminator for correcting spatiotemporal errors in ensemble forecast precipitation. The discriminator consists of a symmetrical residual block stack structure with two paths and a probability output module. Each path consists of 3 residual blocks, and each residual block contains two paths. One path connects to a 2D convolutional layer (kernel size 1×1), and the other path connects two sets of leaky activation function layers (scale factor 0.2) to a 2D convolutional layer (kernel size 3×3, with zero padding). Finally, the features of the two paths are concatenated. The probability output module consists of a leaky activation function layer (scale factor 0.2), a 2D convolutional layer (kernel size 3×3, with zero padding), global average pooling, a fully connected layer (output length 32), a fully connected layer (output length 1), and an activation function layer connected together.

[0114] The discriminator's inputs are the generator's output set of predicted precipitation information (false image) and the condition variable c2 of actual observed precipitation. These two inputs are connected to the ends of the two paths in the symmetrical residual block stack structure, and then to the probability output module, as shown in the following formula:

[0115]

[0116]

[0117]

[0118] In the formula: and The first and second pathways are respectively Output characteristics of each residual block; and These are the input features of the first and second pathways, respectively. and The first and second pathways are respectively The weights of the first convolutional layer in each residual block; , and , The first and second pathways are respectively the first and second pathways. The weights of the second and third convolutional layers in each residual block; This is to leak the activation layer function; Output probability; It is a fully connected layer; This is a global average pooling layer; These are the weights of the two-dimensional convolutional layer in the probability output module.

[0119] S4): Connect the generator and discriminator to construct a conditional generative adversarial network, where the generator receives random variables. With condition variables It outputs ensemble forecast precipitation false images, and the discriminator receives the ensemble forecast precipitation false images and condition variables. The output probability is determined. A phased training strategy is adopted. First, the generator is pre-trained based on the training samples using MSE as the loss function. Then, the output probability is determined by... To fix the loss function, the generator is fixed, the discriminator is trained, and finally... The discriminator is fixed by the loss function, and the generator is trained to obtain a stable ensemble forecast precipitation statistical post-processing model. The loss function formula is as follows:

[0120]

[0121]

[0122] in:

[0123]

[0124] in:

[0125]

[0126] In the formula: and These are the discriminator and the generator, respectively. and These are the parameters for the discriminator and the generator, respectively. The weights are used for gradient penalty. For actual precipitation observation; and The mean probability of generating true or false images and The penalty factor; This represents the function for calculating the mean square error; Fairness skill score calculation function.

[0127] The test samples are placed into the trained stable ensemble forecast precipitation statistical post-processing model, and the output is the statistically post-processed ensemble forecast precipitation data.

[0128] S5): Accuracy Evaluation. For the accuracy evaluation of statistical post-processing results, assessments can be conducted using three indicators: ETS, KGE', and SSIM, focusing on classification accuracy, quantitative accuracy, and spatiotemporal structural similarity accuracy. The formulas are as follows:

[0129]

[0130]

[0131]

[0132] In the formula: This indicates the number of events in the forecast that resulted in precipitation exceeding a certain threshold; This indicates the contribution of random forecasts included in the number of hit events. N represents the total number of forecasts; M represents the number of forecasts that missed reporting of precipitation exceeding a certain threshold. This indicates the number of events in the forecast that falsely reported precipitation exceeding a certain threshold; This represents the correlation coefficient between observed and forecasted precipitation; and These represent the average values ​​of the forecast and observed precipitation, respectively. and These represent the standard deviations of forecasted and observed precipitation, respectively. and This represents the average of the forecasted and observed precipitation within the field of view. and The standard deviation of the predicted and observed precipitation within the field of view; Evaluate the covariance of observed and forecasted precipitation within the field of view; , Using a constant value avoids the problem of a denominator of 0 caused by equal precipitation within the field of view. and The default values ​​are 0.01 and 0.03 respectively. The rainfall was extremely poor.

