A typhoon wind field downscaling method and system fusing terrain segmentation and diffusion generation model

Abstract

The present application relates to a typhoon wind field downscaling method and system integrating terrain segmentation and diffusion generation model, first, integrate typhoon wind field data, typhoon background information and terrain data, and carry out pretreatment, then use SAM model to automatically segment terrain data, calculate DEM average of terrain mask area, and modulate terrain embedded noise based on this, then process data through encoder and decoder, integrate Gate-ConvNeXt module, spatial self-attention layer, terrain cross-attention layer, fuse multi-scale information through jump connection, at the same time, embed typhoon background information through feedforward neural network; finally, the diffusion model adds Gaussian noise through forward diffusion, generates the residual between high-resolution wind field and low-resolution wind field through backward denoising, and combines it with the up-sampling result of low-resolution wind field to generate high-resolution wind field prediction; solve the problem of insufficient accuracy, real-time performance and adaptability of existing wind field downscaling method in processing complex terrain and typhoon weather conditions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of typhoon wind field reconstruction and simulation, and relates to a typhoon wind field downscaling method and system fusing terrain segmentation and diffusion generation model, in particular to a typhoon fine wind field downscaling method and system fusing terrain segmentation and diffusion generation model under the influence of mountainous terrain. BACKGROUND

[0002] With the intensification of global climate change, typhoons landing in China show significant characteristics such as high frequency, great intensity and strong disaster-causing, which seriously threaten the safety of people's lives and property in coastal areas. Typhoon near-surface wind field is a key factor determining its disaster-causing effect (such as strong wind, extreme precipitation and storm surge), and its fine degree directly affects the prediction accuracy and risk response efficiency. However, the fine prediction ability of typhoon near-surface wind field is limited by the insufficient resolution of wind field data. The adaptability of traditional dynamic models in complex terrain areas and the analytical ability of local wind field structure still have obvious shortcomings, although deep learning methods are expected to break through this limitation, but existing research mainly focuses on good wind background, which is difficult to adapt to the complex background of typhoon. And mostly use the L2 loss optimization based deterministic model, which is easy to cause smooth prediction.

[0003] Among them, Chinese invention patent CN118761026A discloses a near-surface wind field downscaling method based on deep learning, which considers the complex nonlinear relationship between wind and terrain and trains the network with meteorological station observations as true values, and can realize wind field downscaling under complex terrain conditions. However, the downscaling method of this patent is only applicable to good wind scenarios, and does not consider the wind speed difference under typhoon weather conditions, and cannot be applied to typhoon wind field downscaling.

[0004] Chinese invention patent CN118734725A discloses a complex micro-terrain influence correction method for typhoon parameterized wind field, belonging to the technical field of typhoon wind field simulation. This method obtains wind field data considering micro-terrain influence and wind field data not considering micro-terrain influence through numerical simulation model, and then constructs a micro-terrain correction model based on deep learning network and trains it. The correction model can capture the spatial features and coupling relationship between micro-terrain and wind field, and improve the accuracy of typhoon parameterized wind field data. However, the optimization L2 loss function used in training the deep learning model of this method still easily leads to the prediction result being too smooth, resulting in the loss of high-frequency details of the wind field.

[0005] Chinese invention patent CN118657082A discloses a method for simulating small-scale wind field under the influence of terrain, which relates to the field of wind field simulation. The method couples WRF model and CFD model to simulate and analyze the wind field affected by islands, realizes the fine depiction of the small-scale spatial distribution of wind field affected by terrain, and reveals the local variation characteristics of typhoon wind field under the influence of islands. However, although this method can finely depict the small-scale spatial distribution of wind field affected by terrain, it relies on the coupling of WRF model and CFD model, and has the problems of large amount of calculation, complex model and poor real-time performance.

[0006] Chinese invention patent CN118037058A discloses a method for predicting the risk of power transmission line based on statistical downscaling of typhoon wind field. The method calculates the error between the wind speed in the mesoscale typhoon forecast wind field and the measured wind speed based on the wind speed and wind direction data in the mesoscale typhoon forecast wind field and the measured wind speed of the meteorological station, and obtains the micro-topographic index of each meteorological station. The logarithmic error is taken as the dependent variable, and the wind speed in the mesoscale typhoon forecast wind field and the micro-topographic index of each meteorological station are taken as the independent variables. Linear regression model is used to fit the logarithmic error. Through the fitting model, the mesoscale typhoon forecast wind field at each power transmission tower is down-scaled to obtain the micro-scale typhoon forecast wind speed field. However, this method relies on linear regression model to fit the wind speed error, which may not fully capture the complex nonlinear relationship between wind field and micro-topography, resulting in insufficient downscaling accuracy and adaptability.

