Two-stage short temporary rainfall prediction diffusion model method based on satellite cloud picture and autoregression mechanism

Abstract

The invention discloses a two-stage short temporary rainfall prediction diffusion model method based on a satellite cloud picture and an autoregression mechanism, and the method specifically comprises the following steps: firstly, inputting the monitoring data of the satellite cloud picture at the previous T1 moment, carrying out the interpolation processing, and finally carrying out the normalization processing; in the first stage, a satellite radar inversion network is utilized to invert an input satellite cloud picture sequence to obtain a radar echo picture sequence at a corresponding moment; in the second stage, aiming at time sequence features of a radar echo map sequence, through a constructed time sequence projection module based on a multi-scale partitioning technology and a KAN network and a motion module based on causal attention, aiming at space texture features, through splicing a previous time step inversion feature and current noise, a new input noise tensor is formed, and the new input noise tensor is input into the radar echo map sequence; fusing the captured features in the denoising network, and outputting a prediction result of the current target time step; and finally, gradually generating a video sequence with complete T2 duration. The problem of error accumulation in prediction can be effectively relieved, and the precision of rainfall prediction is improved.

Description

Technical Field

[0001] The present invention belongs to the field of meteorological forecasting research, and specifically designs a two-stage short-term precipitation prediction diffusion model based on satellite cloud images and an autoregressive mechanism. Background Art

[0002] Precipitation plays a crucial role in the Earth's hydrological cycle. Various precipitation processes play a crucial role in the circulation of water, energy, and matter within the Earth system. Precipitation activities of varying scales and intensities also have a certain impact on human activities and socioeconomic development. my country, with its vast territory, extensive latitude and longitude, diverse climate types, and complex geographical environment, has experienced frequent extreme precipitation disasters under the influence of climate change in recent years, posing a severe challenge to comprehensive meteorological disaster prevention and mitigation. Therefore, further development of precipitation forecasting technology remains an essential requirement for public services such as agriculture, forestry, disaster management, and infrastructure development.

[0003] Traditional precipitation prediction methods based on radar echo extrapolation are difficult to carry out in remote areas or offshore areas where ground radar observation points are sparse. Based on this, inverting radar echoes through satellite cloud images to achieve precipitation prediction has become a research hotspot.

[0004] Although precipitation prediction methods based on deep learning have achieved certain results, current methods still tend to directly use radar data for extrapolated predictions. Precipitation prediction methods based on a single satellite data source still have problems such as image blurring and missing details caused by the loss function, as well as feature information loss caused by the model being too deep, which in turn affects prediction accuracy.

[0005] To address this issue, the present invention proposes a two-stage short-term precipitation prediction diffusion model method based on satellite cloud imagery and an autoregressive mechanism. This method utilizes historical satellite cloud image data and enhances the diversity of results through diffusion model algorithm training, thereby improving the accuracy of precipitation prediction and the realism and detail expression of radar echo images. Summary of the Invention

[0006] The purpose of the present invention is to provide a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism. This scheme can not only improve the accuracy of precipitation prediction, but also help to enhance the realism and detail expression of radar echo images.

[0007] To achieve the above-mentioned object, the present invention adopts the following technical solution: a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism, comprising the following steps:

[0008] S1: Data input and preprocessing: input the monitoring data of 16 channels of satellite cloud images at the previous time T1, perform interpolation processing, then select 6 channels of data including 4 infrared microwave channels and 2 brightness temperature difference channels, and finally perform normalization processing to obtain the cleaned data;

[0009] S2: Satellite radar inversion, using the satellite radar inversion network to invert the input satellite cloud image sequence to obtain the radar echo image sequence X at the corresponding time;

[0010] S3: Modeling the radar echo sequence time series projection, building a time series projection module Patch-KAN module based on multi-scale block technology and KAN network to capture its local spatiotemporal details and output the result Z;

[0011] S4: Radar echo sequence motion trajectory modeling uses a causal attention method to directly perform causal modeling on cross-frame tokens in the latent space, ensuring the unidirectionality of information flow, looking only at the past and not the future, meeting the requirements of autoregressive generation, and obtaining representation results.

[0012] S5: Radar echo sequence spatial texture modeling, combining the inversion feature X(t-1) of the previous time step with the current noise Concatenate in the channel dimension to form a new input tensor x t At the same time, X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj .

[0013] S6: Multi-channel feature fusion, using Z as the key and value, and using the Cross-Attention mechanism to fuse with the original features of the denoising diffusion network; Directly inject into the bottleneck layer of denoising U-Net; F proj Multiple scale feature maps {f i} to perform conditional modulation on x′ t Denoising prediction results

[0014] S7: Autoregression generates a complete video sequence and converts the prediction result of the current time step As the generation condition of the next time step, the video sequence with a complete T2 length is finally generated step by step

[0015] In the above step S1, the satellite cloud image data at the previous time T1 is first input for data sequence preprocessing, including the following steps: the spatial resolution of all channels is interpolated to the level of 0.5 km to ensure the spatial consistency of the data; then, the Pearson correlation coefficient is used to perform correlation analysis, analyzing the information of the correlation between 16 channels and precipitation, and finally selecting channels 8, 10, 11, and 12, as well as the channel brightness temperature difference 8-13 and the visible light channel difference 3-5; the selected channel data is normalized, and the original data is linearly normalized using maximum and minimum normalization to map the data to the range of (0, 1).

[0016] In the above step S2, the satellite radar inversion includes the following steps: selecting the inversion model as the fully convolutional network U-Net, adopting the encoder-decoder structure, and inverting the satellite cloud image data Y at time T1 to obtain the radar echo sequence X at time T1.

[0017] In the above step S3, the process of radar echo sequence time series projection modeling is as follows: predefine multiple patch scales s, and divide them into a multi-scale set according to spatial position, time length, and channel scale after patch flattening For each scale s and each position i, the Patch time series is expanded into a one-dimensional vector along the time dimension to obtain the input vector:

[0018]

[0019] Where T s represents the length of time at scale s, is the Patch vector at scale s, position i, and time step t.

[0020] The input vectors are respectively sent to the corresponding KAN model projection to model the nonlinear temporal dependency corresponding to the scale, and the temporal feature encoding z at scale s is obtained. (s,i) , the whole process formula is as follows:

[0021]

[0022] Among them, ψ q,j (·) is the input variable x j A learnable unary function that performs nonlinear transformations, φ q (·) is a function that further nonlinearly maps each set of weighted results, Q s Indicates the number of channels / basis functions used in this layer, P s Indicates the dimension of a single Patch at the corresponding scale, T s Indicates the length of time at scale s.

[0023] The patch embedding vectors at all scales are concatenated to obtain a unified spatiotemporal feature representation z i , the process is:

[0024]

[0025] In the above step S4, the radar echo sequence motion trajectory modeling process is as follows: given the radar echo sequence X of the input historical frame, the latent space features of each frame are flattened into a token sequence in the temporal direction Merge to get the entire historical time series input Then, causal self-attention is used to Perform linear transformation to obtain Get the query vector Q motion , key vector K motion Sum value vector V motion Three vectors:

[0026]

[0027] Among them, W Q Q motion The weight matrix, W Q K motion The weight matrix, W V V motion The weight matrix,

[0028] Therefore, the causal attention mechanism is expressed as follows:

[0029]

[0030] Where M is the triangular causal mask, which ensures that the representation of the t-th frame can only access historical information of time steps ≤ t, and d is the vector Q motion , K motion The dimension of , which is used to scale the dot product to ensure numerical stability;

[0031] Reshape the token representation of Causal Attention back to the spatial structure of the (t-1) frame and merge or aggregate it to generate the motion representation of the current frame

[0032] In the above step S5, the radar echo sequence spatial texture modeling process is as follows: the inversion feature X(t-1) of the previous time step is combined with the current noise Concatenate in the channel dimension to form a new input tensor

[0033]

[0034] At the same time, X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj Control x′ t The denoising modulation process.

[0035] In the above step S6, the time series projection modeling result Z and the motion trajectory modeling result and the control signal F obtained from spatial texture modeling proj Work together for input x′ t The denoising backbone network is used to obtain the predicted denoising result, and then the predicted result is decoded from the latent space back to the pixel space to obtain the final prediction result.

