Improved weather radar echo short-time nowcasting method based on diffusion model

Abstract

The invention relates to an improved weather radar echo short-time nowcasting method based on a diffusion model, and the method comprises the steps: obtaining a to-be-measured radar echo image, combining the to-be-measured radar echo image with time sequence information, inputting the to-be-measured radar echo image into a radar echo extrapolation model, and obtaining a denoised radar echo image; the radar echo extrapolation model is obtained by using a training set to train a diffusion model; the training set is a combined reflectivity data set; and processing the image blocks through a core denoising module in the radar echo extrapolation model, and predicting noise. According to the invention, the precision and anti-interference capability of short temporary rainfall prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of meteorological forecasting and artificial intelligence, and in particular to a weather radar echo short-term nowcasting method based on an improved diffusion model. Background Art

[0002] Weather radar echo short-term nowcasting technology is an important part of the field of meteorological forecasting, especially playing a key role in the early warning of disastrous weather. Weather radar is the core means of refined severe convective weather monitoring and short-term forecasting. Forecasters use radar echo images to judge the rainstorm area and evolution trend, and make decision recommendations. Therefore, achieving accurate radar echo extrapolation has become the key to disaster prevention and mitigation in the reservoir area. With the development of meteorological radar technology, the new generation of weather radar systems has been able to provide high-resolution observation data, providing richer information for short-term nowcasting. However, existing forecasting methods still face many challenges in practical applications, and the main defects are as follows:

[0003] (1) Limitations of data processing and algorithm optimization:

[0004] Timeliness and accuracy issues: Existing radar echo extrapolation methods need to improve their timeliness and accuracy when processing large amounts of data. For example, traditional optical flow and Lagrangian satellite cloud image extrapolation methods are prone to errors when processing echo data under complex meteorological conditions, resulting in inaccurate forecast results.

[0005] Blind spots in complex terrain and weather conditions: Existing methods have certain blind spots and errors in complex terrain and weather conditions, affecting the accuracy and effectiveness of forecasts. For example, although X-band radar has certain recognition and tracking capabilities in severe weather forecasting, it still has blind spots and errors in complex terrain and weather conditions.

[0006] (2) Shortcomings of multi-source data fusion:

[0007] Imperfect data fusion technology: In practical applications, existing multi-source data fusion technology needs to improve the temporal and spatial resolution and consistency of data. For example, although the temporal resolution of the new generation of weather radar network mosaic systems has been reduced from 6 minutes to 3 minutes, further optimization is still needed.

[0008] Data quality control: Radar data quality control is one of the key factors affecting forecast accuracy. Existing data quality control methods are prone to data noise and outliers when processing large amounts of data, affecting the accuracy of forecast results.

[0009] (3) Model generalization and adaptability:

[0010] Insufficient model generalization: Existing deep learning models lack generalization capabilities when processing different types of meteorological data, resulting in significant differences in application results across different regions. For example, the radar echo extrapolation model based on deep learning exhibits significant variations in accuracy and false alarm rates when processing precipitation data from different regions.

[0011] Model adaptability: Existing models lack adaptability to rapidly changing meteorological conditions, resulting in reduced timeliness of forecast results. For example, while traditional convolutional neural networks and recurrent neural networks improve forecast accuracy when processing short-term heavy rainfall data, they still have certain limitations when responding to sudden weather events. Summary of the Invention

[0012] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a short-term nowcasting method for weather radar echoes based on an improved diffusion model. By improving the diffusion model and combining advanced data processing and multi-source data fusion technology, the present invention aims to improve the timeliness and accuracy of the short-term nowcasting of weather radar echoes, solve the defects in the prior art, and provide more reliable technical support for weather forecasting.

[0013] To achieve the above object, the present invention provides the following solutions:

[0014] An improved weather radar echo short-term nowcasting method based on a diffusion model includes:

[0015] Obtaining a radar echo image to be measured, inputting the radar echo image to be measured into a radar echo extrapolation model in combination with time series information, obtaining a denoised radar echo image, and performing a short-term rainfall forecast based on the denoised radar echo image; the radar echo extrapolation model is obtained by training a diffusion model using a training set; the training set is a combined reflectivity dataset;

[0016] The image blocks are processed by the core denoising module in the radar echo extrapolation model to predict the noise.

[0017] Optionally, obtaining the combined reflectivity dataset includes:

[0018] Radar echo data of different times and different regions are obtained, and the combined reflectivity data set is obtained by combining the radar echo data with automatic weather station data and satellite remote sensing data.

