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

Abstract

The invention relates to a typhoon wind field downscaling method and system fusing terrain segmentation and a diffusion generation model, and the method comprises the steps: integrating typhoon wind field data, typhoon background information and terrain data, carrying out the preprocessing, carrying out the automatic segmentation of the terrain data through an SAM model, calculating a DEM mean value of a terrain mask region, and carrying out the downscaling of a typhoon wind field. The method comprises the following steps of: firstly, modulating terrain embedding noise on the basis of the terrain embedding noise, then processing data through an encoder and a decoder, integrating a Gate-ConvNeXt module, a spatial self-attention layer and a terrain cross attention layer, fusing multi-scale information through jump connection, and embedding typhoon background information through a feedforward neural network; finally, Gaussian noise is added to the diffusion model through forward diffusion, a residual error between the high-resolution wind field and the low-resolution wind field is generated through reverse denoising, the residual error is combined with a sampling result of the low-resolution wind field, and high-resolution wind field prediction is generated. The problem that an existing wind field downscaling method is insufficient in precision, real-time performance and adaptability under the condition of processing complex terrains and typhoon weather is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of typhoon wind field reconstruction and simulation, and relates to a typhoon wind field downscaling method and system that integrates terrain segmentation and diffusion generation models, and in particular to a typhoon fine wind field downscaling method and system that integrates terrain segmentation and diffusion generation models under the influence of mountainous terrain. Background Art

[0002] As global climate change intensifies, typhoons making landfall in my country have shown significant characteristics such as high frequency, high intensity, and strong disaster-causing potential, seriously threatening the safety of life and property in coastal areas. The near-surface wind field of a typhoon is a key factor in determining its disaster-causing effects (such as strong winds, extreme precipitation, and storm surges), and its precision directly affects the accuracy of forecasts and the efficiency of risk response. However, the current ability to accurately forecast typhoon near-surface wind fields is limited by the insufficient resolution of wind field data. Traditional dynamic models still have obvious deficiencies in adaptability to complex terrain areas and the ability to analyze local wind field structures. Although deep learning methods are expected to break through this limitation, existing research mainly focuses on benign wind backgrounds and is difficult to adapt to the complex background of typhoons. In addition, most of the deterministic models based on L2 loss optimization are used, which easily leads to smooth predictions.

[0003] Among them, Chinese invention patent CN118761026A discloses a near-surface wind field downscaling method based on deep learning. This method considers the complex nonlinear relationship between wind and terrain and uses meteorological station observations as ground truth to train the network. This method can achieve wind field downscaling under complex terrain conditions. However, this patent's downscaling method is only applicable to benign wind scenarios and does not consider wind speed differences in typhoon weather conditions, making it unsuitable for typhoon wind field downscaling.

[0004] Chinese invention patent CN118734725A discloses a method for correcting the complex micro-topography effects on typhoon parameterized wind fields, which belongs to the field of typhoon wind field simulation technology. This method uses a numerical simulation model to obtain wind field data that takes into account the influence of micro-topography and wind field data that does not take into account the influence of micro-topography, and then constructs a micro-topography correction model based on a deep learning network and trains it. This correction model can capture the spatial characteristics and coupling relationship between micro-topography and wind field, and improve the accuracy of typhoon parameterized wind field data. However, the optimized L2 loss function used in this method when training the deep learning model still easily leads to overly smooth prediction results, 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 fields under the influence of terrain, which relates to the field of wind field simulation technology. This method couples the WRF model and the CFD model to perform downscaling simulation and analysis of wind fields affected by islands and reefs, achieving a detailed depiction of the small-scale spatial distribution of wind fields affected by terrain, and revealing the local variation characteristics of typhoon wind fields under the influence of islands and reefs. However, although this method can accurately depict the small-scale spatial distribution of wind fields affected by terrain, it relies on the coupling of the WRF model and the CFD model, and has problems such as large computational complexity, complex model, and poor real-time performance.