[0007] Currently, various methods have been proposed for typhoon wind field simulation and downscaling modeling in complex terrain areas, and some progress has been made. However, there are still the following shortcomings: First, although the existing deep learning-based wind field downscaling method (CN118761026A) can effectively model the nonlinear relationship between wind and terrain, and train with meteorological station observation data as true value, it is mainly suitable for stable and good weather conditions, and does not model the case of large changes in wind speed under extreme weather such as typhoon, resulting in insufficient accuracy and adaptability of downscaling in high-intensity typhoon wind field. Second, the research on typhoon wind field micro-terrain correction (CN118734725A) learns the spatial coupling characteristics between wind field and micro-terrain by deep learning method, which improves the fine modeling ability of wind field, but the L2 loss function used in the training process is easy to cause the prediction result to be smooth, and the local high-frequency change details of the wind field are lost, which affects the accuracy of the actual wind disaster assessment. The method of coupling numerical model and CFD model to simulate small-scale wind field changes (CN118657082A) can accurately depict the influence of islands, complex terrain on local wind field, but it relies on high-cost and large-scale numerical calculation as a whole, resulting in poor real-time performance, which is difficult to meet the needs of rapid forecasting and engineering application. In addition, the typhoon wind field downscaling method based on statistical regression (CN118037058A) mainly relies on linear regression model when correcting mesoscale wind field errors, which cannot fully express the complex nonlinear relationship between wind field and micro-terrain, resulting in limited adaptability and accuracy of the downscaling model in complex terrain conditions.

[0008] In summary, the existing typhoon wind field downscaling methods have the following problems: (1) The modeling ability of extreme wind speed characteristics of typhoon is insufficient, and it is difficult to accurately express the high wind speed change process; (2) The modeling ability of local high-frequency details of wind field is limited, and the prediction result is over-smoothed; (3) The calculation complexity is high, and it is difficult to balance the accuracy and real-time requirements; (4) The modeling method cannot fully describe the nonlinear relationship between terrain and wind field, and the model adaptability is limited. These shortcomings limit the application effect of existing technologies in the rapid and fine simulation of typhoon wind field in complex terrain areas. SUMMARY

[0009] Therefore, the present application provides a typhoon wind field downscaling method and system fusing terrain segmentation and diffusion generation model to solve the problem that the existing wind field downscaling methods have insufficient accuracy, real-time performance and adaptability in dealing with complex terrain and typhoon weather conditions, affecting their application effect in the rapid and fine simulation of typhoon wind field in complex terrain areas.

[0010] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0011] A typhoon wind field downscaling method fusing terrain segmentation and diffusion generation model, comprising the following steps:

[0012] S1, determine the data source and pretreatment; determine the source of low resolution typhoon wind field data, high resolution typhoon wind field data, typhoon background information and high resolution terrain data used for input typhoon wind field downscaling model, and pretreat the obtained raw data;

[0013] S2, modulate the terrain embedded noise; the preprocessed high resolution terrain data is automatically segmented using a segmentation all model SAM, the DEM mean value of the region is calculated based on the terrain mask, and the terrain embedded noise modulation is carried out based on the mean value;

[0014] S3, construct a denoising network based on U-Net architecture; the network is composed of an encoder and a decoder, the encoder contains 4 down-sampling blocks, and the decoder contains 4 up-sampling blocks; the down-sampling block and the up-sampling block are integrated with Gate-ConvNeXt module, spatial self-attention layer, terrain cross-attention layer and up-sampling layer; the encoder is responsible for compressing the input data into a compact latent space representation, and the decoder restores it to the original spatial distribution; the skip connection mechanism is introduced between the encoder and the decoder to realize the fusion of low-level details and high-level semantics, and the typhoon background information is embedded into each Gate-ConvNeXt module after projection by the feedforward neural network;