[0019] Optionally, constructing the diffusion model includes: replacing the U-Net architecture in the original diffusion model with the core denoising module to obtain the diffusion model.

[0020] Optionally, training the diffusion model using the training set includes:

[0021] Through the forward process of the diffusion model, the radar echo observation image of the previous target time is combined with the terrain characteristics as conditional information input to generate a Gaussian noise field with the same shape as the predicted target echo image;

[0022] Adding the Gaussian noise field to the target echo image to form a noisy radar echo image;

[0023] splicing the radar echo observation image and the noisy radar echo image to obtain an image field of the target channel;

[0024] Tiling the image field to form image blocks, inputting the image blocks into the core denoising module for position encoding to obtain an image set, inputting the image set and a timestamp into the visual Transformer block for encoding, and processing the encoding results to obtain predicted noise;

[0025] The Gaussian noise field and the predicted noise are averaged by MSE, the MSE average result is used as the current training loss, and the model weight is adjusted through the inverse process of the diffusion model.

[0026] Optionally, the visual Transformer block includes:

[0027] Self-attention mechanism unit:

[0028]

[0029] Among them, Q, K, and V represent query, key, and value matrices respectively. is the dimension of the key vector, is the matrix transpose, is the value matrix;

[0030] Feedforward neural network unit:

[0031]

[0032] Among them, W1, W2 and b1, b2 represent the weight matrix and bias term respectively. is the input vector.

[0033] Optionally, adjusting the model weight includes:

[0034] Through the reverse process, the model weights are adjusted by minimizing the denoising score:

[0035]

[0036] in, is the actual noise, is the total optimization function, For the model in parameters Under the reconstructed noise, is the scaling factor, is the initial sample, is the time step.

[0037] Optionally, the forward process includes:

[0038]

[0039] in, Forward diffusion distribution A sample of Forward diffusion distribution A sample of is the identity matrix, is the standard normal distribution, The preset variance schedule determines the amount of noise added at each step.

[0040] Optionally, the reverse process includes:

[0041]

[0042] in, is the reverse diffusion distribution, is the mean of the Gaussian distribution, is a constant.

[0043] Optionally, acquiring the denoised radar echo image includes:

[0044] Step 1: Acquire a radar echo image to be measured, and use the radar echo image to be measured in combination with time series information to obtain the radar echo image to be measured at the previous target time;

[0045] Step 2: Using the radar echo image to be measured at the time of the preceding target as a condition and splicing the radar echo image to be measured after adding noise to form an image field to be measured;

[0046] Step 3: Input the image field to be measured into the core denoising module in the radar echo extrapolation model to obtain a noise prediction result, and subtract the noise prediction result from the currently input noisy radar echo to be measured to obtain a new image field, and use the new image field as the next noisy radar echo image to be measured;

[0047] Step 4: Repeat steps 1 to 3. When the cycle threshold is reached, the denoised radar echo image is obtained.

[0048] The beneficial effects of the present invention are:

[0049] This invention improves the accuracy and timeliness of radar's identification of hazardous weather events by introducing advanced data processing algorithms. For example, deep learning methods are used to optimize radar data processing, improving the spatial and temporal resolution and consistency of the data. Algorithm parameters are optimized to reduce the impact of data noise and outliers, thereby enhancing the accuracy and reliability of forecast results.

[0050] This paper uses a multi-source data fusion method, combining data from ground observation stations, meteorological satellites, and radar to improve the temporal and spatial resolution and consistency of data. For example, based on the optimal overlap region fusion method, it realizes the mosaic of weather radars with different bands, improving the precision of the data. The introduction of multi-source observation data such as sounding and raindrop spectrometers adds information such as early storm characteristics, optimizes the severe convection product algorithm, and improves the accuracy and timeliness of forecasts.

[0051] This invention improves the diffusion model's generalization and adaptability, enabling it to better handle different types of meteorological data. For example, by employing a high-dimensional, multivariate fusion model, the model's effectiveness in different regions is enhanced. The model's training and validation processes are optimized to improve its adaptability to rapidly changing meteorological conditions, ensuring the timeliness and accuracy of forecast results.