[0133] Figure 4 The following data are presented: the actual observed precipitation benchmark at 12:00 UTC on August 31, 2021; the downscaled ECMWF-d; the traditional statistical post-processing method BGMG; and the ensemble forecast precipitation mean with a lead time of 0–1 day output by the method of this invention, EP-cGAN. As shown in the figure, in terms of capturing the dynamic spatial distribution of rainy and dry precipitation, ECMWF-d suffers from a large-scale precipitation false alarm problem. BGMG and EP-cGAN reduce the false alarm area of ​​precipitation in the middle and lower reaches of the river, and EP-cGAN performs best overall in correctly identifying precipitation areas.

[0134] Figure 5 The spatiotemporal distribution patterns of precipitation ETS indices from the output ensembles of different statistical post-processing methods are presented. For deterministic classification forecasts of rain or no rain (…),… Figure 5 In the middle (a) range, the ETS0.1 of ECMWF-d was generally below 0.3. BGMG and EP-cGAN improved the ETS0.1 for the five forecast periods of 3h, 27h, 51h, 27h, and 99h. The improvement in ETS0.1 was particularly significant in the middle and upper reaches of the river. EP-cGAN showed a better improvement effect than BGMG, with some grid cells achieving an ETS0.1 exceeding 0.4. For deterministic classification forecasts of heavy precipitation ( Figure 5In (b), the spatial distribution patterns of ETS95% of BGMG and ECMWF-d are highly similar and the index values ​​are close. However, the ETS95% of EP-cGAN is similar to the other two methods in the 3-hour forecast period, but the rate of decline is faster as the forecast period is extended, and the low value area shifts from the middle and lower reaches of the 27-hour forecast period to the upper reaches of the 99-hour forecast period.

[0135] Figure 6 The spatiotemporal distribution of KGE' for precipitation forecasts from the output ensembles of different statistical post-processing methods is presented. ECMWF-d shows a spatial pattern of significantly lower KGE' in the northern and middle reaches of the basin and the lower reaches, and higher KGE' in the upper reaches; BGMG significantly reduced the low-value area in the northern middle reaches, increased the KGE' in the lower reaches, and also significantly improved the index value in the upper reaches, with some grids exceeding 0.6; EP-cGAN basically eliminated the two low-value areas present in ECMWF-d and improved the KGE' for the entire basin.

[0136] Figure 7 The spatiotemporal distribution of SSIM (Spatial Similarity Scale) for ensemble forecast precipitation outputs using different statistical post-processing methods is presented. All three methods show a higher SSIM in the middle and lower reaches than in the upper reaches. BGMG slightly improves the SSIM in the middle and lower reaches during the 3-hour lead time of ECMWF-d, but the SSIM across the entire basin rapidly declines after 75 hours, generally falling below 0.5, with significant deterioration in some areas of the middle and upper reaches compared to ECMWF-d. EP-cGAN significantly improves the SSIM compared to BGMG and ECMWF-d across all lead times. During the 3-hour lead time, the SSIM of most grids approaches 0.8, and even after extending the lead time to 99 hours, the SSIM in the middle and lower reaches remains close to or above 0.6. Overall, the EP-cGAN model demonstrates a stronger ability to improve the spatiotemporal structural similarity of the ensemble forecast precipitation mean than BGMG, which is closely related to the fact that the former uses image patches rather than grids as the analysis unit. In conclusion, the method presented in this invention improves the spatiotemporal resolution of ensemble forecast precipitation information compared to traditional statistical post-processing methods, and has advantages in classification, quantification, and structural similarity accuracy.

[0137] As described above, although the invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the invention itself. Various changes in form and detail may be made without departing from the spirit and scope of the invention as defined in the appended claims.