[0015] S4, construct a diffusion model; the diffusion model sampling process includes forward diffusion process and reverse denoising process, the forward diffusion process of the diffusion model is a Markov process, given the original data distribution x0~q(x0) in T time steps, the standard Gaussian noise is gradually added, so that the original data distribution is gradually converted into a standard Gaussian distribution;

[0016] S5, train the denoising model; in the training process of the diffusion model, the model parameters are optimized by using the back propagation algorithm, the L2 loss between the model predicted noise and the real noise is minimized, and the effective learning of the diffusion model is realized;

[0017] S6, wind field generation stage; sample random noise from the standard Gaussian distribution in step S4 and combine the target terrain embedded features as the starting point of reverse denoising; use the trained diffusion model and control conditions to gradually denoise by reverse diffusion, generate W HR_Norm and the residual between W LR_Up ; combine the residual with W LR_Up to generate the final high resolution wind field prediction.

[0018] Further, the processing of the raw data in step S1 specifically includes: missing value filling, data standardization, low resolution data up-sampling and category variable encoding.

[0019] Further, the typhoon background information in step S1 includes typhoon center longitude Lon, center latitude Lat, intensity I, near-center minimum pressure PRES, and near-center 2-minute average maximum sustained wind speed WND, and the high-resolution terrain data includes digital elevation model DEM, terrain slope Slope, and slope aspect Aspect; the preprocessed low-resolution typhoon wind field data W LR_Norm , the sampled low-resolution typhoon wind field data W LR_Up , the high-resolution wind field data W HR_Norm , the typhoon background information TC Norm , the high-resolution terrain data Topo HR_Norm .

[0020] Further, the automatic segmentation process in step S2 is as follows:

[0021] [Mask1, Mask2, …, Mask K ] = SAM(Topo HR_Norm )

[0022] wherein Mask i (1≤i≤k) represents the i-th mask obtained by SAM segmentation.

[0023] The mean value calculation process is as follows:

[0024]

[0025] wherein A i represents the region corresponding to the i-th mask, μ(A i ) represents the DEM mean value of the region, |A i | represents the number of data points in the region A i , and DEM x represents the elevation value of each data point x in the region A i .

[0026] The terrain embedding noise modulation process is as follows:

[0027] E Topo = Concat(μ(A1), μ(A2), …, μ(A K ))

[0028] ∈' = ∈ + E Topo

[0029] wherein E Topo represents terrain embedding, Concat(·) represents concatenation operation, ∈ represents standard Gaussian noise, and ∈' represents terrain embedding noise.

[0030] Further, the Gate-ConvNeXt module calculation formula in step S3 is as follows:

[0031] x Context ,x Gate =Chunk(LN(Conv(x) GC )))

[0032] F Context =GRN(GELU(x) Context ))

[0033] F Gate =SiLU(Conv(x) Gate ))

[0034] O GC =Conv(F Context ⊙F Gate )+x GC

[0035] Where, x GC x represents the input of the Gate-ConvNeXt module. Context Represents preprocessing context features, x Gate F represents the preprocessing gating feature. Context F represents contextual features. Gate Represents gating characteristics, O GC represents the output of the Gate-ConvNeXt module, LN represents the layer normalization layer, Chunk(·) represents the feature separation operation, GRN represents the global response normalization layer, SiLU represents the SiLU activation layer, and ⊙ represents the dot product operation.

[0036] The formula for embedding typhoon background information is as follows:

[0037] E Lonlat =Linear(SiLU(Linear(Lon Norm ,Lat Norm )))

[0038] E Tp =Linear(SiLU(Linear(I Norm PRES Norm WND Norm )))

[0039] O GC_Embed =O GC +E Lonlat +E Tp

[0040] Among them, E Lonlat E represents the embedding of the typhoon center's latitude and longitude. Tp Represents embedded typhoon background information, O GC_EmbedThe output of the Gate-ConvNeXt module after adding the embedded information is represented by O

[0041] The calculation formula of the spatial self-attention layer is as follows:

[0042] Q, K, V = Conv(O GC_Embed )

[0043]

[0044] O Spa_Attn = O GC_Embed + Conv(F Spa_Attn )

[0045] wherein F Spa_Attn represents the spatial attention feature, and O Spa_Attn represents the output of the spatial attention layer;