[0006] Chinese invention patent CN118037058A discloses a transmission line risk prediction method based on typhoon wind field statistical downscaling. This method is based on the wind speed and wind direction data in the mesoscale typhoon forecast wind field and the measured wind speed at the meteorological station, calculates the error between the wind speed in the mesoscale typhoon forecast wind field and the measured wind speed, and simultaneously obtains the micro-topography index of each meteorological station. With the logarithmic error as the dependent variable and the wind speed in the mesoscale typhoon forecast wind field and the micro-topography index of each meteorological station as the independent variables, a linear regression model is used to fit the logarithmic error. Through this fitting model, the mesoscale typhoon forecast wind field at each transmission tower is downscaled to obtain the micro-scale typhoon forecast wind speed field. However, this method relies on the linear regression model to fit the wind speed error, which may not fully capture the complex nonlinear relationship between the wind field and the 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, the following shortcomings remain: First, while existing deep learning-based wind field downscaling methods (CN118761026A) can effectively model the nonlinear relationship between wind and terrain and use meteorological station observation data as ground truth for training, they are primarily applicable to stable and benign meteorological conditions and do not model the large variations in wind speed during extreme weather conditions such as typhoons. This results in insufficient downscaling accuracy and adaptability in high-intensity typhoon wind fields. Second, research on typhoon wind field micro-topography correction (CN118734725A) uses deep learning methods to learn the spatial coupling characteristics between wind fields and micro-topography, improving the ability to fine-tune wind field modeling. However, the L2 loss function used in the training process can easily lead to smooth prediction results and lose details of local high-frequency changes in the wind field, affecting the accuracy of actual wind disaster assessments. While the method of simulating small-scale wind field variations by coupling numerical models with CFD models (CN118657082A) can accurately depict the impact of islands, reefs, and complex terrain on local wind fields, it relies on high-cost, large-scale numerical calculations, resulting in poor real-time performance and difficulty meeting the needs of rapid forecasting and engineering applications. Furthermore, the typhoon wind field downscaling method based on statistical regression (CN118037058A) primarily relies on linear regression models to correct for mesoscale wind field errors. This method cannot fully express the complex nonlinear relationship between wind fields and microtopography, limiting the universality and accuracy of the downscaling model under complex terrain conditions.

[0008] In summary, existing typhoon wind field downscaling methods generally have the following problems: (1) insufficient adaptive modeling capabilities for typhoon extreme wind speed characteristics, making it difficult to accurately express the process of drastic changes in high wind speed; (2) limited modeling capabilities for local high-frequency details of the wind field, resulting in oversmoothing of the prediction results; (3) high computational complexity, making it difficult to balance accuracy and real-time requirements; (4) insufficient modeling of the nonlinear relationship between terrain and wind field, resulting in limited model adaptive capabilities. These deficiencies limit the application of existing technologies in the rapid and detailed simulation of typhoon wind fields in complex terrain areas. Summary of the Invention

[0009] In view of this, in order to solve the problem that the existing wind field downscaling methods lack accuracy, real-timeness and adaptability when dealing with complex terrain and typhoon weather conditions, which affects their application effect in the rapid and precise simulation of typhoon wind fields in complex terrain areas, the present invention provides a typhoon wind field downscaling method and system that integrates terrain segmentation and diffusion generation model.

[0010] In order to achieve the above object, the present invention provides the following technical solutions:

[0011] A typhoon wind field downscaling method integrating terrain segmentation and diffusion generation model includes the following steps:

[0012] S1. Determine the data source and preprocess it; determine the sources of low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution terrain data used to input the typhoon wind field downscaling model, and preprocess the obtained raw data;

[0013] S2. Modulating the terrain embedding noise: Automatically segment the pre-processed high-resolution terrain data using the Segmentation Everything Model (SAM), calculate the DEM mean of the area based on the terrain mask, and modulate the terrain embedding noise based on the mean;

[0014] S3. Construct a denoising network based on the U-Net architecture. This network 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 Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up- and downsampling layers. The encoder is responsible for compressing the input data into a compact latent space representation, while the decoder restores it to its 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 embedded in each Gate-ConvNeXt module after projection through a feedforward neural network.

[0015] S4. Construct a 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 x0~q(x0), standard Gaussian noise is gradually added in T time steps, thereby gradually converting the original data distribution into a standard Gaussian distribution.

[0016] S5. Train the denoising model. During the training process, the diffusion model uses the back propagation algorithm to optimize the model parameters. By minimizing the L2 loss between the model's predicted noise and the actual noise, the diffusion model is effectively learned.