[0036] Where Z is the Key and Value, and the query vector Q of Z is obtained through linear transformation Z , key vector K Z Sum value vector V Z , the Cross-Attention mechanism is used to output the denoising UNet fusion through residual connection, LayerNorm and denoising backbone network StableDiffusion to obtain new features The formula is as follows:

[0037]

[0038] Motion trajectory modeling results Directly inject it into the bottleneck layer of U-Net and model it as motion information;

[0039] The control signal F obtained from spatial texture modeling proj , in the U-Net internal multiple scale feature maps {f i} to perform conditional modulation:

[0040]

[0041] Among them, γ i and β i By controlling the signal F proj The predicted scaling and translation parameters control the feature amplitude and distribution shift respectively.

[0042] In the above step S7, the autoregressive process generates a complete video sequence and the prediction result of the current time step is As the generation condition for the next time step, the following generation process is used to complete the generation of the complete target sequence of the autoregressive formula, and finally gradually generate the video sequence of the complete T2 length

[0043]

[0044] in is the predicted image of the t+i frame, represents the autoregressive context window for the current generation, p θ (·) represents the sampling distribution of the reverse diffusion process under the current conditions.

[0045] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and an autoregressive mechanism is implemented.

[0046] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and an autoregressive mechanism.

[0047] Compared with the prior art, the advantages of the present invention are as follows:

[0048] 1) The present invention adopts a dual temporal modeling structure of temporal projection modeling and motion trajectory modeling, which can achieve complementary modeling between global temporal and local spatial details. In the prediction task of the next hour, the PSNR of the present invention method is improved by 12.4% and 7.21% compared with the Earthformer and AA-TransUnet models, respectively.

[0049] 2) This invention introduces a causal self-attention mechanism for motion trajectory modeling, directly performing causal modeling of cross-frame tokens in the latent space, thereby obtaining a more flexible and information-rich temporal context representation. It also ensures the unidirectionality of information flow, focusing only on the past and not the future, meeting the requirements of autoregressive generation.

[0050] 3) This paper proposes a time projection module that utilizes multi-scale block segmentation and the nonlinear modeling capabilities of KAN. Multi-scale patch segmentation supports capturing local and global changes at different time-space granularities. KAN nonlinear projection characterizes complex time evolution patterns through a combination of unary functions, significantly enhancing the nonlinear modeling capabilities of the model to effectively capture the temporal evolution and spatial detail characteristics of images.

[0051] 4) The spatial texture channel of the present invention takes into account the Markov characteristics of the diffusion process and reasonably introduces X(t-1) as a conditional input, which can inject global structure priors in the early stage of denoising. At the same time, it further refines the local texture modeling through feature modulation, thereby improving the overall restoration quality.

[0052] 5) Compared with the traditional prediction model structure, the present invention adopts an autoregressive strategy to generate future sequences frame by frame, thereby improving the prediction accuracy while effectively alleviating the fuzzy accumulation problem in the long-term time series prediction process and enhancing the coherence and stability of the generated sequence. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of the two-stage short-term precipitation prediction diffusion model based on satellite cloud images and autoregressive mechanism of the present invention.

[0054] Figure 2 This is a schematic diagram of data preprocessing of the present invention,

[0055] Figure 3 Schematic diagram of the two-stage short-term precipitation prediction diffusion model network based on the autoregressive mechanism of the present invention, Figure 4 This is a schematic diagram of satellite radar inversion.

[0056] Figure 5 This is a schematic diagram of the time series projection module.

[0057] Figure 6 Schematic diagram of causal attention. DETAILED DESCRIPTION

[0058] In order to deepen the knowledge and understanding of the present invention, the present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0059] Example 1: This specific embodiment discloses a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism, such as Figures 2 to 6 As shown, the following steps are included:

[0060] S1: If Figure 2 As shown in the figure, the monitoring data of 16 channels of satellite cloud images at the time T1 before are input, interpolated, and then 6 channels of data including 4 infrared microwave channels and 2 brightness temperature difference channels are selected. Finally, normalization is performed to obtain the cleaned data.

[0061] S2: If Figure 4 As shown, satellite radar inversion uses a satellite radar inversion network to invert the input satellite cloud image sequence to obtain the radar echo image sequence X at the corresponding time;

[0062] S3: If Figure 5 As shown, radar echo sequence temporal projection modeling is performed, and a temporal projection module Patch-KAN module based on multi-scale block technology and KAN network is constructed to capture its local spatiotemporal details and output the result Z;

[0063] S4: As Figure 6As shown, the radar echo sequence motion trajectory modeling uses a CausalAttention to directly perform causal modeling on cross-frame tokens in the latent space, ensuring the unidirectionality of information flow, only looking at the past and not the future, meeting the requirements of autoregressive generation, and obtaining the representation result.