[0046] The calculation formula of the terrain cross-attention layer is as follows:

[0047] Q = Conv(O Spa_Attn )

[0048] K, V = Conv(Topo HR_Norm )

[0049]

[0050] O Topo_Attn = O Spa_Attn + Conv(F Topo_Attn )

[0051] wherein Q represents the query matrix, K represents the key matrix, V represents the value matrix, x Topo_Attn represents the input of the terrain cross-attention layer, softmax represents the softmax activation layer, d represents the dimension number of the query and the key, F Topo_Attn represents the terrain cross-attention feature, and O Topo_Attn represents the output of the terrain cross-attention layer.

[0052] Further, the original data in step S4 is defined as the residual between the high-resolution wind field data W HR_Norm and the sampled low-resolution typhoon wind field data W LR_Up The target generated by the diffusion model is the residual between W HT_Norm and W LR_Up The mean and variance of the standard Gaussian noise added by the traditional diffusion model at each time step t (0≤t≤T) are determined by the hyperparameter β t , and the sample x t after adding the noise is generated accordingly; the terrain embedded noise generated in step S2 is used to replace the standard Gaussian noise, and the process of the forward diffusion process is represented as:

[0053]

[0054] Where I represents the identity matrix. Conversely, the reverse denoising process gradually restores the noise samples to the target wind field, and this process can be represented as:

[0055]

[0056] Where c represents the diffusion control conditions (in this invention, these include typhoon background information and high-resolution terrain data). The denoising network estimates the mean function predicted by the neural network, and its inputs are the noise sample at the current time and the diffusion control conditions. This is the noise variance hyperparameter.

[0057] Furthermore, the calculation process of the loss function in step S5 is as follows:

[0058]

[0059] Where L represents the loss function, ∩ represents the true noise, and ∩ θ (x t ,x0,E Topo c) represents model prediction noise. To enhance the model's ability to model high-frequency details of the wind field, a high-frequency loss attention mechanism is introduced to weight and adjust the loss function; the calculation process of the weighted loss is shown below:

[0060] Weight = LN(DWT(W) HR_Norm ))

[0061] L Final =Weight⊙L

[0062] Where DWT(·) represents Discrete Wavelet Transform, Weight represents the loss weight, and L Final This represents the final loss.

[0063] This typhoon wind field downscaling system, which integrates terrain segmentation and diffusion generation models, includes:

[0064] Data preprocessing module: Integrates low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution topographic data to complete missing value imputation, standardization, upsampling, and coding preprocessing;

[0065] Terrain embedding noise modulation module: The preprocessed high-resolution terrain data is automatically segmented using the Segmentation Model (SAM), the mean DEM value of the terrain mask region is calculated, and the terrain embedding noise is modulated based on this value.

[0066] A denoising model based on the U-Net architecture is constructed: the encoder compresses the input data into a latent space representation through 4 downsampling blocks, and the decoder restores it to the original spatial distribution through 4 upsampling blocks; the network integrates Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up-down sampling layers, and fuses multi-scale information through a skip connection; the typhoon background information is embedded through a feedforward neural network;

[0067] Diffusion model: the original data is gradually added Gaussian noise through a forward diffusion process to convert it into a standard Gaussian distribution; the reverse denoising process gradually removes noise to generate a high-resolution wind field W HR_Nrom and the residual of the low-resolution wind field W LR_Up interpolation result;

[0068] Wind field generation module: starting with standard Gaussian noise and terrain embedding features, combining the trained diffusion model, generating residuals and combining the low-resolution wind field up-sampling result, finally generating a high-resolution wind field prediction.

[0069] An electronic device comprising: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the typhoon wind field downscaling method of the fusion terrain segmentation and diffusion generation model described above.

[0070] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the typhoon wind field downscaling method of the fusion terrain segmentation and diffusion generation model described above.