[0052] In summary, the present invention aims to improve the timeliness and accuracy of weather radar echo short-term nowcasting by improving the diffusion model and combining advanced data processing and multi-source data fusion technology, thereby solving the defects in the existing technology and providing more reliable technical support for weather forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 This is a diagram of the Vision Transformer architecture of an embodiment of the present invention;

[0055] Figure 2 This is a diagram of the RadarDiT architecture of an embodiment of the present invention;

[0056] Figure 3 This is a flow chart of a method for short-term weather radar echo nowcasting based on an improved diffusion model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0058] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0059] like Figure 1 As shown, this embodiment discloses a weather radar echo short-term nowcasting method based on an improved diffusion model, including: obtaining a radar echo image to be measured, inputting the radar echo image to be measured into a radar echo extrapolation model in combination with time series information, and obtaining a denoised radar echo image; the radar echo extrapolation model is obtained by training a diffusion model using a training set; the training set is a combined reflectivity data set; and processing image blocks using a core denoising module in the radar echo extrapolation model to predict noise.

[0060] Specifically, this embodiment discloses a method for short-term weather radar echo nowcasting based on an improved diffusion model. The method includes: constructing a diffusion process constrained by meteorological physics based on an existing meteorological big data cloud platform database to enhance the modeling of spatiotemporal evolution; constructing a diffusion denoising model with ViT as the denoising core, and the radar echo image generated by denoising is the radar echo extrapolation result generated by the model. Based on the ZR relationship, the echo extrapolation result can be converted to rainfall intensity extrapolation, thereby achieving short-term rainfall forecasting. Radar echo noise interference is suppressed, and a common data set is used to evaluate the model's data distribution learning ability to verify model usability. An adaptive time step control algorithm is designed to improve extrapolation efficiency by more than 50%. The terrain feature embedding module is integrated to improve prediction accuracy in complex terrain areas.

[0061] The terrain feature embedding module extracts terrain features, including altitude, slope, aspect, and relief, from a high-resolution digital elevation model (DEM). These features are normalized and combined with radar echo data as input to the diffusion model. During the model input phase, terrain features are used as additional input channels and concatenated with the radar echo data to form a multi-channel input tensor.

[0062] Second, within the diffusion model, a dedicated module is designed to handle the interaction between terrain features and radar echo data. Convolutional layers or attention mechanisms are used to capture the spatial and temporal correlations between terrain features and radar echo data. Through multi-scale convolution operations, the model can extract joint features at different scales, thereby better simulating the propagation and evolution of radar echoes in complex terrain.

[0063] Finally, the model is trained on a joint dataset containing terrain features and radar echo data, and the model performance is verified on an independent test set.

[0064] 1. Constructing a combined reflectivity dataset for the rainfall process:

[0065] To train the generative diffusion model, we first need to construct a high-quality dataset of combined rainfall reflectivity. This dataset collects and integrates radar echo data from different times and regions, combining dual-polarization radar observations, automatic weather station data, and satellite remote sensing data to form a collection of samples covering a wide range of rainfall types and intensities. The dataset must ensure temporal and spatial continuity, diversity, and accuracy of the data to provide rich learning material for model training.

[0066] 2. Construct a diffusion denoising model based on ViT:

[0067] This paper proposes a diffusion model (RadarDiT) with the Vision Transformer (ViT) as its core denoising module. This model replaces the U-Net architecture in traditional diffusion models with the ViT architecture, leveraging ViT's ability to process image patches to improve the model's denoising performance for radar echo images. Pre-training on common image datasets (such as ImageNet) verifies the model's ability to learn data distributions, ensuring its applicability to radar echo extrapolation tasks.

[0068] 3. Introducing image condition constraints to adapt to radar extrapolation generation tasks:

[0069] To better adapt the diffusion model to the radar echo extrapolation task, an image conditional constraint mechanism was introduced. Specifically, the current radar echo image is used as a conditional input, combined with time series information, to guide the model in generating radar echo images for future moments. This approach enables the model to better capture the spatiotemporal evolution of rainfall processes. Ultimately, a radar echo extrapolation model, named RadarDiT, was developed and trained and analyzed using an extrapolation dataset.

[0070] 4. Model Performance Verification

[0071] To validate the extrapolation performance of the RadarDiT model, the generated forecasts were evaluated using various metrics (such as CSI, FAR, and POD skill score). Experiments compared the model with other existing radar echo extrapolation models (such as PhyDNet, PySTEPS, PredRNN, ConvLSTM, DGMR, and SmatUnet). Testing on the same extrapolation dataset analyzed the differences in prediction accuracy, generalization, and computational efficiency. The experimental results demonstrate that the RadarDiT model demonstrates superior accuracy and stability in radar echo extrapolation tasks.