[0064] S5: If Figure 3 As shown on the left, the radar echo sequence spatial texture modeling is performed by combining the inversion feature X(t-1) of the previous time step with the current noise Concatenate in the channel dimension to form a new input tensor x t At the same time, X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj .

[0065] S6: As Figure 3 As shown in the figure, multi-channel feature fusion takes Z as the key and value, and adopts the Cross-Attention mechanism to fuse with the original features of the denoising diffusion network; Directly inject into the bottleneck layer of denoising U-Net; F proj Multiple scale feature maps {f i} to perform conditional modulation on x′ t Denoising prediction results

[0066] S7: As Figure 3 As shown, the autoregressive generates a complete video sequence and the prediction result of the current time step is As the generation condition of the next time step, the video sequence with a complete T2 length is finally generated step by step

[0067] In step S1, Figure 2 As shown, first, the satellite cloud image data at the previous T1 moment is input for data sequence preprocessing, which includes the following steps: the spatial resolution of all channels is interpolated to the level of 0.5 km to ensure the spatial consistency of the data; then the Pearson correlation coefficient is used to perform correlation analysis, analyze the information of the correlation between 16 channels and precipitation, and finally select channels 8, 10, 11, 12 and channel brightness temperature difference 8-13 and visible light channel difference 3-5; the selected channel data is normalized, and the original data is linearly normalized using maximum and minimum normalization to map the data to the range of (0, 1).

[0068] In step S2, as Figure 4As shown in the figure, the satellite radar inversion includes the following steps: the inversion model is selected as the fully convolutional network U-Net, and the encoder-decoder structure is adopted to invert the satellite cloud image data Y at time T1 to obtain the radar echo data X at time T1.

[0069] In step S3, Figure 5 As shown in the figure, the process of radar echo sequence time series projection modeling is as follows: predefine multiple patch scales S, and obtain a multi-scale set according to the spatial position, time length, and channel scale after patch flattening For each scale s and each position i, the Patch time series is expanded into a one-dimensional vector along the time dimension to obtain the input vector:

[0070]

[0071] Where T s represents the length of time at scale s, is the Patch vector at scale s, position i, and time step t.

[0072] The input vector x (s,i) The corresponding KAN model projections are respectively sent to model the nonlinear temporal dependency corresponding to the scale, and the temporal feature encoding z at scale s is obtained. (s,i) , the whole process formula is as follows:

[0073]

[0074] Among them, ψ q,j (·) is the input variable x j A learnable unary function that performs nonlinear transformations, φ q (·) is a function that further nonlinearly maps each set of weighted results, Q s Indicates the number of channels / basis functions used in this layer, P s Indicates the dimension of a single Patch at the corresponding scale, T s Indicates the length of time at scale s.

[0075] The patch embedding vectors at all scales are concatenated to obtain a unified spatiotemporal feature representation z i , the process is:

[0076]

[0077] In step S4, Figure 6 As shown, the radar echo sequence motion trajectory modeling process is as follows: given the radar echo sequence X of the input historical frame, the latent space features of each frame are flattened into a token sequence in the temporal direction Merge to get the entire historical time series input Then, causal self-attention is used to Perform linear transformation to obtain Get the query vector Q motion , key vector K motion Sum value vector V motion Three vectors:

[0078]

[0079] Among them, W Q Q motion The weight matrix, W Q K motion The weight matrix, W V V motion The weight matrix,

[0080] Therefore, the causal attention mechanism is expressed as follows:

[0081]

[0082] Where M is the triangular causal mask, which ensures that the representation of the t-th frame can only access historical information of time steps ≤ t, and d is the vector Q motion , K motion The dimension of , which is used to scale the dot product to ensure numerical stability;

[0083] Reshape the token representation of Causal Attention back to the spatial structure of the (t-1) frame and merge or aggregate it to generate the motion representation of the current frame

[0084] In step S5, if Figure 3 As shown on the left side of the figure, the spatial texture modeling process of the radar echo sequence is as follows: the inversion feature X(t-1) of the previous time step is combined with the current noise Concatenate in the channel dimension to form a new input tensor

[0085]

[0086] At the same time, X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj Control x′ t De-noising modulation process.