Claims

1. A method for correcting spatiotemporal errors in ensemble forecast precipitation information based on image super-resolution, characterized in that, Includes the following steps: Obtain raw ensemble forecast precipitation information and actual observed precipitation information within the defined rectangular boundary of the watershed, and determine the temporal resolution, spatial resolution, start time, members, and lead time information of the raw ensemble forecast precipitation information, as well as the temporal resolution and spatial resolution of the actual observed precipitation information; Acquire 2m air temperature, total cloud cover and DEM data within the defined rectangular boundary of the watershed, and determine the spatial resolution of the three. Using the spatial resolution of actual observed precipitation information as the target resolution, bilinear interpolation was used to downscale the original ensemble forecast precipitation information, 2m temperature, total cloud cover and DEM data to register the spatial resolution. Using the temporal resolution of actual observed precipitation as the target resolution, linear interpolation is used to downscale the temporal resolution of ensemble forecast precipitation information. Then, the mean of downscaled ensemble forecast precipitation information with the same forecast period at different start times is combined, and time registration is performed according to time, actual observed precipitation information, 2m air temperature, and total cloud cover. Training samples and test samples are divided according to time, effective samples in the training samples are selected and sample augmentation is performed, and data homogenization processing is carried out on this basis. A generator for correcting the spatiotemporal error of ensemble forecast precipitation is constructed. Using a U-shaped network structure combined with the input module G, a false image of ensemble forecast precipitation that approximates the actual spatial distribution of precipitation is generated. A discriminator for ensemble forecasting of precipitation spatiotemporal error correction is constructed. The residual network structure is combined with the input module D to determine the probability that the generated precipitation spatial distribution image is a true image. The generator and discriminator are connected to form a conditional generative adversarial network. Based on the training samples, the network is trained alternately to obtain a stable ensemble forecast precipitation statistical post-processing model. The test samples are input into the ensemble forecast precipitation statistical post-processing model to generate statistically post-processed ensemble forecast precipitation results.

2. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 1, characterized in that, The selection of effective samples from the training samples and the subsequent sample augmentation are based on the following rules: In the original ensemble forecast precipitation information, the proportion of grid cells with forecast precipitation greater than 0.1 mm per unit time scale exceeds 50%; in the original ensemble forecast precipitation information for at least 50% of the forecast period, the proportion of grid cells with forecast precipitation greater than 2 mm exceeds 50%; in the original ensemble forecast precipitation information for at least 50% of the forecast period, the proportion of grid cells with forecast precipitation greater than 5 mm exceeds 50%; and in the original ensemble forecast precipitation information for at least 25% of the forecast period, the proportion of grid cells with forecast precipitation greater than 10 mm exceeds 50%. The effective training samples are augmented by rotating them counterclockwise by 90°, 180° and 270° respectively, thereby expanding the size of the effective samples.

3. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 1, characterized in that, The aforementioned data homogenization process specifically includes: For both training and test samples, the mean of the forecast precipitation information of the set is subjected to logarithmic transformation and normalization operations that retain zero values, while only normalization operations are performed on the 2m temperature, total cloud cover and DEM data.

4. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 1, characterized in that, The input module G1 of the generator is a Gaussian random variable. , Input module G2 is a condition variable , ; in, This represents a normal distribution with a mean of 0 and a variance of 1. For digital elevation; This represents the mean of the downscaled ensemble forecast precipitation. and These are the downscaled 2m air temperature and total cloud cover, respectively.

5. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 4, characterized in that, The generator for constructing spatiotemporal error correction of ensemble precipitation forecasts specifically includes: The encoder of the generator is constructed. It performs feature downsampling on the input tensors of input modules G1 and G2. It consists of several downsampling modules connected together. Each downsampling module comprises two 2D convolutional layers, a max-pooling layer, a batch normalization layer, and a dropout layer, as shown in the following expression: ； In the formula, For the first The output characteristics of each downsampling module; For discarding layers; For batch normalization layer; This is a max pooling layer; and The first The weights of the first and second convolutional layers in each downsampling module; This is the input to the i-th downsampling module; This is a convolution operation; The decoder is constructed to build the generator. The decoder upsamples the input features and consists of several interconnected upsampling modules, ensuring that the number of upsampling modules exceeds one downsampling module. These upsampling modules serve as the connection between the encoder and decoder. Each upsampling module consists of two 2D convolutional layers, a transposed convolutional layer, a batch normalization layer, and a dropout layer, as shown in the following expression: ； In the formula, For the first Each upsampling module outputs features; and The first The weights of the first and second convolutional layers in each upsampling module; For the first The output characteristics of each downsampling module; For the first The output characteristics of each upsampling module; This is a transpose convolution operation; This is a feature splicing operation.

6. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 5, characterized in that, The input modules G1 and G2 are connected to the encoder, and then sequentially connected to the connection module and the decoder to construct a U-shaped network. The output features of the downsampling module in the encoder are connected one by one to the two-dimensional convolutional layer of the upsampling module corresponding to the U-shaped network to form a complete generator for spatiotemporal error correction of ensemble forecast precipitation.

7. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 1, characterized in that, The discriminant for constructing the spatiotemporal error correction of ensemble precipitation forecasts specifically includes: The input module D1 of the discriminator is the output feature of the generator. Input module D2 is a condition , , This indicates actual observed precipitation; A symmetrical residual block stack structure consisting of two paths is constructed. Each residual block is composed of a two-dimensional convolutional layer and two sets of leaky activation function layers connected to two-dimensional convolutional layers, as shown in the following expression: ； ； In the formula, and The first and second pathways are respectively Output characteristics of each residual block; and These are the input features of the first and second pathways, respectively. and The first and second pathways are respectively The weights of the first convolutional layer in each residual block; , and , The first and second pathways are respectively the first and second pathways. The weights of the second and third convolutional layers in each residual block; To leak the activation function layer; Construct the probability output module in the discriminator. The probability output module is constructed by concatenating... and The probability that the output of the leaky activation function layer, the 2D convolutional layer, the global average pooling layer, the fully connected layer, and the activation function layer is true is expressed as follows: ； In the formula, Output probability; It is a fully connected layer; This is a global average pooling layer; These are the weights of the two-dimensional convolutional layer in the probability output module; Input modules D1 and D2 are sequentially connected to a symmetrical residual block stacking structure and a probability output module to form a complete discriminator for ensemble forecast precipitation spatiotemporal error correction.

8. The method for correcting spatiotemporal errors of ensemble forecast precipitation information based on image super-resolution according to claim 1, characterized in that, A conditional generative adversarial network is constructed by connecting a generator and a discriminator, where the generator receives random variables. With condition variables It outputs ensemble forecast precipitation false images, and the discriminator receives the ensemble forecast precipitation false images and condition variables. Output probability; Define the loss function of the discriminator respectively Loss function of generator : ； in, ; ； in, ; ； In the formula, and These are the discriminator and the generator, respectively. and These are the parameters for the discriminator and the generator, respectively. The weights are used for gradient penalty. For actual precipitation observation; and The mean probability of generating true or false images and The penalty factor; This represents the function for calculating the mean square error; A function for calculating fair skill scores; A phased training strategy is adopted. First, the generator is pre-trained based on the training samples using MSE as the loss function, and then... To fix the loss function, the generator is fixed, the discriminator is trained, and finally... The discriminant is fixed by the loss function, and the generator is trained to obtain a stable ensemble forecast precipitation statistical post-processing model. The test samples are placed into the trained stable ensemble forecast precipitation statistical post-processing model, and the output is the statistically post-processed ensemble forecast precipitation data.

9. A spatiotemporal error correction system for ensemble forecast precipitation information based on image super-resolution, used to execute the spatiotemporal error correction method for ensemble forecast precipitation information based on image super-resolution as described in any one of claims 1 to 8, characterized in that, include: The data pre-extraction module is used to acquire multi-source element data within the defined rectangular boundary of the watershed, including aggregated forecast precipitation information, actual observed precipitation information, 2m air temperature, total cloud cover and DEM data, and to perform spatiotemporal scale registration. The sample preprocessing module is used to divide training samples and test samples according to time, filter effective samples in the training samples and perform sample augmentation, and carry out data homogenization processing on this basis. Generator; The generator uses a U-shaped network structure combined with the input module G to generate ensemble forecast precipitation false images that approximate the actual spatial distribution of precipitation; Discriminator; the discriminator uses a residual network structure combined with input module D to determine the probability that the generated precipitation spatial distribution image is a real image; Post-processing module; The generator and discriminator are connected to form an adversarial network, and the post-processing module is obtained by alternating training based on the training samples; The error correction result output module is used to input the test samples into the post-processing module to generate statistically post-processed ensemble forecast precipitation results.

Images