[0017] S6, wind field generation stage; sampling random noise from the standard Gaussian distribution in step S4 and combining it with the target terrain embedding features as the starting point for reverse denoising; using the trained diffusion model and control conditions, gradually denoising through reverse diffusion to generate W HR_Norm With W LR_Up The residual between LR_Up Combined, the final high-resolution wind field forecast is generated.

[0018] Furthermore, the processing of the original data in step S1 specifically includes: missing value filling, data standardization, low-resolution data upsampling and categorical variable encoding.

[0019] Furthermore, in step S1, the typhoon background information includes the typhoon center longitude Lon, center latitude Lat, intensity I, the lowest pressure near the center PRES, and the 2-minute average maximum sustained wind speed near the center WND. The high-resolution terrain data includes the digital elevation model DEM, terrain slope Slope, and slope aspect Aspect. The pre-processed low-resolution typhoon wind field data W is obtained. LR_Norm , low-resolution typhoon wind field data W after sampling LR_Up , high-resolution wind field data W HR_Norm , Typhoon Background Information TC Norm , high-resolution terrain data Topo HR_Norm .

[0020] Furthermore, the automatic segmentation process in step S2 is:

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

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

[0023] The mean calculation process is:

[0024]

[0025] Among them, A i Represents the area corresponding to the i-th mask, μ(A i ) represents the DEM mean of the area, |A i |Represents Region A i The number of data points in the DEM x Representing Region A i The elevation value of each data point x in .

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

[0030] Furthermore, the calculation formula of the Gate-ConvNeXt module 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] Among them, x GC Represents the Gate-ConvNeXt module input, x Context represents the preprocessing context features, x Gate represents the preprocessing gate feature, F Context represents context features, F Gate represents the gated feature, 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 calculation formula for typhoon background information embedding 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 represents the longitude and latitude of the typhoon center, E Tp Represents the typhoon background information embedding, O GC_EmbedRepresents the output of the Gate-ConvNeXt module after adding the embedding information;

[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] Among them, F Spa_Attn represents the spatial attention feature, O Spa_Attn Represents the output of the spatial attention layer;

[0046] The terrain cross attention layer calculation formula 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] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and x Topo_Attn represents the terrain cross attention layer input, softmax represents the softmax activation layer, d represents the number of dimensions of query and key, F Topo_Attn represents the topographic cross attention feature, O Topo_Attn Represents the output of the terrain cross attention layer.

[0052] Furthermore, the raw data in step S4 is defined as high-resolution wind field data W HR_Norm Compared with the low-resolution typhoon wind field data W after sampling LR_Up The residual between the two, the target generated by the diffusion model is W HT_Norm With W LR_Up The residual between the two; the mean and variance of the standard Gaussian noise added to each time step t (0≤t≤T) of the traditional diffusion model are determined by the hyperparameter β t Determine and generate the noise-added sample x accordingly t ; Using the terrain embedding noise generated in step S2 to replace the standard Gaussian noise, the process of the forward diffusion process is expressed as:

[0053]

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

[0055]

[0056] Wherein, c is the diffusion control condition (including typhoon background information and high-resolution terrain data in this invention), The denoising network is estimated as the mean function of the neural network prediction, and its input is the current noise sample and diffusion control condition. is the noise variance hyperparameter.

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

[0058]

[0059] Among them, L represents the loss function, ∩ represents the real noise, ∩ θ (x t ,x0,E Topo ,c) represents the model prediction noise. To enhance the model’s ability to model high-frequency details of the wind farm, a wind farm high-frequency loss attention mechanism is introduced to perform weighted adjustment on the loss function; the calculation process of the weighted loss is as follows:

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

[0061] L Final =Weight⊙L

[0062] Among them, DWT(·) represents discrete wavelet transform, Weight represents loss weight, L Final Represents the final loss.

[0063] The 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 terrain data, and completes missing value filling, standardization, upsampling, and encoding preprocessing;

[0065] Terrain embedding noise modulation module: Use the Segmentation Everything Model (SAM) to automatically segment the pre-processed high-resolution terrain data, calculate the DEM mean of the terrain mask area, and modulate the terrain embedding noise based on this;

[0066] A denoising model based on the U-Net architecture was constructed: the encoder compresses the input data into a latent space representation using four downsampling blocks, and the decoder restores it to its original spatial distribution using four upsampling blocks. The network integrates Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up- and downsampling layers. Multi-scale information is fused through skip connections, and typhoon background information is embedded through a feedforward neural network.