[0087] In step S6, Figure 3 As shown, the time series projection modeling result Z and the motion trajectory modeling result and the control signal F obtained from spatial texture modeling projThe work is done on the input x′ t The denoising backbone network is used to obtain the predicted denoising result, and then the predicted result is decoded from the latent space back to the pixel space to obtain the final prediction result.

[0088] Where Z is the Key and Value, and the query vector Q of Z is obtained through linear transformation Z , key vector K Z Sum value vector V Z , the Cross-Attention mechanism is used to output the denoising UNet fusion through the residual connection with LayerNorm and the denoising backbone network StableDiffusion. The formula is as follows:

[0089]

[0090] Motion trajectory modeling results Directly inject it into the bottleneck layer of U-Net and model it as motion information;

[0091] The control signal F obtained from spatial texture modeling proj , in the U-Net internal multiple scale feature maps {f i} to perform conditional modulation:

[0092]

[0093] Among them, γ i and β i By controlling the signal F proj The predicted scaling and translation parameters control the feature amplitude and distribution shift respectively.

[0094] In the above step S7, Figure 3 As shown, the autoregressive generates a complete video sequence and the prediction result of the current time step is As the generation condition for the next time step, the following generation process is used to complete the generation of the complete target sequence of the autoregressive formula, and finally gradually generate the video sequence of the complete T2 length

[0095]

[0096] in is the predicted image of the t+i frame, represents the autoregressive context window for the current generation, p θ (·) represents the sampling distribution of the reverse diffusion process under the current conditions.

[0097] Example 1:

[0098] The validation dataset for this method uses the geosynchronous orbiting meteorological satellite Sunflower-8, which covers the range of 60°N to 60°S and 80°W to 160°W, encompassing several areas of Jiangsu Province. The dataset spans 31 days from June to July 2019, with a data interval of 10 minutes and a total of 7,992 frames.

[0099] The proposed method is compared with the baseline model. Tables 1 and 2 show the image quality index and precipitation prediction accuracy index, respectively. The best performance is highlighted in bold. Compared with other SOTA models, our model achieves the best performance level in terms of evaluation indicators.

[0100] Table 1 Comparison of image quality accuracy between the present invention and existing algorithms

[0101]

[0102] Table 2 Comparison of precipitation prediction accuracy between the present invention and existing algorithms

[0103]

[0104] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.

Claims

1. A two-stage short-term precipitation forecast diffusion model method based on satellite cloud images and autoregressive mechanism, characterized by: The method comprises the following steps: S1: Data input and preprocessing: input the monitoring data of 16 channels of satellite cloud images at the previous time T1, perform interpolation processing, then select 6 channels of data including 4 infrared microwave channels and 2 brightness temperature difference channels, and finally perform normalization processing to obtain the cleaned data; S2: Satellite radar inversion, using the satellite radar inversion network to invert the input satellite cloud image sequence to obtain the radar echo image sequence X at the corresponding time; S3: Modeling the radar echo sequence time series projection, building a time series projection module Patch-KAN module based on multi-scale block technology and KAN network to capture its local spatiotemporal details and output the result Z; S4: Radar echo sequence motion trajectory modeling, using a causal attention, directly in the latent space to causally model cross-frame tokens, ensuring the unidirectionality of information flow, only looking at the past, not the future, meeting the requirements of autoregressive generation, and obtaining representation results S5: Radar echo sequence spatial texture modeling, combining the inversion feature X(t-1) of the previous time step with the current noise Concatenate in the channel dimension to form a new input tensor x′ t At the same time, X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj , S6: Multi-channel feature fusion, using Z as the key and value, and using the Cross-Attention mechanism to fuse with the original features of the denoising diffusion network; Directly inject into the bottleneck layer of denoising U-Net; F proj Multiple scale feature maps {f i } to perform conditional modulation on the new input noise tensor x′ t Denoising prediction results S7: Autoregression generates a complete video sequence and converts the prediction result of the current time step As the generation condition of the next time step, the video sequence with a complete T2 length is finally generated step by step 2. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The data preprocessing method in step S1 is: S11: Input data: Process 16 channels of satellite cloud image data with a duration of T1, including infrared and microwave channels; S12: Interpolation processing: interpolate the spatial resolution of all channels to 0.5km; S13: Channel selection: The Pearson correlation coefficient was calculated to analyze the correlation between the 16 channels and precipitation. Finally, channels 8, 10, 11, and 12, as well as the channel brightness temperature difference of 8-13 and the visible light channel difference of 3-5 were selected. S14: Normalization processing: The raw data are linearly normalized using the minimum-maximum method.

3. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The specific steps of step S2 are as follows: S21: Select the satellite radar inversion model as the related improved model based on U-Net, and output the inverted radar echo sequence X.

4. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The specific steps of step S3 are: S31: Predefine multiple Patch scales s, split according to spatial position, time length, and channel scale after Patch flattening, and obtain a Patch time series for each spatial position i Where T s represents the length of time at scale s, Patch vector at scale s, position i, and time step t S32: For each scale s and each position i, the Patch time series Splice along the time dimension to get the input vector x (s,i) : S33: Send the input vectors to the corresponding KAN model projection to model the nonlinear temporal dependency corresponding to the scale, and obtain the temporal feature encoding z at scale s (s,i) , the whole process formula is as follows: Among them, ψ q,j (·) is the input variable x j A learnable unary function that performs nonlinear transformations, φ q (·) is a function that further nonlinearly maps each set of weighted results, Q s Indicates the number of channels / basis functions used in this layer, P s Indicates the dimension of a single Patch at the corresponding scale, T s Indicates the length of time under scale s; S34: Encode the temporal features of all scales s to z (s,i) The unified spatiotemporal feature obtained by splicing in the spatial dimension i is represented as z i , the process is:

5. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The specific steps of step S4 are: S41: Given a radar echo sequence X of input historical frames, the latent space features of each frame are flattened into a token sequence in the temporal direction Merge to get the entire historical time series input S42: Using N-layer causal self-attention, including causal attention mechanism, feedforward network, residual connection, layer normalization, historical time series input After the linear connection layer and the causal self-attention mechanism, the residual connection and layer normalization are performed, and then the output vector passes through the feedforward fully connected network, and the residual connection and back-projection Linear operation are performed again. Will Perform linear transformation to obtain the query vector Q motion , key vector K motion Sum value vector V motion Three vectors: Among them, W Q Q motion The weight matrix, W Q K motion The weight matrix, W V V motion The weight matrix, Therefore, the causal attention mechanism is expressed as follows: Where M is the triangular causal mask, which ensures that the representation of the t-th frame can only access historical information of time steps ≤ t, and d is the vector Q motion , K motion The dimension is used to scale the dot product to ensure numerical stability. S43: The output of the causal attention structure is a dynamic weighted representation of the historical latent variables. The overall processing process is as follows: Reshape the token representation back to the spatial structure of the (t-1) frame and merge or aggregate it to generate the motion representation of the current frame 6. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The specific steps of step S5 are: S51: Combine the inversion feature X(t-1) of the previous time step with the current noise Concatenate in the channel dimension to form a new input tensor S52: X(t-1) is sent to an independent ControlNet branch, and local features are extracted through shallow convolution to obtain the control signal F proj .

7. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: The specific steps of step S6 are: S61: Use the result Z obtained in step S3 as the Key and Value, and obtain the query vector Q of Z through linear transformation Z , key vector K Z Sum value vector V Z , the Cross-Attention mechanism is used to output the new features by fusing the original features with the residual connection and LayerNorm The formula is as follows: S62: The result obtained in step S4 The features are directly injected into the bottleneck layer of U-Net and modeled as motion information; S63: F obtained in step S5 proj Control signal, feature map of multiple scales within U-Net {f i } to perform conditional modulation: Among them, γ i and β i By controlling the signal F proj The predicted scaling and translation parameters control the feature amplitude and distribution offset respectively. S64: Input tensor x′ formed in step S5 t , as the noise input of the denoising network, it injects coarse-grained spatial structure information from the source to improve the initial denoising effect; x′ t Under the combined effect of S61, S62, S63 and Stable Diffusion denoising Unet, the prediction results of a single frame are output.

8. The two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism according to claim 1 is characterized by: In step S7, the complete target sequence of the autoregressive formula is generated according to the following generation process: in is the predicted image of the t+i frame, represents the autoregressive context window for the current generation, p θ (·) represents the sampling distribution of the reverse diffusion process under the current conditions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism is implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, a two-stage short-term precipitation prediction diffusion model method based on satellite cloud images and autoregressive mechanism is implemented as described in any one of claims 1-8.

Images