[0071] The beneficial effects of the present application are:

[0072] 1. In the prior art, some typhoon wind field downscaling methods rely on simplified assumptions or smoothing models, which cannot accurately capture the high-frequency details of the wind field, resulting in insufficient downscaling accuracy, and complex calculations and poor real-time performance. To solve this problem, the present application uses a diffusion model and combines the fine-grained segmentation capability of the segmentation everything large model to guide the model to generate a wind field highly correlated with local terrain distribution. In addition, a wind field high-frequency loss attention mechanism is proposed to guide the model to enhance its modeling capability for high-frequency components of the wind field, effectively preserving small-scale change characteristics and improving the deficiency of the current model output being too smooth. The present application can more accurately simulate the influence of complex terrain on typhoon wind field, has the characteristics of high precision, high operation efficiency and high adaptability, and is suitable for various scenarios such as weather forecasting and disaster warning.

[0073] 2. The typhoon wind field downscaling method disclosed in this invention, which integrates terrain segmentation and diffusion generation models, accurately generates wind fields highly correlated with local terrain by introducing the diffusion model and the fine-grained segmentation capability of the Segmentation All-Around Model (SAM). This significantly improves downscaling accuracy, captures high-frequency details of the wind field, and overcomes the shortcomings of existing technologies that rely on simplified assumptions or smoothed models. Simultaneously, the proposed high-frequency loss attention mechanism effectively guides the model to enhance its ability to model high-frequency components, avoiding overly smooth outputs, preserving small-scale variation characteristics, and improving the simulation accuracy of the impact of complex terrain. Furthermore, this invention boasts high computational efficiency and strong real-time performance, making it suitable for scenarios requiring rapid response, such as weather forecasting and disaster early warning. It also has broad applicability, exhibiting stronger adaptability under extreme weather conditions such as typhoons, and can be used for theoretical research as well as playing a crucial role in practical meteorological monitoring and disaster early warning.

[0074] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0075] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0076] Figure 1 This is a flowchart of the typhoon wind field downscaling method that integrates terrain segmentation and diffusion generation models according to the present invention.

[0077] Figure 2 This is a flowchart of the terrain embedding noise modulation module of the present invention;

[0078] Figure 3 The flowchart illustrates the construction of a denoising model based on the U-Net architecture for this invention. Detailed Implementation

[0079] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0080] like Figure 1 The typhoon wind field downscaling method shown, which integrates terrain segmentation and diffusion generation models, includes the following steps:

[0081] S1, determine data sources and pre-process. Determine the sources of low-resolution typhoon wind field data, high-resolution wind field data, typhoon background information (including typhoon center longitude Lon, center latitude Lat, intensity I, near center minimum pressure PRES, near center 2-minute average maximum sustained wind speed WND) and high-resolution terrain data (including digital elevation model DEM, terrain slope Slope, slope aspect Aspect) used for input into the typhoon wind field downscaling model.

[0082] Pre-process the obtained raw data, including missing value filling, data standardization, low-resolution data upsampling and category variable encoding. Obtain pre-processed low-resolution typhoon wind field data W LR_Norm , upsampled low-resolution typhoon wind field data W LR_Up , high-resolution wind field data W HR_Norm , typhoon background information TC Norm , high-resolution terrain data Topo HR_Norm .

[0083] S2, modulate terrain embedding noise. As shown in the figure, this module first uses the Segment Anything Model (SAM) to automatically segment the pre-processed high-resolution terrain data, then calculates the DEM mean value of the region based on the terrain mask, and modulates the terrain embedding noise based on the mean value. Figure 2

[0084] The automatic segmentation process is as follows:

[0085] [Mask1,Mask2,…,Mask K ]=SAM(Topo HR_Norm )

[0086] Where Mask i (1≤i≤K) represents the i-th mask obtained by SAM segmentation.

[0087] The mean value calculation process is as follows:

[0088]

[0089] Where A i represents the region corresponding to the i-th mask, μ(A i ) represents the DEM mean value of the region, |A i | represents the number of data points in region A i , DEM x represents the elevation value of each data point x in region A i .

[0090] The terrain embedding noise modulation process is as follows: ​

[0091] E Topo =Concat(μ(A1),μ(A2),…,μ(A K ))

[0092] ∈'=∈+E Topo

[0093] Among them, E Topo Represents terrain embedding, Concat(·) represents the concatenation operation, ∈ represents standard Gaussian noise, and ∈' represents terrain embedding noise.