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

[0068] Wind field generation module: Starting from standard Gaussian noise and terrain embedding features, combined with the trained diffusion model, generates residuals and combines them with the low-resolution wind field upsampling results to finally generate high-resolution wind field forecasts.

[0069] An electronic device comprises: a memory for storing a computer program; and a processor for implementing the steps of the above-mentioned typhoon wind field downscaling method integrating terrain segmentation and diffusion generation model when executing the computer program.

[0070] A 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 that integrates terrain segmentation and diffusion generation models.

[0071] The beneficial effects of the present invention 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, complex calculations, and poor real-time performance. To address this problem, the present invention adopts a diffusion model, combined with the fine-grained segmentation capability of segmenting all large models, to guide the model to generate a wind field that is highly correlated with the local terrain distribution. In addition, a high-frequency loss attention mechanism for wind fields is proposed to guide the model to enhance its modeling capability for high-frequency components of wind fields, effectively retain small-scale variation characteristics, and improve the deficiency of the current model output being too smooth. The invention can more accurately simulate the impact of complex terrain on typhoon wind fields, has the characteristics of high precision, high computational 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 the present invention, which integrates terrain segmentation and diffusion generation model, can accurately generate a wind field that is highly correlated with the local terrain by introducing the diffusion model and the fine-grained segmentation capability of the Segment All Large Model (SAM), significantly improve the downscaling accuracy, capture the high-frequency details of the wind field, and overcome the shortcomings of the existing technology that relies on simplified assumptions or smooth models. At the same time, the proposed wind field high-frequency loss attention mechanism effectively guides the model to enhance the modeling ability of high-frequency components, avoids excessively smooth output, retains small-scale variation characteristics, and improves the simulation accuracy of complex terrain effects. In addition, the present invention has high computational efficiency and strong real-time performance, and is suitable for scenarios that require rapid response, such as weather forecasting and disaster warning. It also has a wide range of applicability, especially showing stronger adaptability under extreme weather conditions such as typhoons. It can be used for theoretical research and can also play an important role in actual meteorological monitoring and disaster warning.

[0074] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0076] Figure 1 This is a flow chart of the typhoon wind field downscaling method integrating terrain segmentation and diffusion generation model of the present invention;

[0077] Figure 2 A flow chart of the terrain-embedded noise modulation module of the present invention;

[0078] Figure 3 A flowchart of the denoising model based on the U-Net architecture is constructed for the present invention. DETAILED DESCRIPTION

[0079] The following describes the embodiments of the present invention through specific examples. Those skilled in the art will readily understand the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the details in this specification may be modified or altered 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 in the figure, which integrates terrain segmentation and diffusion generation model, includes the following steps:

[0081] S1. Determine data sources and preprocess them. Determine the sources of low-resolution typhoon wind data, high-resolution wind 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, and aspect) used as input into the typhoon wind field downscaling model.

[0082] The obtained raw data is preprocessed, including missing value filling, data standardization, low-resolution data upsampling and category variable encoding. The preprocessed low-resolution typhoon wind field data W is obtained. LR_Norm , low-resolution typhoon wind field data W after upsampling LR_Up , high-resolution wind field data W HR_Norm , Typhoon Background Information TC Norm , high-resolution terrain data Topo HR_Norm .

[0083] S2, modulated terrain embedded noise. Figure 2 As shown, the module first automatically segments the preprocessed high-resolution terrain data using the SegmentAnything Model (SAM), then calculates the DEM mean of the area based on the terrain mask, and performs terrain embedding noise modulation based on the mean.

[0084] The automatic segmentation process is:

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

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

[0087] The mean calculation process is:

[0088]

[0089] Among them, A i Represents the area corresponding to the i-th mask, μ(A i ) represents the DEM mean of the area, |A i |Represents Region A i The number of data points in the DEM x Representing Region A i The elevation value of each data point x in .

[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, build a denoising network based on U-Net architecture. Figure 3 The network shown in the figure 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 Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up- and downsampling layers. The encoder compresses the input data into a compact latent space representation, while the decoder restores it to its original spatial distribution. A skip connection mechanism is introduced between the encoder and decoder to integrate low-level details with high-level semantics. Furthermore, typhoon background information is projected through a feedforward neural network and embedded into each Gate-ConvNeXt module.