[0072] Furthermore, obtaining the combined reflectivity dataset includes: obtaining radar echo data at different times and in different regions, and obtaining the combined reflectivity dataset by combining the radar echo data with automatic weather station data and satellite remote sensing data.

[0073] Specifically, data preprocessing:

[0074] Purpose: Ensure the quality and consistency of input data and provide a reliable data foundation for model training.

[0075] Data Collection: Echo data from multiple weather radars within a designated area will be collected with a temporal resolution of 6 minutes and a spatial resolution of 1 km. Radar echo data will include dual-polarization radar variables such as reflectivity (Z), differential reflectivity (Zdr), and specific differential phase shift (Kdp). Minute-level precipitation observations from automatic stations will be collected over the past five years. ERA5 reanalysis data will also be collected, including environmental information such as temperature, humidity, and wind, with a temporal resolution of 1 hour and a spatial resolution of 0.25° × 0.25°.

[0076] Based on rainfall thresholds, we preliminarily screen minute-by-minute automatic station rainfall observation data to obtain automatic station monitoring results for the rainfall process. Rainfall pattern analysis, rainfall intensity matching, and maximum value assessment are performed on the screened automatic station data to comprehensively determine the rainfall type to which the data belongs and construct a sample set.

[0077] Quality Control: Radar echo data is quality controlled to remove non-meteorological echoes (such as ground clutter and sea clutter). A threshold-based approach is used, such as removing areas with a reflectivity factor below 15 dBZ as non-meteorological echoes. Data quality control is performed using temporal continuity and spatial consistency checks. Reanalysis data are interpolated to ensure that their temporal and spatial resolutions are consistent with those of the radar echo data.

[0078] Data Interpolation: Radar echo data are interpolated into a unified Cartesian coordinate system to ensure data consistency. Bilinear interpolation is used to convert the radar echo data from polar coordinates to Cartesian coordinates. Reanalysis data are interpolated in time and space to ensure that their temporal and spatial resolutions are consistent with those of the radar echo data. Bilinear interpolation is used for spatial interpolation, and linear interpolation is used for temporal interpolation.

[0079] Function: Through data preprocessing, ensure the quality and consistency of input data and provide a reliable data foundation for model training.

[0080] Furthermore, constructing the diffusion model includes: replacing the U-Net architecture in the original diffusion model with the core denoising module to obtain the diffusion model.

[0081] Diffusion improvement model construction and training:

[0082] The diffusion model is a generative model that learns the data distribution by gradually adding noise to the data through a Markov process and then removing the noise through the reverse process. During training, the model learns how to gradually remove the noise from pure data and restore the original data, thereby generating high-quality images or other data. During the generation process, sampling is performed by reversing the random diffusion process.

[0083] The diffusion model is divided into a forward process and a reverse process. The forward process is fixed, usually Gaussian distributed, and discretized into t∈1,…,T time steps, and its recursive definition is:

[0084] (1)

[0085] Where: Forward diffusion distribution samples; I is the identity matrix; N is the Gaussian distribution, and the parameters Controls the amount of noise added to each step of the conversion, and its value depends on the specific forward process used. Iterating through Equation (1) will increase Gaussian noise. When T is large, x T The inverse process is parameterized by the learned model parameters θ and adopts a Gaussian form, which can be expressed as:

[0086] (2)

[0087] Where: variance Usually set to a constant; It is a reverse diffusion distribution; is the mean of the Gaussian distribution. Here, use Expressed as:

[0088] (3)

[0089] Where: is the noise ratio at time step t; =∏s=1t αs, which is the noise accumulation coefficient up to time step t; is the latent space feature at time t and ∈N(0,1). Function is called the score function, which can be interpreted as a noise estimator for Denoising is performed to produce The diffusion model is usually trained by minimizing the denoising score, which can be expressed as:

[0090] (4)

[0091] Where: is the actual noise. You can initialize x with Gaussian noise t And iteratively apply the inverse learning process given in Equation (2) to extract samples from the model. Providing additional inputs can make the model conditional. For the diffusion denoising model, noise prediction is achieved by training a deep neural network, such as UNet. In this paper, a multi-layer Vision Transformer is used to construct this noise prediction network. The Vision Transformer is described in detail below.