[0094] S3. Construct a denoising network based on the U-Net architecture. For example... Figure 3 The network shown consists of an encoder and a decoder. The encoder contains four downsampling blocks, and the decoder contains four upsampling blocks. Both the downsampling and upsampling blocks integrate a Gate-ConvNeXt module, a spatial self-attention layer, a terrain cross-attention layer, and upsampling / downsampling layers. The encoder is responsible for compressing the input data into a compact latent spatial representation, while the decoder restores it to the original spatial distribution. A skip connection mechanism is introduced between the encoder and decoder to achieve the fusion of low-level details and high-level semantics. In addition, typhoon background information is projected into each Gate-ConvNeXt module through a feedforward neural network.

[0095] The calculation formula for the Gate-ConvNeXt module is as follows:

[0096] x Context ,x Gate =Chunk(LN(Conv(x) GC )))

[0097] F Context =GRN(GELU(x) Context ))

[0098] F Gate =SiLU(Conv(x) Gate ))

[0099] Q GC =Conv(F Context ⊙F Gate )+x GC

[0100] Where, x GC x represents the input of the Gate-ConvNeXt module. Context Represents preprocessing context features, x Gate F represents the preprocessing gating feature. Context F represents contextual features. Gate Represents gating characteristics, O GCThe Gate-ConvNeXt module output is represented, LN represents a layer normalization layer, Chunk(·) represents a feature separation operation, GRN represents a global response normalization layer, SiLU represents a SiLU activation layer, and represents a dot multiplication operation.

[0101] The typhoon background information embedding calculation formula is as follows:

[0102] E Lonlat = Linear(SiLU(Linear(Lon Norm ,Lat Norm )))

[0103] E Tp = LineR(SiLU(Linear(I Norm ,PRES Norm ,WND Norm )))

[0104] O GC_Embed = O GC +E Lonlat +E Tp

[0105] Wherein, E Lonlat represents the typhoon center longitude and latitude embedding, E Tp represents the typhoon background information embedding, and O GC_Embed represents the Gate-ConvNeXt module output after adding the embedding information.

[0106] The spatial self-attention layer calculation formula is as follows:

[0107] Q, K, V = Conv(O GC_Embed )

[0108]

[0109] O Spa_Attn = O GC_Embed + Conv(F Spa_Attn )

[0110] Wherein, F Spa_Attn represents the spatial attention feature, and Q Spa_Attn represents the spatial attention layer output.

[0111] The terrain cross-attention layer calculation formula is as follows:

[0112] Q = Conv(Q Spa_Attn )

[0113] K, V = Conv(Topo HR_Norm )

[0114]

[0115] O Topo_Attn =O Spa_Attn +Conv(F Topo_Attn )

[0116] where Q represents a query matrix, K represents a key matrix, V represents a value matrix, x Topo_Attn represents a terrain cross-attention layer input, softmax represents a softmax activation layer, d represents a dimension number of a query and a key, F Topo_Attn represents a terrain cross-attention feature, O Topo_Attn represents a terrain cross-attention layer output.

[0117] S4, constructing a diffusion model. The diffusion model sampling process includes a forward diffusion process and a backward denoising process. The diffusion model forward diffusion process is a Markov process, and given the original data distribution x0~q(x0), the standard Gaussian noise is gradually added in T time steps, so that the original data distribution is gradually converted into a standard Gaussian distribution.

[0118] In the present application, the original data is defined as the high-resolution wind field data W HR_Norm and the residual error between the low-resolution typhoon wind field data W LR_Up after sampling, so the target generated by the diffusion model in the present application is the residual error between W HR_Norm and W LR_Up . The mean and variance of the standard Gaussian noise added by the traditional diffusion model at each time step t (0≤t≤T) are determined by the hyperparameter β t , and the sample x t after adding noise is generated accordingly. The terrain embedding noise generated in step S2 is used to replace the standard Gaussian noise in the present application, and the process of the forward diffusion process is represented as:

[0119]

[0120] where I represents an identity matrix. On the contrary, the backward denoising process gradually restores the noise sample to the target wind field, and its process can be represented as:

[0121]

[0122] where c is a diffusion control condition (including typhoon background information and high-resolution terrain data in the present application), represents a mean function estimated by a denoising network as a neural network prediction, and the input of the denoising network is the current time noise sample and the diffusion control condition, is a noise variance hyperparameter.