[0095] The calculation formula of 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] Among them, x GC Represents the Gate-ConvNeXt module input, x Context represents the preprocessing context features, x Gate represents the preprocessing gate feature, F Context represents context features, F Gate represents the gated feature, O GCrepresents 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.

[0101] The calculation formula for typhoon background information embedding 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] Among them, E Lonlat represents the longitude and latitude of the typhoon center, E Tp Represents the typhoon background information embedding, O GC_Embed Represents the output of the Gate-ConvNeXt module after adding the embedding information.

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

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

[0108]

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

[0110] Among them, F Spa_Attn represents the spatial attention feature, Q Spa_Attn Represents the output of the spatial attention layer.

[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] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and x Topo_Attn represents the terrain cross attention layer input, softmax represents the softmax activation layer, d represents the number of dimensions of query and key, F Topo_Attn represents the topographic cross attention feature, O Topo_Attn Represents the output of the terrain cross attention layer.

[0117] S4. Construct a 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 x0~q(x0), standard Gaussian noise is gradually added over T time steps, thereby gradually transforming the original data distribution into a standard Gaussian distribution.

[0118] In the present invention, the original data is defined as high-resolution wind field data W HR_Norm Compared with the low-resolution typhoon wind field data W after sampling LR_Up Therefore, the target of the diffusion model generation in the present invention is W HR_Norm With W LR_Up The mean and variance of the standard Gaussian noise added to each time step t (0≤t≤T) in the traditional diffusion model are determined by the hyperparameter β t Determine and generate the noise-added sample x accordingly t The present invention uses the terrain embedded noise generated in step S2 to replace the standard Gaussian noise, and the process of the forward diffusion process is expressed as:

[0119]

[0120] Where I represents the identity matrix. In contrast, the reverse denoising process gradually restores the noise samples to the target wind field, and the process can be expressed as:

[0121]

[0122] Wherein, c is the diffusion control condition (including typhoon background information and high-resolution terrain data in this invention), The denoising network is estimated as the mean function of the neural network prediction, and its input is the current noise sample and diffusion control condition. is the noise variance hyperparameter.

[0123] S5. Train the denoising model. During the training process, the backpropagation algorithm is used to optimize the model parameters. By minimizing the L2 loss between the model's predicted noise and the actual noise, the diffusion model is effectively learned. The calculation process of the loss function is as follows:

[0124]

[0125] Among them, L represents the loss function, ∈ represents the real noise, ∈ θ (x t ,x0,E Topo ,c) represents the model prediction noise. To enhance the model's ability to model high-frequency details of the wind farm, a wind farm high-frequency loss attention mechanism is introduced to perform weighted adjustment on the loss function. The calculation process of the weighted loss is as follows:

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

[0127] L Final =Weight⊙L

[0128] Among them, DWT(·) represents discrete wavelet transform, Weight represents 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 it with the embedded features of the target terrain area, and uses it as the starting point of the reverse denoising process. Then, by using the trained diffusion model and diffusion control conditions, the reverse diffusion process is used to gradually denoise and generate W HR_Norm With W LR_Up Finally, the residual is compared with W LR_Up Combined with the final prediction, the generation of high-resolution near-surface wind field is achieved.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention 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 invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A typhoon wind field downscaling method integrating terrain segmentation and diffusion generation model, characterized in that: The following steps are involved: S1. Determine the data source and preprocess it; determine the sources of low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution terrain data used to input the typhoon wind field downscaling model, and preprocess the obtained raw data; S2, modulated terrain embedded noise; The pre-processed high-resolution terrain data is automatically segmented using the Segmentation Everything Model (SAM). The DEM mean of the area is calculated based on the terrain mask, and terrain embedding noise modulation is performed based on the mean. S3. Construct a denoising network based on the U-Net architecture. This network 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 Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up- and downsampling layers. The encoder is responsible for compressing the input data into a compact latent space representation, while the decoder restores it to its 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 embedded in each Gate-ConvNeXt module after projection through a feedforward neural network. 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 exist In each time step, standard Gaussian noise is gradually added to gradually transform the original data distribution into a standard Gaussian distribution; S5. Train the denoising model. During the training process, the diffusion model uses the back propagation algorithm to optimize the model parameters. By minimizing the L2 loss between the model's predicted noise and the actual noise, the diffusion model is effectively learned. S6, wind field generation stage; sampling random noise from the standard Gaussian distribution in step S4 and combining it with the target terrain embedding features as the starting point for reverse denoising; using the trained diffusion model and control conditions, gradually denoising through reverse diffusion to generate and The residual between Combined, the final high-resolution wind field forecast is generated.