[0092] The Vision Transformer (ViT) is a deep learning model that uses the Transformer architecture to process images. ViT divides an image into a series of small patches, which are treated as elements in a sequence. Each patch is first mapped to a high-dimensional space. These mapped patches (with positional encoding to preserve spatial information) are then fed into a standard Transformer model for processing. This approach allows ViT to leverage the Transformer's powerful self-attention mechanism to capture complex relationships between patches, thereby learning rich image feature representations.

[0093] like Figure 2 As shown in Figure 2, the core calculation formula of ViT involves two main parts of Transformer: the Self-Attention mechanism and the Feed-Forward Neural Network (FFNN).

[0094] The self-attention mechanism can be expressed as the following formula:

[0095] (2.1)

[0096] Here, Q, K, and V represent query, key, and value matrices, respectively, which are obtained by linear transformation of the input. is the dimension of the key vector, which is used to scale the dot product to prevent the gradient from disappearing. The attention mechanism enables the model to focus on different parts of the input image, thereby capturing the inherent global dependencies. This self-attention architecture of the Transformer enables ViT to learn powerful visual representations by directly modeling the relationship between image blocks without any convolution operations. The feedforward neural network (FFNN) can be expressed as the following formula:

[0097] (2.2)

[0098] FFNN is applied after each self-attention layer and typically consists of two linear transformations and a ReLU activation function. W1, W2 and b1, b2 represent the weight matrix and bias terms, respectively. By stacking multiple self-attention and FFNN layers, ViT has shown great potential in processing visual data and is capable of handling complex image tasks.

[0099] ViT has demonstrated impressive performance in visual tasks such as image classification, object detection, and image segmentation. Compared to CNNs, which rely on local convolution operations, ViT directly captures long-range dependencies between image blocks through a global self-attention mechanism, enabling it to process image data more efficiently. In addition, another advantage of the ViT model is its scalability, which can improve performance by increasing the depth and width of the model, making ViT particularly effective when processing large-scale image datasets. Therefore, the present invention selects it as the core component of the denoising network.

[0100] Furthermore, using the training set to train the diffusion model includes: through the forward process of the diffusion model, using the radar echo observation image of the previous target time as conditional information input, generating a Gaussian noise field consistent with the shape of the predicted target echo image; adding the Gaussian noise field to the target echo image to form a noisy radar echo image; splicing the radar echo observation image and the noisy radar echo image to obtain an image field of the target channel; tile the image field to form image blocks, input the image blocks into the core denoising module for position encoding, obtain an image set, input the image set and the timestamp into the visual Transformer block for encoding, and process the encoding result to obtain predicted noise; MSE average the Gaussian noise field and the predicted noise, use the MSE average result as the current training loss, and adjust the model weight through the reverse process of the diffusion model.

[0101] Specifically, the previous five time periods are used to predict the next five time periods. For a random time step TimeStep, a Gaussian noise field Noise with the same shape as the predicted result is generated, and then added to the predicted target echo image to form the noisy radar echo image X t The echo images of the first five time series are used as conditions and X t Do concatenation Concat to form a 10-channel image field X as the input of the denoising model.

[0102] Patch the X image into image blocks to form image block tokens, according to V i By default, the patch is positionally encoded and added to the tokens. The token set and timestamp are passed to a multi-layer VisionTransformer block for encoding. The encoded result is processed and converted to a 5x300x400 pixel value, which is the predicted noise. The mean square error (MSE) of the original Gaussian noise field and the predicted noise is averaged to obtain the current training loss. Backpropagation is used to adjust the model weights, thereby achieving model training.

[0103] Furthermore, based on the diffusion model, denoising is introduced into Vision Transformer to form DiT, and the radar echo observation data of the first five time periods are input as conditional information, thereby realizing the conditional diffusion model to generate RadarDiT, as shown in Figure 3 shown.

[0104] Furthermore, obtaining the denoised radar echo image includes: step 1, obtaining the radar echo image to be measured, and using the radar echo image to be measured in combination with time series information to obtain the radar echo image to be measured at the previous target time; step 2, using the radar echo image to be measured at the previous target time as a condition to splice the radar echo image to be measured after adding noise to form an image field to be measured; step 3, inputting the image field to be measured into the core denoising module in the radar echo extrapolation model to obtain a noise prediction result, and subtracting the noise prediction result from the currently input noisy radar echo to be measured to obtain a new image field, and using the new image field as the next noisy radar echo image to be measured; step 4, repeating steps 1 to 3, and when a cycle threshold is reached, obtaining the denoised radar echo image.