[0123] S5, training the denoising model. During the training process, the model parameters are optimized using the backpropagation algorithm, and the effective learning of the diffusion model is achieved by minimizing the L2 loss between the model predicted noise and the true noise. The calculation process of the loss function is as follows:

[0124]

[0125] Wherein, L represents the loss function, ∈ represents the true noise, ∈ θ (x t ,x0,E Topo ,c) represents the model predicted noise. In order to enhance the modeling ability of the model to the high frequency details of the wind field, the high frequency loss attention mechanism of the wind field is introduced, and the loss function is weighted and adjusted. The calculation process of the weighted loss is as follows:

[0126] Weight=LN(DWT(W HR_Norm ))

[0127] L Final =Weight⊙L

[0128] Wherein, DWT(·) represents the discrete wavelet transform, Weight represents the loss weight, L Final represents the final loss.

[0129] S6, wind field generation stage. This step first samples a random noise with the same dimension as the target wind field from the standard Gaussian distribution in step S4 and combines the embedding features of the target terrain area as the starting point of the reverse denoising process. Then by using the trained diffusion model and the diffusion control condition, the residual between W HR_Norm and W LR_Up is generated by the reverse diffusion process. Finally, the residual and W LR_Up are combined to generate the final prediction, realizing the generation of high resolution near surface wind field.

[0130] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the purpose and scope of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A method for downscaling typhoon wind fields by integrating terrain segmentation and diffusion generation models, characterized in that, Includes the following steps: S1. Determine the data sources and preprocess them; Determine the sources of low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution topographic data used as input to the typhoon wind field downscaling model, and preprocess the obtained raw data. S2, Modulated terrain embedding noise; The preprocessed high-resolution terrain data is automatically segmented using the Segmentation Model (SAM). The mean DEM value of the region is calculated based on the terrain mask, and terrain embedding noise modulation is performed based on this mean value. S3. Construct a denoising network based on the U-Net architecture. This network consists of an encoder and a decoder. The encoder contains 4 downsampling blocks, and the decoder contains 4 upsampling blocks. Both the downsampling and upsampling blocks integrate Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and upsampling and downsampling layers. The encoder is responsible for compressing the input data into a compact latent spatial representation, while the decoder restores it to the original spatial distribution. A skip connection mechanism is introduced between the encoder and decoder to achieve the fusion of low-level details and high-level semantics. Typhoon background information is projected into each Gate-ConvNeXt module through a feedforward neural network. S4. Constructing the diffusion model; the diffusion model sampling process includes a forward diffusion process and a reverse denoising process. The forward diffusion process of the diffusion model is a Markov process, given the original data distribution. exist In each time step, standard Gaussian noise is gradually added, thereby gradually transforming the original data distribution into a standard Gaussian distribution; S5. Training the denoising model; During the training process, the diffusion model uses the backpropagation algorithm to optimize the model parameters, and achieves effective learning of the diffusion model by minimizing the L2 loss between the model's predicted noise and the actual noise. S6, Wind Field Generation Stage: Random noise is sampled from the standard Gaussian distribution in step S4 and combined with the target terrain embedding features as the starting point for reverse denoising; using the trained diffusion model and control conditions, denoising is gradually performed through reverse diffusion to generate... and The residuals between; the residuals and Combined, a final high-resolution wind field prediction is generated.

2. The typhoon wind field downscaling method as described in claim 1, characterized in that, The processing of raw data in step S1 specifically includes: missing value imputation, data standardization, low-resolution data upsampling, and categorical variable encoding.

3. The typhoon wind field downscaling method as described in claim 1, characterized in that, The typhoon background information in step S1 includes the longitude of the typhoon center. Central latitude ,strength Lowest air pressure near the center 2-minute average maximum sustained wind speed near the center High-resolution terrain data includes digital elevation models. terrain slope slope direction Obtain preprocessed low-resolution typhoon wind field data. Sampling low-resolution typhoon wind field data High-resolution wind field data Typhoon background information High-resolution terrain data .

4. The typhoon wind field downscaling method as described in claim 1, characterized in that, The automatic segmentation process in step S2 is as follows: in, The first segment obtained from SAM segmentation A mask; The mean calculation process is as follows: in, This represents the region corresponding to the i-th mask. This represents the mean DEM value for the region. Representative area The number of data points in the data. Representative area Each data point Elevation value; The terrain embedding noise modulation process is as follows: in, Represents terrain embedding, This represents a splicing operation. Represents standard Gaussian noise. This represents terrain embedding noise.