2. The typhoon wind field downscaling method according to claim 1, characterized in that: The processing of the original data in step S1 specifically includes: missing value filling, data standardization, low-resolution data upsampling and categorical variable encoding.

3. The typhoon wind field downscaling method according to claim 1, characterized in that: The typhoon background information in step S1 includes the typhoon center longitude , central latitude ,strength , the lowest pressure near the center , 2-minute average maximum sustained wind speed near the center , high-resolution terrain data including digital elevation models , terrain slope , slope direction ; Obtain pre-processed low-resolution typhoon wind field data , low-resolution typhoon wind field data after sampling , high-resolution wind field data , Typhoon background information , high-resolution terrain data .

4. The typhoon wind field downscaling method according to claim 1, characterized in that: The automatic segmentation process in step S2 is: in, ( ) represents the first Mask; The mean calculation process is: in, Represents the area corresponding to the i-th mask, represents the DEM mean of the area, Representative area The number of data points in , Representative area Each data point in The elevation value of The terrain embedding noise modulation process is as follows: in, represents the terrain embedding, Represents the splicing operation, represents standard Gaussian noise, Represents terrain embedding noise.

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

6. The typhoon wind field downscaling method according to claim 5, characterized in that: The raw data in step S4 is defined as high-resolution wind field data Compared with the low-resolution typhoon wind field data after sampling The residual between the two, the target generated by the diffusion model is and The residual between each time step of the traditional diffusion model The mean and variance of the added standard Gaussian noise are determined by the hyperparameters Determine and generate noise-added samples accordingly ; The standard Gaussian noise is replaced by the terrain embedding noise generated in step S2, and the process of the forward diffusion process is expressed as: in, Represents the identity matrix. On the contrary, the reverse denoising process gradually restores the noise samples to the target wind field. The process can be expressed as: in, is the diffusion-controlled condition, The denoising network is estimated as the mean function of the neural network prediction, and its input is the current noise sample and diffusion control condition. is the noise variance hyperparameter.

7. The typhoon wind field downscaling method according to claim 6, characterized in that: The calculation process of the loss function in step S5 is as follows: in, represents the loss function, represents the real noise, The representative model predicts noise. To enhance the model's ability to model high-frequency details of the wind farm, a high-frequency loss attention mechanism for the wind farm is introduced to perform weighted adjustment on the loss function. The calculation process of the weighted loss is as follows: in, Discrete Wavelet Transform, represents the loss weight, Represents the final loss.

8. A typhoon wind field downscaling system integrating terrain segmentation and diffusion generation model, characterized by: include: Data preprocessing module: Integrates low-resolution typhoon wind field data, high-resolution typhoon wind field data, typhoon background information, and high-resolution terrain data, and completes missing value filling, standardization, upsampling, and encoding preprocessing; Terrain embedding noise modulation module: Use the Segmentation Everything Model (SAM) to automatically segment the pre-processed high-resolution terrain data, calculate the DEM mean of the terrain mask area, and modulate the terrain embedding noise based on this; A denoising model based on the U-Net architecture was constructed: the encoder compresses the input data into a latent space representation using four downsampling blocks, and the decoder restores it to its original spatial distribution using four upsampling blocks. The network integrates Gate-ConvNeXt modules, spatial self-attention layers, terrain cross-attention layers, and up- and downsampling layers. Multi-scale information is fused through skip connections, and typhoon background information is embedded through a feedforward neural network. Diffusion model: The original data is gradually added with Gaussian noise through the forward diffusion process to convert 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 residual of the sampling result; Wind field generation module: Starting from standard Gaussian noise and terrain embedding features, combined with the trained diffusion model, generates residuals and combines them with the low-resolution wind field upsampling results to finally generate high-resolution wind field forecasts.

9. An electronic device, characterized in that: include: memory for storing computer programs; A processor, configured to implement the steps of the typhoon wind field downscaling method integrating terrain segmentation and diffusion generation model as described in any one of claims 1 to 7 when executing a computer program.

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 that integrates terrain segmentation and diffusion generation models as described in any one of claims 1 to 7.

Images