[0105] Specifically, the random noise is denoised 1000 times in a DDPM manner to obtain the final generated result. t In the first denoising generation, that is, TimeStep is 1, a 5*300*400 Gaussian noise field Noise is randomly generated, and the other TimeSteps are the previous generation results. The echo images of the first 5 time series are used as conditions and X t Do Concat to form a 10-channel image field X as the input of the denoising model. Then input the above trained model for noise prediction and convert the input X t Subtract the predicted noise to get the image field and use it as the next input X t .

[0106] The above denoising process is repeated 1000 times to obtain the final generated result.

[0107] Forecast results evaluation:

[0108] Evaluation metrics: In terms of the performance evaluation of the unconditional diffusion model data distribution learning, we use mean, variance RMSE and spatial similarity to evaluate the model data distribution learning performance.

[0109] MSE (mean squared error) is used to measure the difference between each pixel of the predicted image and the true image. The smaller the MSE, the smaller the difference between the generated image and the true image, and the better the model performance. The calculation formula is as follows:

[0110]

[0111] In the formula, is the true value, is the generated value.

[0112] To evaluate the performance of radar echo extrapolation models, we use commonly used forecasting and analysis metrics in the meteorological field: the Critical Success Index (CSI), the False Alarm Rate (FAR), and the Probability of Detection (POD). The CSI is a comprehensive metric that takes into account both FAR and POD and directly reflects the quality of the model. A higher CSI indicates better performance.

[0113] When calculating the above indices, the generated and observed echo patterns are converted into binary matrices by setting thresholds. If the radar echo value is greater than the given threshold, the corresponding value is set to 1; otherwise, it is set to 0. By traversing each pixel position, the true negative prediction number (TN), false negative prediction number (FN), true positive prediction number (TP), and false positive prediction number (FP) can be calculated. The specific correspondence between the prediction and the true value for each result is shown in Table 1:

[0114] Table 1 Contingency table

[0115] Prediction = 0 Prediction = 1 True = 0 TN (true negative) FN (false positive) True = 1 FN (false negative) TP (True Positive)

[0116] The calculation formulas for CSI, FAR and POD are as follows:

[0117] CSI = TP / (TP+FN+FP)

[0118] FAR=FN / (TP+FN)

[0119] POD = TP / (TP+FP)

[0120] At the same time, MSE is used as an evaluation criterion to intuitively reflect the degree of similarity of the extrapolated images. By comparing with existing methods (such as PhyDNet, PySTEPS, PredRNN, ConvLSTM, DGMR, and SmatUnet), the advantages of the diffusion model in forecasting accuracy and reliability are verified.

[0121] Encoder-decoder structure: The Vision Transformer (ViT) is used as the encoder, combined with the denoising capability of the generative diffusion model to improve the feature extraction and generation of radar echo images.

[0122] Conditional input mechanism: The radar echo image at the current moment is introduced as conditional input, combined with time series information to enhance the model's ability to capture spatiotemporal evolution.

[0123] Optimized training and evaluation methods:

[0124] Loss function design: A combination of mean square error (MSE), perceptual loss, and adversarial loss is used to optimize the training effect of the model.

[0125] Performance Verification: Through comparative experiments with existing methods, the RadarDiT model is verified to have significantly improved performance in forecasting short-term heavy rainfall and thunderstorms.

[0126] Multi-source data fusion and dynamic boundary processing:

[0127] Multi-source data fusion: Combine radar data, meteorological observation data, satellite remote sensing data and other data to improve the accuracy and reliability of radar echo extrapolation.

[0128] Dynamic boundary processing: Solve the dynamic boundary problem in radar echo extrapolation by introducing dynamic boundary conditions or adaptive grid methods;

[0129] Specifically, physical quantities (such as wind fields, temperature fields) or radar echo characteristics at the boundary are predicted through meteorological physical models (such as WRF) or data-driven methods (such as deep learning models), so that the boundary conditions change dynamically over time and space, which is closer to the actual movement of weather systems.

[0130] Adaptive Grid Method: Dynamically adjusts grid resolution based on the temporal and spatial distribution of radar echoes. High-resolution grids are used in areas with drastic radar echo variations or severe convective weather, while low-resolution grids are used in areas with more gradual variations. Furthermore, the grid resolution is dynamically adjusted based on the speed of radar echo variation, with high resolution used for rapidly varying echoes and low resolution for slowly varying echoes.