5. The typhoon wind field downscaling method as described in claim 4, characterized in that, The calculation formula for the Gate-ConvNeXt module in step S3 is as follows: in, Represents the Gate-ConvNeXt module input. Represents preprocessing context features, Represents preprocessing gating features, Represents contextual features, Represents gating characteristics. Represents the output of the Gate-ConvNeXt module. Representative layer normalization layer, Representative feature separation operation, Represents the global response normalization layer. represent Activation layer, Represents the dot product operation; The formula for embedding typhoon background information is as follows: in, Represents the embedding of the typhoon center's latitude and longitude. Represents embedded typhoon background information. This represents the output of the Gate-ConvNeXt module after adding embedded information; The formula for calculating the spatial self-attention layer is as follows: in, Represents spatial attention characteristics. Represents the output of the spatial attention layer; The formula for calculating the terrain cross-attention layer is as follows: in, Represents the query matrix. V represents the key matrix, and V represents the value matrix. The input represents the terrain-crossing attention layer. represent Activation layer, The number of dimensions representing the query and the key. Represents terrain cross-attention features. This represents the output of the terrain cross-attention layer.

6. The typhoon wind field downscaling method as described in claim 5, characterized in that, In step S4, the raw data is defined as high-resolution wind field data. Compared with sampled low-resolution typhoon wind field data The residuals between them, the objective of the diffusion model is to generate the diffusion model. and The residuals between time steps; traditional diffusion model at each time step The mean and variance of the added standard Gaussian noise are determined by the hyperparameters. The decision is made, and samples with added noise are generated accordingly. The forward diffusion process, which replaces the standard Gaussian noise with the terrain embedding noise generated in step S2, is represented as follows: in, Representing the identity matrix, conversely, the reverse denoising process gradually restores the noise samples to the target wind field, and this process can be represented as: in, For diffusion control conditions, The denoising network estimates the mean function predicted by the neural network, and its inputs are the noise sample at the current time and the diffusion control conditions. This is the noise variance hyperparameter.

7. The typhoon wind field downscaling method as described in claim 6, characterized in that, The calculation process of the loss function in step S5 is as follows: in, Represents the loss function. Represents real noise. The model predicts noise; to enhance the model's ability to model high-frequency details of the wind field, a high-frequency loss attention mechanism is introduced to weight and adjust the loss function; the calculation process of the weighted loss is shown below: in, Discrete wavelet transform, Represents the loss weight. This represents the final loss.

8. A system using a typhoon wind field downscaling method that integrates terrain segmentation and diffusion generation models as described in claim 1, characterized in that, include: Data preprocessing module: Integrates low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution topographic data to complete missing value imputation, standardization, upsampling, and coding preprocessing; Terrain embedding noise modulation module: The preprocessed high-resolution terrain data is automatically segmented using the Segmentation Model (SAM), the mean DEM value of the terrain mask region is calculated, and the terrain embedding noise is modulated based on this value. A denoising model based on the U-Net architecture is constructed: the encoder compresses the input data into a latent spatial representation through four downsampling blocks, and the decoder restores it to the original spatial distribution through four upsampling blocks; the network integrates the Gate-ConvNeXt module, spatial self-attention layer, terrain cross-attention layer, and upsampling and downsampling layers, and fuses multi-scale information through skip connections, while typhoon background information is embedded through a feedforward neural network; Diffusion model: Gaussian noise is gradually added to the original data through a forward diffusion process to transform it into a standard Gaussian distribution; the reverse denoising process gradually removes the noise to generate a high-resolution wind field. With low-resolution wind field The residuals of the sampling results; Wind field generation module: Starting with standard Gaussian noise and terrain embedding features, combined with a trained diffusion model, it generates residuals and combines them with low-resolution wind field upsampling results to finally generate high-resolution wind field predictions.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute a computer program to implement the steps of the typhoon wind field downscaling method according to any one of claims 1 to 7, which integrates terrain segmentation and diffusion generation models.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, is used to implement the steps of the typhoon wind field downscaling method according to any one of claims 1 to 7, which integrates terrain segmentation and diffusion generation models.

Images