[0131] Combining the two: dynamic boundary conditions guide adaptive grid adjustment, increasing the grid resolution in the boundary change area in advance; adaptive grid supports dynamic boundary condition updates, using high-resolution grid data to more accurately update boundary conditions, thereby integrating multi-source data information to enhance the quality and information content of radar echo images.

[0132] Real-time performance and computational efficiency:

[0133] Efficient computing: Improve the real-time performance and computing efficiency of the model through optimized algorithm design, parallel computing, and hardware acceleration.

[0134] Low-energy reasoning: Adopts an efficient diffusion model architecture to reduce the energy consumption of single-token reasoning.

[0135] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A weather radar echo short-term nowcasting method based on an improved diffusion model, characterized in that: include: Obtaining a radar echo image to be measured, inputting the radar echo image to be measured into a radar echo extrapolation model in combination with time series information, obtaining a denoised radar echo image, and performing a short-term rainfall forecast based on the denoised radar echo image; the radar echo extrapolation model is obtained by training a diffusion model using a training set; the training set is a combined reflectivity dataset; The image blocks are processed by the core denoising module in the radar echo extrapolation model to predict the noise.

2. The improved weather radar echo short-term nowcasting method based on the diffusion model according to claim 1 is characterized in that: Acquiring the combined reflectance dataset includes: Radar echo data of different times and different regions are obtained, and the combined reflectivity data set is obtained by combining the radar echo data with automatic weather station data and satellite remote sensing data.

3. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 1 is characterized in that: Constructing the diffusion model includes: replacing the U-Net architecture in the original diffusion model with the core denoising module to obtain the diffusion model.

4. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 1 is characterized in that: Training the diffusion model using the training set includes: Through the forward process of the diffusion model, the radar echo observation image of the previous target time is combined with the terrain characteristics as conditional information input to generate a Gaussian noise field with the same shape as the predicted target echo image; Adding the Gaussian noise field to the target echo image to form a noisy radar echo image; splicing the radar echo observation image and the noisy radar echo image to obtain an image field of the target channel; Tiling the image field to form image blocks, inputting the image blocks into the core denoising module for position encoding to obtain an image set, inputting the image set and a timestamp into the visual Transformer block for encoding, and processing the encoding results to obtain predicted noise; The Gaussian noise field and the predicted noise are averaged by MSE, the MSE average result is used as the current training loss, and the model weight is adjusted through the inverse process of the diffusion model.

5. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 4 is characterized in that: The visual Transformer block includes: Self-attention mechanism unit: Among them, Q, K, and V represent query, key, and value matrices respectively. is the dimension of the key vector, is the matrix transpose, is the value matrix; Feedforward neural network unit: Among them, W1, W2 and b1, b2 represent the weight matrix and bias term respectively. is the input vector.

6. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 4 is characterized in that: Adjusting the model weights includes: Through the reverse process, the model weights are adjusted by minimizing the denoising score: in, is the actual noise, is the total optimization function, For the model in parameters Under the reconstructed noise, is the scaling factor, is the initial sample, is the time step.

7. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 4 is characterized in that: The forward process includes: in, Forward diffusion distribution A sample of Forward diffusion distribution A sample of is the identity matrix, is the standard normal distribution, The preset variance schedule determines the amount of noise added at each step.

8. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 4 is characterized in that: The reverse process includes: in, is the reverse diffusion distribution, is the mean of the Gaussian distribution, is a constant.

9. The improved weather radar echo short-term nowcasting method based on diffusion model according to claim 1, characterized in that: Acquiring the denoised radar echo image includes: Step 1: Acquire a radar echo image to be measured, and use the radar echo image to be measured in combination with time series information to obtain the radar echo image to be measured at the previous target time; Step 2: Using the radar echo image to be measured at the time of the preceding target as a condition and splicing the radar echo image to be measured after adding noise to form an image field to be measured; Step 3: Input the image field to be measured into the core denoising module in the radar echo extrapolation model to obtain a noise prediction result, and subtract the noise prediction result from the currently input noisy radar echo to be measured to obtain a new image field, and use the new image field as the next noisy radar echo image to be measured; Step 4: Repeat steps 1 to 3. When the cycle threshold is reached, the denoised radar echo image is obtained.

Images